Plants are vital in sustaining life on Earth by converting carbon dioxide (CO2) into oxygen (O2) and serving as a primary food source for most living organisms while also playing a crucial role in various industrial sectors, from pharmaceuticals to food products. Even for experts, accurately identifying and cataloging plant species is challenging due to the vast biodiversity and unique characteristics. To address this, current state-of-the-art approaches rely on supervised deep learning models, which, while effective, still depend on vast annotated datasets. In this study, we propose a self-taught learning approach to mitigate the reliance on annotated plant species data. We employ a convolutional auto-encoder to learn unsupervised deep representations from unlabeled databases across different and similar domains. These learned representations, known as latent vectors, are utilized to extract deep features from target datasets such as the well-known Flavia and Leafsnap, which are then used to train domain-specific classifiers. Furthermore, we explore a fusion of different architectures for representation learning. The contribution of this work lies in developing a self-taught learning pipeline for plant species recognition, achieving competitive results without the need for annotated datasets when compared to deep supervised approaches.

Heart failure (HF) represents a global clinical and economic challenge, with hospitalizations accounting for 65% of disease-related costs. This study proposes an approach to predict HF hospitalizations using Large Language Models (LLMs) trained on chronological data from Brazilian health insurance beneficiaries. By converting administrative records (consultations, medications, diagnoses) into temporal narratives, models like RoBERTa and Open-Cabrita3B were fine-tuned to identify clinical deterioration patterns. The HealthHistoryRoBERTapt model, trained with historical health insurance data and specifically adjusted for HF, achieved an AUC-ROC of 0.93-0.95 in prediction windows from 5 to 180 days, significantly outperforming other studies (AUC 0.63-0.76) using static clinical data or basic demographics, and those combining clinical-administrative data (AUC 0.82). It is noteworthy that the ability of the model to maintain an F1-score greater than 0.85 and sensitivity of 0.87 in predictions of up to 150 days, revealing that administrative variables (e.g., history of hospitalizations, frequency of consultations) function as effective proxies for socioeconomic and behavioral factors, traditionally neglected. Compared to other works, this study demonstrates that longitudinal health insurance data combined with NLP techniques capture non-linear risk trajectories, enabling precise predictions for strategic health planning.

Kidney stones significantly impact healthcare systems, with diagnosis typically requiring time-consuming Computed Tomography (CT) scan consultations between physicians and radiologists, often delaying patient care. Achieving a quick and accurate diagnosis is essential to ensure timely and effective treatment, which has motivated the development of Deep Neural Network (DNN)-based approaches for automated kidney stone detection. However, building effective models remains challenging, as it often requires access to large and diverse datasets that are siloed across institutions, and sharing such medical data is rarely feasible due to strict privacy regulations and patient confidentiality concerns. This paper proposes a privacy-preserving Federated Learning (FL) framework that enables multiple medical institutions to collaboratively train a DNN model without sharing sensitive patient data. Each institution trains a local model on its private dataset, and a centralized trusted server securely aggregates model parameters. We evaluate our approach using abdominal CT scan image datasets from two distinct institutions. Experimental results demonstrate that our proposed model achieves high classification accuracy within the same training environment, with an F1-score of up to 0.94. In addition, in cross-dataset evaluations, our approach outperforms traditional centralized baselines, showing significantly lower performance degradation while preserving patient privacy.

The society produces textual data online in several ways, e.g., via reviews and social media posts. Therefore, numerous researchers have been working on discovering patterns in textual data that can indicate peoples’ opinions, interests, and so on. Most tasks regarding natural language processing are addressed using traditional machine learning methods and static datasets. This setting can lead to several problems, e.g., outdated datasets and models, which degrade in performance over time. This is particularly true regarding concept drift, in which the data distribution changes over time. Furthermore, text streaming scenarios also exhibit further challenges, such as the high speed at which data arrive over time. Models for stream scenarios must adhere to the aforementioned constraints while learning from the stream, thus storing texts for limited periods and consuming low memory. This study presents a systematic literature review regarding concept drift adaptation in text stream scenarios. Considering well-defined criteria, we selected 48 papers published between 2018 and August 2024 to unravel aspects such as text drift categories, detection types, model update mechanisms, stream mining tasks addressed, and text representation methods and their update mechanisms. Furthermore, we discussed drift visualization and simulation and listed real-world datasets used in the selected papers. Finally, we brought forward a discussion on existing works in the area, also highlighting open challenges and future research directions for the community.

Pre-trained language models (LMs) have been used in several scenarios and data mining tasks due to their good-quality representations and their use readiness. Although LMs constitute a significant gain in usability, they are frequently utilized statically over time, meaning that these models can suffer from concept drift and semantic shift, which correspond to changes in data distribution and word meanings. These phenomena are more noticeable when new texts become gradually available. This paper evaluates the impact of updating pre-trained SentenceBERT models overtime on a Brazilian news post classification task in text streaming fashion, a paradigm suitable for learning from data streams. While we update the SBERT model yearly with a reduced number of recent posts, we compare it with scenarios using static LMs. We used the adaptive random forest for classification and evaluated it regarding macro F1-score and elapsed time. The experimental results show that regularly leveraging sampled texts from the recent past for fine-tuning LMs can improve performance metrics over time, reaching better results than using static LMs in most years analyzed. We also evaluated the run times, which suggests that fine-tuning LMs over time provides a good trade-off between performance and run time.

Fuels are crucial for any country's development and economy, impacting various sectors such as transportation, industry, and electricity generation. Accurate prediction of monthly fuel demand can improve supply chain management, strategic decision-making, and financial planning for businesses while helping governments develop decarbonization policies and estimate pollutant emissions. This paper explores machine learning models to forecast fossil fuels and biofuel demand 12 months ahead, using univariate time series data representing the historical sales of 27 Brazilian states, one of the world's leading producers and consumers of fuels. We evaluate different time series feature sets, machine learning regression models, and prediction strategies to address the complexity of fuel sales influenced by factors such as economic conditions and geopolitical events. Our comprehensive evaluation aims to determine an effective setting for predictive models in the fuel domain. Our results show that popular feature extractors for time series, such as Catch22 and TsFresh, cannot improve the original data representation for most forecasting models. Although focused on Brazil, our findings apply to other countries, since the trained models do not rely on external variables, such as micro and macroeconomic indicators.

Audio replay attacks present a significant challenge to automatic speaker verification systems (ASVs), emphasizing the need for effective detection methods. Traditionally, embedding-based approaches, such as those leveraging Convolutional Neural Networks (CNNs), have been used. However, dissimilarity-based methods emerge as a promising alternative, offering potential advantages in detecting subtle differences between genuine and spoofed audio. This study evaluates dissimilarity strategies for detecting genuine versus spoofed audio signals using a well-known benchmark dataset and established metrics, including accuracy and Equal Error Rate (EER). We provide a comparative performance assessment of various CNN architectures and dissimilarity strategies, finding that while dissimilarity approaches are competitive with embedding-based methods, the Dissimilarity Vectors strategy outperforms the Dissimilarity Space strategy.

Plant diseases are one of the factors that compro-mise food production goals. Tomatoes are one of the world's most consumed vegetables and are widely affected by various diseases. Tomato cultivation in greenhouses enables continuous production. In this context, this research focuses on identifying diseases in greenhouse tomato cultivation scenarios. For this study, new datasets were created with two image sizes: the Tomato Leaf Image Dataset (TLID) with image sizes of 256 x 256 pixels and 15,256 images, and the Patch Tomato Leaf Image Dataset (PTLID) with patch sizes of 32 x 32 pixels and 227,218 images. Both datasets comprise seven classes, including four types of diseases, two combinations of diseases on the same leaf, and the healthy leaf. Machine Learning techniques have been widely used to identify plant diseases. This work presents two machine learning methods tested with both datasets. In the proposed models, three convolutional neural networks were combined: a customized CNN, VGG19, and Resnet50, along with two voting classification methods using Hard and Soft de-cisions. The evaluation conducted on the datasets demonstrated that using patches significantly improves results, achieving an accuracy of 90.48%. This technique enables the identification of the disease stage.

The number of Android malicious applications keeps growing as time passes, even paving their way to official app markets. In recent years, a promising malware detection approach makes use of the compiled app source codes (dex), through convolutional neural networks (CNN) as an image classification task. Unfortunately, current proposals often rely on unrealistic datasets, focusing their detection on the mal-ware families, while neglecting the detection of malware apps in the first place. In this paper, we propose a reliable and hierarchical Android malware detection through an image-based CNN scheme, implemented twofold. First, Android malware classification is performed in a hierarchically-structured local manner, initially identifying malware apps, then, their related family. Second, to ensure reliability and improve classification accuracy, only highly confident classified apps are reported, in a classification with reject option rationale. Experiments performed in a new dataset with over 26 thousand Android apps, divided into 29 malware families, compounding over 13 GB of app dex images, have shown that current image-based CNN for malware detection is unable to provide high detection accuracies. In contrast, our proposed model is able to reliably detect malware apps, improving the true-negative rates by up to 5.5%, and the average true-positive rate of the malware families of accepted apps by up to 12.7%, while rejecting only 10% of Android apps.

The information provided by the vehicle’s sensors allows the estimation of critical parameters in a pre-crash scenario. However, the available open multimodal datasets focused on optical sensors usually present a low-resolution radar. Moreover, the sensors’ positioning does not provide a representative view of the objects in the last meters around them. With this in mind, we introduce a new approach to fuse camera, lidar, and radar of high resolution at a very early stage using a fully convolutional network and an anchor-free strategy to detect the front side of a vehicle, estimating distance and orientation in a single step. A robust experimental protocol based on a new multimodal dataset shows that the proposed fusion of the three sensors brings more stability and accuracy in detection and regression tasks by better representing far and near ranges around the vehicle where the sensors response to objects change significantly.

Social media has been a data source for various applications, given its characteristic of working as a social sensor. Many applications in several areas, such as brand reputation and online opinion monitoring, use this valuable resource to understand the users of services and products. This paper describes an application in the soccer domain, considering data collected from a social media textual data stream. The goal is to detect possible sentiment drifts related to actual events in a soccer match. This task is challenging as we resort to short texts made available during a short time (match length). We evaluated four drift detectors using four metrics: false alarms, delay (considering the number of posts), delay, and missing drifts. Our results show that ADWIN had a stable performance in sentiment drift detection compared to other methods in timely detecting the flagged drifts, raising a small number of false alarms. Given the drifts detected, we used Incremental Word-Vectors to monitor words of interest and check their relatedness to actual events in the match. We empirically assert that the closest words trace back to the sentiment drift generator events.

Data normalization is an essential task when modeling a classification system. When dealing with data streams, data normalization becomes especially challenging since we may not know in advance the properties of the features, such as their minimum/maximum values, and these properties may change over time. We compare the accuracies generated by eight well-known distance functions in data streams without normalization, normalized considering the statistics of the first batch of data received, and considering the previous batch received. We argue that experimental protocols for streams that consider the full stream as normalized are unrealistic and can lead to biased and poor results. Our results indicate that using the original data stream without applying normalization, and the Canberra distance, can be a good combination when no information about the data stream is known beforehand.

This work draws attention to the large fraction of near-duplicates in the training and test sets of datasets widely adopted in License Plate Recognition (LPR) research. These duplicates refer to images that, although different, show the same license plate. Our experiments, conducted on the two most popular datasets in the field, show a substantial decrease in recognition rate when six well-known models are trained and tested under fair splits, that is, in the absence of duplicates in the training and test sets. Moreover, in one of the datasets, the ranking of models changed considerably when they were trained and tested under duplicate-free splits. These findings suggest that such duplicates have significantly biased the evaluation and development of deep learning-based models for LPR. The list of near-duplicates we have found and proposals for fair splits are publicly available for further research at https://raysonlaroca.github.io/supp/lpr-train-on-test/.

Feature extraction regards transforming unstructured or semi-structured data into structured data that can be used as input for classification and sentiment analysis algorithms, among other applications. This task becomes even more challenging and relevant when textual data becomes available over time as a continuous data stream since the lexicon and semantics can be ever-evolving. Data streams are, by definition, potentially infinite sequences of data that may have ephemeral characteristics, that is, where the data behavior changes, it leads to a phenomenon named concept drift. Textual data streams are specialized data streams, in which texts arrive over time from a continual data source, such as social media, raising challenges in which feature extractors are of great help. In this paper, we benchmark different feature extraction algorithms, i.e., Hashing Trick, Word2Vec, BERT, and Incremental Word-Vectors; in textual data stream classification, considering different stream lengths. The evaluation was performed over a binary and a multiclass classification task, considering two different datasets. Results show that pre-trained models, such as BERT, achieve interesting results, while Hashing Trick also performs competitively. We also observe that incremental methods such as Word2Vec and Incremental Word-Vectors are the most prepared for changing scenarios, yet, they are much more computationally intensive compared to the former when applied to larger streams.

School dropout is a problem that impacts many socioeconomic aspects, including inequality.Dropout prediction algorithms can help remediate this problem, although several past attempts in the literature did so using small datasets.This paper brings forward an experimental approach of machine learning for school dropout prediction in Brazilian schools.The data used for this study was first retrieved from the academic systems of a group of Brazilian private schools, which was later enriched with socio-economic data extracted from governmental sources.Using the dataset to train different types of classifiers, we obtained up to 95.2% precision rates when predicting dropout at different year and educational stages, thus allowing schools to plan and apply retention strategies.

This paper describes a method to perform the tasks of Image Retrieval and Pattern Spotting in a collection of historical documents, in a zero-shot manner, i.e. without prior knowledge or any training on the pattern to be searched. The proposed method measures the similarity between images using representation schemes based on feature maps provided by intermediate layers of a CNN. Moreover, to improve the time response and reduce the storage need, we propose to binarize these image representations. Experimental results obtained on the DocExplore dataset show that the proposed method improves the mAP on this dataset by 38.46% for Image Retrieval and 134.6% for Pattern Spotting compared to the state-of-the-art methods. The proposed binarization strategy provides a reduction of memory usage by a factor of 16 with a slight decrease of less than one percentage point in the mAP for both tasks and a reduction of 13.3% in search time when using the binary representation.

This work describes different strategies to generate unsupervised representations obtained through the concept of self-taught learning for facial emotion recognition (FER). The idea is to create complementary representations promoting diver-sity by varying the autoencoders' initialization, architecture, and training data. SVM, Bagging, Random Forest, and a dynamic ensemble selection method are evaluated as final classification methods. Experimental results on JAFFE and Cohn-Kanade datasets using a leave-one-subject-out protocol show that FER methods based on the proposed diverse representations compare favorably against state-of-the-art approaches that also explore unsupervised feature learning.

When using vision-based approaches to classify individual parking spaces between occupied and empty, human experts often need to annotate the locations and label a training set containing images collected in the target parking lot to fine-tune the system. We propose investigating three annotation types (polygons, bounding boxes, and fixed-size squares), providing different data representations of the parking spaces. The rationale is to elucidate the best trade-off between handcraft annotation precision and model performance. We also investigate the number of annotated parking spaces necessary to fine-tune a pre-trained model in the target parking lot. Experiments using the PKLot dataset show that it is possible to fine-tune a model to the target parking lot with less than 1,000 labeled samples, using low precision annotations such as fixed-size squares.

Over the last years, several works have proposed highly accurate Android malware detection techniques. Surprisingly, modern malware apps can still pave their way to official markets, thus, demanding the provision of more robust and accurate detection approaches. This paper proposes a new multi-view Android malware detection through image-based deep learning, implemented threefold. First, apps are evaluated according to several feature sets in a multi-view setting, thus, increasing the information provided for the classification task. Second, extracted feature sets are converted to an image format while maintaining the principal components of the data distribution, keeping the information for the classification task. Third, built images are jointly represented in a single shot, each in a predefined image channel, enabling the application of deep learning architectures. Experiments on a new version of a publicly available Android malware dataset composed of over 11 thousand Android apps have shown our proposal's feasibility. It reaches true-negative rates of up to 99.5% when implemented with a single-view approach with our new image-building technique. In addition, if our proposed multi-view scheme is used, the classification accuracies of malware families become more stable, reaching a true-positive rate of up to 98.7%.

Hospitalizations represent an expressive part of total health costs and, therefore, reducing the number of hospitalizations, when possible, can generate both economic gains and enhanced quality of life of patients. Several works have been striving to use machine learning to create models for hospitalization predictions. Most of them require specialized knowledge in the health area, mainly in the stages of data preparation and selection of features. This feature engineering is not always perfect and may fail to select relevant features for the model training process. In this paper, to fill this gap, we explore three sources of information to extract features, i.e., medical specialty, event description, and the International Classification of Diseases. In addition, we introduce a dataset composed of 38,524 records of medical events from 34,930 patients. To assess and set a baseline for this new dataset, we have used two well-known ensemble methods (Random Forest and Gradient Boosting). The best results, AUC = 0.82, were achieved by combining the models generated from the three feature set tested and gradient boosting. We believe that researchers will find this dataset a valuable tool in their work on hospitalization prediction. It will also make future benchmarking and evaluation possible.

This paper presents a deep learning approach for image retrieval and pattern spotting in digital collections of historical documents. First, a region proposal algorithm detects object candidates in the document page images. Next, deep learning models are used for feature extraction, considering two distinct variants, which provide either real-valued or binary code representations. Finally, candidate images are ranked by computing the feature similarity with a given input query. A robust experimental protocol evaluates the proposed approach considering each representation scheme (real-valued and binary code) on the DocExplore image database. The experimental results show that the proposed deep models compare favorably to the state-of-the-art image retrieval approaches for images of historical documents, outperforming other deep models by 2.56 percentage points using the same techniques for pattern spotting. Besides, the proposed approach also reduces the search time up to 200$\times$, and the storage cost up to 6,000$\times$ when compared to related works based on real-valued representations.

The plant species classification using leaf images is a challenge due to the lack of annotation, imbalanced classes and similarities in the data representation. For such problems, Siamese Neural Networks (SNN’s) have been used to overcome these bottlenecks in several contexts. In light of this, this work evaluates different architectures trained in Siamese manner for classifying plant species from the leaf image. Besides, we combined features from the intermediate convolutional layers to improve representations. Experiments on the well-known Flavia and MalayaKew databases have shown that the fusion of intermediate features results in a relevant gain in performance.

Mining data streams is a hot topic in the machine learning (ML) community. In addition to learning and updating accurate models over time, these techniques must respect constraints that are not necessarily as strong in batch mode, such as time processing and memory consumption efficiency. A successful family of techniques in batch ML is dynamic classifier selection (DCS). However, these are roughly overlooked in data stream mining. In this paper, we propose a novel dynamic classifier selection framework for data streams called Double Dynamic Classifier Selection (DDCS). We compare DDCS against state-of-art methods for mining data streams in both synthetic and realworld datasets. Results depict that DDCS not only outperforms the state-of-art ensemble methods for data stream classification in terms of accuracy but is also significantly more efficient in terms of processing time and memory consumption.

The use and development of new Advanced Driver Assistance Systems (ADAS) based on fusion of various sensors, aiming to achieve a higher robustness, are becoming more common. In pre-crash scenarios, the extraction of information from an incoming vehicle should be done as near as possible to the moment of impact "t0", where relevant issues such as disturbances in the sensor signal and partial occlusion of the vehicle play an important role. The presented work introduces a novel fusion approach between a Lidar and a camera to extract the distance and the approach angle of an oncoming vehicle. It detects the bullet vehicle's license plate using a deep learning neural network and fuses the Lidar information to estimate its distance and approach angle. The proposed architecture was evaluated on data from a benchmark dataset - KITTI, and with a pre-crash scenario dataset collected inside the CARISSMA Research Center indoor hall facilities. The algorithm reached an accuracy of 0.16m±0.26m for the distance and 5.48°±4.99° for the angle estimation on the benchmark dataset. On the pre-crash dataset, the experimental results yielded an accuracy of 0.022m ± 0.022m and 2.82°± 1.98° for the distance and angle, respectively. The system can extract parameters until 0.4m before a collision, and can overcome blurriness and partial occlusion conditions, being also easily reproducible in other sensor setups.

Usually, in a real-world scenario, few signature samples are available to train an automatic signature verification system (ASVS). However, such systems do indeed need a lot of signatures to achieve an acceptable performance. Neuromotor signature duplication methods and feature space augmentation methods may be used to meet the need for an increase in the number of samples. Such techniques manually or empirically define a set of parameters to introduce a degree of writer variability. Therefore, in the present study, a method to automatically model the most common writer variability traits is proposed. The method is used to generate offline signatures in the image and the feature space and train an ASVS. We also introduce an alternative approach to evaluate the quality of samples considering their feature vectors. We evaluated the performance of an ASVS with the generated samples using three well-known offline signature datasets: GPDS, MCYT-75, and CEDAR. In GPDS-300, when the SVM classifier was trained using one genuine signature per writer and the duplicates generated in the image space, the Equal Error Rate (EER) decreased from 5.71% to 1.08%. Under the same conditions, the EER decreased to 1.04% using the feature space augmentation technique. We also verified that the model that generates duplicates in the image space reproduces the most common writer variability traits in the three different datasets.

This paper describes a classifier pool generation method guided by the diversity estimated on the data complexity and classifier decisions. First, the behavior of complexity measures is assessed by considering several subsamples of the dataset. The complexity measures with high variability across the subsamples are selected for posterior pool adaptation, where an evolutionary algorithm optimizes diversity in both complexity and decision spaces. A robust experimental protocol with 28 datasets and 20 replications is used to evaluate the proposed method. Results show significant accuracy improvements in 69.4% of the experiments when Dynamic Classifier Selection and Dynamic Ensemble Selection methods are applied.

Automatic facial expression recognition (FER) is an important research area in the emotion recognition and computer vision. Applications can be found in several domains such as medical treatment, driver fatigue surveillance, sociable robotics, and several other human-computer interaction systems. Therefore, it is crucial that the machine should be able to recognize the emotional state of the user with high accuracy. In recent years, deep neural networks have been used with great success in recognizing emotions. In this paper, we present a new model for continuous emotion recognition based on FER by using an unsupervised learning approach based on transfer learning and autoencoders. The proposed approach also includes preprocessing and post-processing techniques which contribute favorably to improving the performance of predicting the concordance correlation coefficient for arousal and valence dimensions. Experimental results for predicting spontaneous and natural emotions on the RECOLA 2016 dataset have shown that the proposed approach based on visual information can achieve concordance correlation coefficient of 0.516 and 0.264 for valence and arousal, respectively.

Data imbalance is a major problem that affects several machine learning (ML) algorithms. Such a problem is troublesome because most of the ML algorithms attempt to optimize a loss function that does not take into account the data imbalance. Accordingly, the ML algorithm simply generates a trivial model that is biased toward predicting the most frequent class in the training data. In the case of histopathologic images (HIs), both low-level and high-level data augmentation (DA) techniques still present performance issues when applied in the presence of inter-patient variability; whence the model tends to learn color representations, which is related to the staining process. In this paper, we propose a novel approach capable of not only augmenting HI dataset but also distributing the inter-patient variability by means of image blending using the Gaussian-Laplacian pyramid. The proposed approach consists of finding the Gaussian pyramids of two images of different patients and finding the Laplacian pyramids thereof. Afterwards, the left-half side and the right-half side of different HIs are joined in each level of the Laplacian pyramid, and from the joint pyramids, the original image is reconstructed. This composition combines the stain variation of two patients, avoiding that color differences mislead the learning process. Experimental results on the BreakHis dataset have shown promising gains vis-à-vis the majority of DA techniques presented in the literature.

This work presents a sky and ground segmentation approach in digital images using the supervised Support Vector Machine (SVM) algorithm based on Whiteness and Blueness indexes, Local Binary Patterns, and Extended morphological profile features. The goal is to separate the image contents in two classes, the sky and the ground. The research is divided into two stages: first, the best features selected by monolithic classifiers using the cross-validation technique. The second stage based on combination of classifiers to segment sky and ground: in a first approach, the strategy consists in segmentation process without dividing the image databases into categories. The second approach carries out segmentation a pre-classification of databases into four categories, City, Highway/Road, Sea/ Harbor, and Nature/Mountain. We used two bases of 1200 images each, containing images with different sky and ground contexts. The first approach is generally adopted in the literature. The second approach, little cited in the literature, presents promising and distinct results for both image bases and allows us to see importance of dividing the images into categories, since there is alteration of the ground context, which leads to different results and greater hit rate.

Convolutional neural networks have been successfully applied for video content detection in the last years. However, such cognitive models usually demand the availability of several gigabytes of memory and present a low detection throughput, as a result, they are not feasible for resource-constrained devices, especially for real-time applications like video streaming. In this paper, we address real-time video content detection in resource-constrained devices in a threefold manner. First, we improve detection throughput by means of a frame sampling technique. Then, we propose a new evaluation measure towards proper deployment of convolutional neural networks in resource-constrained devices. Finally, we address the accuracy degradation caused by the porting of the convolutional neural network, applying a lightweight classification verification technique. The evaluation results, through a real-time demanding application, show that the proposed approach can detect up to 301 frames/sec, demanding only 9 megabytes of memory while reaching up to 89.3% of accuracy. Besides, we can increase the detection throughput by up to 10 times, with no effects on accuracy, and further increase accuracy without effects on processing demands.

An end-to-end solution for handwritten numeral string recognition is proposed, in which the numeral string is considered as composed of objects automatically detected and recognized by a YoLo-based model. The main contribution of this paper is to avoid heuristic-based methods for string preprocessing and segmentation, the need for task-oriented classifiers, and also the use of specific constraints related to the string length. A robust experimental protocol based on several numeral string datasets, including one composed of historical documents, has shown that the proposed method is a feasible end-to-end solution for numeral string recognition. Besides, it reduces the complexity of the string recognition task considerably since it drops out classical steps, in special preprocessing, segmentation, and a set of classifiers devoted to strings with a specific length.

This paper shows the susceptibility of spectrogram-based audio classifiers to adversarial attacks and the transferability of such attacks to audio waveforms. Some commonly used adversarial attacks to images have been applied to Mel-frequency and short-time Fourier transform spectrograms, and such perturbed spectrograms are able to fool a 2D convolutional neural network (CNN). Such attacks produce perturbed spectrograms that are visually imperceptible by humans. Furthermore, the audio waveforms reconstructed from the perturbed spectrograms are also able to fool a 1D CNN trained on the original audio. Experimental results on a dataset of western music have shown that the 2D CNN achieves up to 81.87% of mean accuracy on legitimate examples and such performance drops to 12.09% on adversarial examples. Likewise, the 1D CNN achieves up to 78.29% of mean accuracy on original audio samples and such performance drops to 27.91% on adversarial audio waveforms reconstructed from the perturbed spectrograms.

A common problem in machine learning is to find representative real-world labeled datasets to put the methods to test. When developing approaches to deal with concept drifts, some datasets such as the Forest Covertype and Nebraska Weather are common choices for testing, even though there is no consensus on whether these exhibit concept drifts or not. We argue that some well-known real-world concept drift datasets present a high serial dependence in the target class and may have only minor changes. With this in mind, we propose the use of Naïve methods that should be used for comparison with methods that deal with concept drifts. The experimental results using six real-world well-known concept drift datasets show that the Naïve approaches can be better than some methods to deal with possible concept drifts in datasets such as the Forest Covertype, Electricity, and Nebraska Weather. These results suggest that some widely used datasets may be trivial from the concept drift standpoint, and thus, should be avoided, or at least the results should be compared with the proposed Naïve methods.

Abstract Ensemble of classifiers can improve classification accuracy by combining several models. The fusion method plays an important role in the ensemble performance. Usually, a criterion for weighting the decision of each ensemble member is adopted. Frequently, this can be done using some heuristic based on accuracy or confidence. Then, the used fusion rule must consider the established criterion for providing a most reliable ensemble output through a kind of competition among the ensemble members. This article presents a new ensemble fusion method, named centrality score‐based fusion, which uses the centrality concept in the context of social network analysis (SNA) as a criterion for the ensemble decision. Centrality measures have been applied in the SNA to measure the importance of each person inside of a social network, taking into account the relationship of each person with all others. Thus, the idea is to derive the classifier weight considering the overall classifier prominence inside the ensemble network, which reflects the relationships among pairs of classifiers. We hypothesized that the prominent position of a classifier based on its pairwise relationship with the other ensemble members could be its weight in the fusion process. A robust experimental protocol has confirmed that centrality measures represent a promising strategy to weight the classifiers of an ensemble, showing that the proposed fusion method performed well against the literature.

Biopsies are the gold standard for breast cancer diagnosis. This task can be improved by the use of Computer Aided Diagnosis (CAD) systems, reducing the time of diagnosis and reducing the inter and intra-observer variability. The advances in computing have brought this type of system closer to reality. However, datasets of Histopathological Images (HI) from biopsies are quite small and unbalanced what makes difficult to use modern machine learning techniques such as deep learning. In this paper we propose a compact architecture based on texture filters that has fewer parameters than traditional deep models but is able to capture the difference between malignant and benign tissues with relative accuracy. The experimental results on the BreakHis dataset have show that the proposed texture CNN achieves almost 90% of accuracy for classifying benign and malignant tissues.

Due to the poor condition of most of historical documents, binarization is difficult to separate document image background pixels from foreground pixels. This paper proposes Convolutional Neural Networks (CNNs) based on predicted two-channel images in which CNNs are trained to classify the foreground pixels. The promising results from the use of multispectral images for semantic segmentation inspired our efforts to create a novel prediction-based two-channel image. In our method, the original image is binarized by the structural symmetric pixels (SSPs) method, and the two-channel image is constructed from the original image and its binarized image. In order to explore impact of proposed two-channel images as network inputs, we use two popular CNNs architectures, namely SegNet and U-net. The results presented in this work show that our approach fully outperforms SegNet and U-net when trained by the original images and demonstrates competitiveness and robustness compared with state-of-the-art results using the DIBCO database.

Video as well as music are potent means to convey emotions. However, despite their importance in several applications, few works deal with the issue of emotion classification in videos. The main reason is possibly the lack of available databases. In this work we extend the CAL500 database by including music videos, since the CAL500 was originally proposed as an audio-only database. The main rationale here is that the music videos must be official as they were developed to convey the same emotion as the song. After adapting the database, we have extracted audio and video features to perform our computational experiments. Our main result is that there is a complementarity between the audio and video features as the best result was achieved using their combination.

In this paper, the concept of representation learning based on deep neural networks is applied as an alternative to the use of handcrafted features in a method for automatic visual inspection of corroded thermoelectric metallic pipes. A texture convolutional neural network (TCNN) replaces hand-crafted features based on Local Phase Quantization (LPQ) and Haralick descriptors (HD) with the advantage of learning an appropriate textural representation and the decision boundaries into a single optimization process. Experimental results have shown that it is possible to reach the accuracy of 99.20% in the task of identifying different levels of corrosion in the internal surface of thermoelectric pipe walls, while using a compact network that requires much less effort in tuning parameters when compared to the handcrafted approach since the TCNN architecture is compact regarding the number of layers and connections. The observed results open up the possibility of using deep neural networks in real-time applications such as the automatic inspection of thermoelectric metal pipes.

This work proposes a classification approach for breast cancer histopathologic images (HI) that uses transfer learning to extract features from HI using an Inception-v3 CNN pre-trained with ImageNet dataset. We also use transfer learning on training a support vector machine (SVM) classifier on a tissue labeled colorectal cancer dataset aiming to filter the patches from a breast cancer HI and remove the irrelevant ones. We show that removing irrelevant patches before training a second SVM classifier, improves the accuracy for classifying malign and benign tumors on breast cancer images. We are able to improve the classification accuracy in 3.7% using the feature extraction transfer learning and an additional 0.7% using the irrelevant patch elimination. The proposed approach outperforms the state-of-the-art in three out of the four magnification factors of the breast cancer dataset.

In this work we present a comprehensive set of experiments aiming to perform music genre classification using learned and handcrafted features plus the fusion of them. Handcrafted features were obtained from the audio signal itself, lyrics, chords and spectrogram images extracted from the audio. The rationale behind this investigation is based on the assumption that one can find some complementarity between classifiers created from these different resources. The experimental protocol was conducted on the Brazilian Music Dataset using the artist filter restriction and they confirm the power of non-handcrafted features to perform audio classification tasks. The experimental results have shown a significant complementarity among the handcrafted features for which the evaluated fusion strategies allowed an improvement in the classification accuracy up to 4 percent points. On the other hand, the fusion of learned and handcrafted features provided similar accuracy than the best individual CNN (0.7815).

This paper presents a novel approach for image retrieval and pattern spotting in document image collections. The manual feature engineering is avoided by learning a similarity-based representation using a Siamese Neural Network trained on a previously prepared subset of image pairs from the ImageNet dataset. The learned representation is used to provide the similarity-based feature maps used to find relevant image candidates in the data collection given an image query. A robust experimental protocol based on the public Tobacco800 document image collection shows that the proposed method compares favor-ably against state-of-the-art document image retrieval methods, reaching 0.94 and 0.83 of mean average precision (mAP) for retrieval and pattern spotting (IoU=0.7), respectively. Besides, we have evaluated the proposed method considering feature maps of different sizes, showing the impact of reducing the number of features in the retrieval performance and time-consuming.

Facial expression recognition is a major problem in the domain of artificial intelligence. One of the best ways to solve this problem is the use of convolutional neural networks (CNNs). However, a large amount of data is required to train properly these networks but most of the datasets available for facial expression recognition are relatively small. A common way to circumvent the lack of data is to use CNNs trained on large datasets of different domains and fine-tuning the layers of such networks to the target domain. However, the fine-tuning process does not preserve the memory integrity as CNNs have the tendency to forget patterns they have learned. In this paper, we evaluate different strategies of fine-tuning a CNN with the aim of assessing the memory integrity of such strategies in a cross-dataset scenario. A CNN pre-trained on a source dataset is used as the baseline and four adaptation strategies have been evaluated: fine-tuning its fully connected layers; fine-tuning its last convolutional layer and its fully connected layers; retraining the CNN on a target dataset; and the fusion of the source and target datasets and retraining the CNN. Experimental results on four datasets have shown that the fusion of the source and the target datasets provides the best trade-off between accuracy and memory integrity.

A fine-grained plant leaf classification method based on the fusion of deep models is described. Complementary global and patch-based leaf features are combined at each hierarchical level (genus and species) by pre-trained CNNs. The deep models are adapted for plant recognition by using data augmentation techniques to face the problem of plant classes with very few samples for training in the available imbalanced dataset. Experimental results have shown that the proposed coarse-to-fine classification strategy is a very promising alternative to deal with the low inter-class and high intra-class variability inherent to the problem of plant identification. The proposed method was able to surpass other state-of-the-art approaches on the ImageCLEF 2015 plant recognition dataset in terms of average classification scores.

Recognition of forest species is a very challenging task thanks to the great intra-class variability. To cope with such a variability, we propose a multiple classifier system based on a two-level classification strategy and microscopic images. By using a divide-and-conquer approach, an image is first divided into several sub-images which are classified independently by each classifier. In a first fusion level, partial decisions for the sub-images are combined to generate a new partial decision for the original image. Then, the second fusion level combines all these new partial decisions to produce the final classification of the original image. To generate the pool of diverse classifiers, we used classical texture-based features as well as keypoint-based features. A series of experiments shows that the proposed strategy achieves compelling results. Compared to the best single classifier, a Support Vector Machine (SVM) trained with a keypoint based feature set, the divide-and-conquer strategy improves the recognition rate in about 4 and 6 percentage points in the first and second fusion levels, respectively. The best recognition rate achieved by this proposed method is 98.47%.

This work introduces a Brazilian Speech Database (BrSD), a novel dataset freely available created to support the development of speech-based recognition tasks. As far as we know, this is the first Portuguese language based database with these characteristics created and made available to the research community. We also describe experiments accomplished on BrSD exploring its different possibilities of classification tasks, i.e., age group and gender classification. We use four well-known acoustic features extracted directly from the audio signal and one texture-based feature extracted from a visual representation of the audio signal, the spectrogram. We considered three different classification scenarios: each feature individually, early fusion of the features, and late fusion of the features. Experiments were conducted using Support Vector Machine (SVM) and Multi-layer Perceptron (MLP) classifiers. The obtained results showed that SVM classifier achieved the best recognition rates both in early and late fusion scenarios. The best recognition rates achieved were 91.25%, 88.75%, and 80.25% for gender, age group, and age-gender classification tasks, respectively.

This work describes a new oracle based Dynamic Ensemble Selection (DES) method in which an Ensemble of Classifiers (EoC) is selected to predict the class of a given test instance (xt). The competence of each classifier is estimated on a local region (LR) of the feature space (Region of Competence - RoC) represented by the most promising k-nearest neighbors (or advisors) related to xt according to a discrimination index (D) originally proposed in the Item and Test Analysis (ITA) theory. The D value is used to better define the advisors of the RoC since they will suggest the classifiers (local oracles) to compose the EoC. A robust experimental protocol based on 30 classification problems and 20 replications have shown that the proposed DES compares favorably with 15 state-of-the-art dynamic selection methods and the combination of all classifiers in the pool.

This paper presents the Centrality Based Fusion (CBF) method for ensemble fusion which is based on the centrality measures in the context of complex network theory. Such a concept has been applied in Social Network Analysis to measure the importance of each person inside of a social network. We hypothesized that the centrality of each classifier inside of an ensemble represented as a complex network could be combined with accuracy to provide the weight for its decision during the ensemble fusion. The main idea is to derive the weight considering the classifier importance inside the ensemble network which reflects the classifiers' diversity. A robust experimental protocol based on 30 datasets has confirmed that the notion of prominence provided employing centrality measures is a promising strategy to weight the classifiers of an ensemble. When compared with 9 fusion methods of the literature, the proposed fusion method won in 189 out of 270 experiments (70%), lost in 61 cases (22.59%) and tied in 20 cases (7.41%).

This paper presents segmentation-free strategies for the recognition of handwritten numeral strings of unknown length. A synthetic dataset of touching numeral strings of sizes 2-, 3- and 4-digits was created to train end-to-end solutions based on Convolutional Neural Networks. A robust experimental protocol is used to show that the proposed segmentation-free methods may reach the state-of-the-art performance without suffering the heavy burden of over-segmentation based methods. In addition, they confirmed the importance of introducing contextual information in the design of end-to-end solutions, such as the proposed length classifier when recognizing numeral strings.

This paper proposes a novel approach for content based graphical object retrieval in document images. The challenge is to search for occurrences of a queried graphical objects in document images that can vary in terms of color, shape, texture and quality, increasing considerably the level of difficulty of the retrieval process. To that end, the manual feature engineering is avoided by learning the image representation for the retrieval task using a Convolutional Neural Network (CNN). However, such a representation should be as compact as possible to allow a fast document image retrieval and storage. Thus, a pretrained CNN model is used to cope with the lack of training data, which is fine tuned to achieve a compact yet discriminant representation of the graphical objects. From experiments conducted on the public Tobacco800 document image collection, we show that the proposed method compares favorably against state-of-the-art document image retrieval methods, reaching 0.72 of average precision (mAP). In addition, an increase of 4 percentage points in the average precision is observed using a compact deep representation in which the number of features is reduced by 16 times, thus allowing a reduction of 47% in terms of computation time by the image retrieval task.

This paper presents a new image metamorphosis method, ADAMM, especially developed to overcome the challenge in Forensic facial reconstruction of missing children and which does not require any type of training or supervision. ADAMM consists first in facial points of interest detection by computer vision, followed by automatic alignment between the corresponding segments in both source images, and finally the final image generation through its morphological median. Experiments were performed on a dataset with real images to compare the proposed method results with those from commercial programs based on traditional morphing methods. The experiments yielded a numeric gain in sharpness in all cases, ranging from +24% to +584:4%.

A novel two-stage age prediction approach with group-specific features is proposed in this paper. Aging process is captured through a highly discriminating feature representation that models shape, appearance, skin spots, and wrinkles. The two-stage method consists of a multi-class Support Vector Machine (SVM) to predict the age bracket while the final age prediction is carried out using Support Vector Regression (SVR). The novelty of our work is that the feature extraction is group-specific and can therefore be tailored to each age bracket in the specific age prediction step. The FG-NET Aging dataset was used to evaluate the proposed method and an impressive mean absolute error (MAE) of 3.98 was achieved. Our approach outperforms the current state-of-the-art while increasing the robustness to blur, expression and lighting variation with local phase features.

This paper proposes a classification approach in which monolithic and multiple classifier systems are combined in a cascading fashion. The rationale behind that is to deal with the existing trade-off between the need for increasing the accuracy, while reducing the complexity of the classification method. In other words, the idea is to offer an interesting strategy to conciliate the different levels of efforts necessary to deal with easy and hard patterns usually observed in a classification problem. The experimental results have shown that for some problems more than 90% of the instances can be processed in the first step of the cascade, saving efforts by avoiding the use of the second step in which a more complex classification method is used. It means that for some problems the reduction of the classification cost achieved more than 70% when compared to the use of an MCS. In addition to this interesting classification cost reduction, the cascade approach has shown to be able of improving the classification accuracy up to 15.19 percentage points.

This paper presents a multiple classifier system (MCS) to identify plants species based on the texture and shape features extracted from leaf images. A diverse pool of SVM and Neural Network classifiers is trained on four different feature sets, namely, Local Binary Pattern (LBP), Histogram of Gradients (HOG), Speed of Robust Features (SURF) and Zernike Moments (ZM). Then, a static classifier selection method is used to search for the ensembles that maximize the average classification score. Experimental results on ImageCLEF 2011 and 2012 datasets have shown that combining different kind of classifiers trained on shape and texture features is an effective strategy for the plant automatic identification. The MCS improves the identification performance in up to 28% relative to the monolithic approach. Furthermore, the proposed approach also compares favourably with the best results reported in the literature for those datasets.

The ever increasing data generation confronts both practitioners and researchers on handling massive and sequentially generated amounts of information, the so-called data streams. In this context, a lot of effort has been put on the extraction of useful patterns from streaming scenarios. Learning from data streams embeds a variety of problems, and by far, the most challenging is concept drift, i.e. changes in data distribution. In this paper, we focus on a specific type of drift uncommonly assessed in the literature: feature drifts. Feature drifts occur whenever a subset of features becomes, or ceases to be, relevant to the concept to be learned. We propose and review several feature drifting data stream generators and use them to benchmark state-of-the-art data stream classification algorithms and their combination with drift detectors. Results show that, although drift detectors enable slight quicker recovery to feature drifts, best results are obtained by Hoeffding Adaptive Tree, the only learner that performs dynamic feature selection as streams progress.

Extracting useful knowledge from data streams is problematic, mainly due to changes in their data distribution, a phenomenon named concept drift. Recently, studies have shown that most of existing algorithms for learning from data streams do not encompass techniques for a specific kind of drift: feature drifts. Feature drifts occur when features become, or cease to be, relevant to the learning task. In this paper, we propose an extension to the k-nearest neighbor classifier, so its distances' computations are weighted according to their current discriminative power. On our proposal, the discriminative power of features is given by entropy, which is swiftly computed over a sliding window. Empirical evidence shows that our approach is able to overcome several existing algorithms in accuracy and feature drift adaptation, while at the expense of bounded processing time and memory space.

This work describes the Dynse framework, which uses dynamic selection of classifiers to deal with concept drift. Basically, classifiers trained on new supervised batches available over time are add to a pool, from which is elected a custom ensemble for each test instance during the classification time. The Dynse framework is highly customizable, and can be adapted to use any method for dynamic selection of classifiers given a test instance. In this work we propose a default configuration for the framework which has provided promising results in a range of problems. The experimental results have shown that the proposed framework achieved the best average rank when considering all datasets, and outperformed the state-of-the-art in three of four tested datasets.

Different dynamic classifier selection techniques have been proposed in the literature to determine among diverse classifiers available in a pool which should be used to classify a test instance. The individual competence of each classifier in the pool is usually evaluated taking into account its accuracy on the neighborhood of the test instance in a validation dataset. In this work we investigate the possible contribution of considering during the classifier evaluation the use of features related to the problem complexity. Since usually the pool generation technique does not assure diversity, the idea is to consider diversity during the selection. Basically, we select a classifier trained in subset of data showing similar complexity than that observed in neighborhood of the test instance. We expect that this similarity in terms of complexity allow us to select a more competent classifier. Experiments on 30 classification problems representing different levels of difficulty have shown that the proposed selection method is comparable to well known dynamic selection strategies. When compared with other DS approaches it was able to win on 123 over 150 experiments. This promising results indicate that further investigation must be done to increase diversity in terms of data complexity during the process of pool generation.

This paper presents a syntactic approach for the authorship attribution to literary texts. To this end, syntactic features were used in verification and identification approaches. We used also, a classification method based on dissimilarity that has been successfully applied in cases of authorship attribution. In addition, to evaluating two models, the writer-dependent model and writer-independent, we tested approaches based on polytomy and dichotomy. We also tested the impact of the redundancy, by varying the number of references for verification and identification, for short and long texts. In order to meet these goals, we conducted experiments with four different databases, which are journalistic and literary texts in Portuguese and English languages. Through a series of experiments, we showed that the proposed approach was able to surpass the results reported in the literature, in both languages. In our experiments the proposed approach achieved results higher than 90% in both languages for verification and above 75% for the identification.

This paper presents a novel approach for bird species identification that relies on both visual features extracted from unconstrained bird images and acoustic features extracted from bird vocalizations. The Scale Invariant Feature Transform (SIFT) detects local features in bird images, which are then used to train a support vector machine classifier. The instances that are not classified with a certain degree of certainty are then rejected and reclassified using Mel-frequency cepstral coefficients (MFCCs) extracted from the bird songs if available. Experiments conducted on a dataset of 50 bird species that comprise images from the CUB200-2011 and audio samples from Xeno-Canto have shown that improvements between 1.2 and 15.7 percentage points are achieved when using an acoustic classifier to re-process the instances rejected by the visual classifier, depending on the rejection level.

SignWriting is a writing system for sign languages. It is based on visual symbols to represent the hand shapes, movements and facial expressions, among other elements. It has been adopted by more than 40 countries, but to ensure the social integration of the deaf community, writing systems based on sign languages should be properly incorporated into the Information Technology. This article reports our first efforts toward the implementation of an automatic reading system for SignWiring. This would allow converting the SignWriting script into text so that one can store, retrieve, and index information in an efficient way. In order to make this work possible, we have been collecting a database of hand configurations, which at the present moment sums up to 7,994 images divided into 103 classes of symbols. To classify such symbols, we have performed a comprehensive set of experiments using different features, classifiers, and combination strategies. The best result, 94.4% of recognition rate, was achieved by a Convolutional Neural Network.

This paper presents a k-nearest oracle-based dynamic ensemble selection method in which overall local accuracy (OLA) and local class accuracy (LCA) are combined into a twostep selection scheme. The OLA and LCA are computed on the neighborhood of the test pattern in a validation set to filter out the classifiers selected by the k-nearest oracles. The complementary information of OLA and LCA has shown to be an interesting alternative to approximate the classification performance to that estimated for the oracle of the initial pool of classifiers. The results were compared with the recognition rates of the majority voting of all classifiers in the initial pool, and also with the recognition rates of related classifier and ensemble selection methods which have inspired the proposed method and its variants. The proposed method achieved the best results on 5 out of 8 experiments using small and large datasets of different applications.

This paper addresses the problem of recognizing gestures which are captured using the Kinect sensor in a educational game devoted to the deaf community. Different strategies are evaluated to deal with the problem of having few samples for training. We have experimented a Leave One Out Training and Testing (LOOT) strategy and an HMM-based ensemble of classifiers. A dataset containing 181 videos of gestures related to nine signs commonly used in educational games is introduced, which is available for research purposes. The experimental results have shown that the proposed ensemble-based method is a promising strategy to deal with problems where few training samples are available.

In this paper we assess the use of textural de-scriptors for the problem of parking space detection. We focus our experiments on two descriptors (Local Binary Patterns and Local Phase Quantization) that have attracted a great deal of attention because of their outstanding performance in a number of applications. We show through a series of comprehensive experiments that both descriptors are able to achieve very low error rates on a database composed of 105,837 images of parking spaces. We also show that the combination of the diverse classifiers developed in this work can bring further improvement achieving an error rate of 0.16%. The results reached in this work compare favorably to other published methods.

This paper presents a method for the location of the identification code of train wagons from digital videos. For this purpose, thresholding and multithresholding based techniques combined with edge detectors, mathematical morphology and filtering by compacity factor are used to segment each video frame, locate the connected components, identify the text-block candidates, and filter them. This combination of different techniques makes the proposed approach promising to locate the identification code in wagons of different formats and presenting a bad maintenance. Experimental results on more than 2500 video frames containing 116 different codes have shown that the proposed method may reach about 85% of correct code location.

This paper presents a dynamic ensemble selection method for music genre classification which employs two pools of diverse classifiers. The pools of classifiers are created by using different features types extracted from three distinct segments of each music piece. From these initial pools of weak classifiers, ensembles of classifiers are dynamically selected for each test pattern using the k-nearest oracles method. The experiments compare the performance of different selection strategies on the Latin Music Database to those related to the use of best single classifier, and to the combination of all classifiers in the pool. It was possible to observe that the most promising selection strategy evaluated allows improving the classification accuracy from 63.71% to 70.31%.

In this paper we present a verification module that has as input the output provided by a word recognizer which is based on the segmentation-recognition paradigm. The word recognizer models words as the concatenation of character hidden Markov models (HMMs) and it provides at the output a list with the Top N best word hypotheses, including their likelihoods and the segmentation points of the words into sub words, which ideally should be characters. The verification module uses the segmentation points provided by the word recognizer for each word hypothesis to extract different features from each sub word. A classifier based on a multilayer perceptron neural network assigns a character class (A-Z) and estimates the a posteriori probability to each sub word that make up a word. Further, both the character class and the a posteriori probabilities are combined with the original output of the word recognizer to re-rank the word hypothesis into the Top N list. Experimental results show that the verification module improves the Top 1 recognition rate in 3.9% for an 85,092-word recognition task.

In this work we address the problem of forest species recognition which is a very challenging task and has several potential applications in the wood industry. The first contribution of this work is a database composed of 22 different species of the Brazilian flora that has been carefully labeled by expert in wood anatomy. In addition, in this work we demonstrate through a series of comprehensive experiments that color-based features are quite useful to increase the discrimination power for this kind of application. Last but not least, we propose a segmentation approach so that a wood can be locally processed to mitigate the intra-class variability featured in some classes. Such an approach also brings important contribution to improve the final performance in terms of classification.

Different strategies for combination of complementary features in an HMM-based method for handwritten character recognition are evaluated. In addition, a noise reduction method is proposed to deal with the negative impact of low probability symbols in the training database. New sequences of observations are generated based on the original ones, but considering a noise reduction process. The experimental results based on 52 classes of alphabetic characters and more than 23,000 samples have shown that the strategies proposed to optimize the HMM-based recognition method are very promising.

An effective way to improve a classification method's performance is to create ensembles of classifiers. Two elements are believed to be important in constructing an ensemble: (a) the performance of each individual classifier; and (b) diversity among the classifiers. Nevertheless, most works based on diversity suggest that there exists only weak correlation between classifier performance and ensemble accuracy. We propose compound diversity functions which combine the diversities with the performance of each individual classifier, and show that there is a strong correlation between the proposed functions and ensemble accuracy. Calculation of the correlations with different ensemble creation methods, different problems and different classification algorithms on 0.624 million ensembles suggests that most compound diversity functions are better than traditional diversity measures. The population-based Genetic Algorithm was used to search for the best ensembles on a handwritten numerals recognition problem and to evaluate 42.24 million ensembles. The statistical results indicate that compound diversity functions perform better than traditional diversity measures, and are helpful in selecting the best ensembles.

Hidden Markov Models (HMMs) have been shown to be useful in handwritten pattern recognition. However, owing to their fundamental structure, they have little resistance to unexpected noise among observation sequences. In other words, unexpected noise in a sequence might "break" the normal transmission of states for this sequence, making it unrecognizable to trained models. To resolve this problem, we propose a leave-one-out-training strategy, which will make the models more robust. We also propose a leave-one-out-testing method, which will compensate for some of the negative effects of this noise. The latter is actually an example of a system with a single classifier and multiple classifications. Compared with the 98.00 percent accuracy of the benchmark HMMs, the new system achieves a 98.88 percent accuracy rate on handwritten digits.

One of disadvantages of Hidden Markov Models (HMMs) is its low resistance to unexpected noises among observation sequences. Unexpected noises in a sequence usually “break” a sequence of observations, and then makes this sequence unrecognizable for trained models. We propose a new HMM training and testing scheme, which compensates some of the negative effects of such noises. We carried out experiment on handwritten digit recognition problem and the result suggests our proposal can be as effective as multiclassifier systems.

Ensemble of Classifiers (EoC) has been shown effective in improving the performance of single classifiers by combining their outputs. By using diverse data subsets to train classifiers, the ensemble creation methods can create diverse classifiers for the EoC. In this work, we propose a scheme to measure the data diversity directly from random subspaces and we explore the possibility of using the data diversity directly to select the best data subsets for the construction of the EoC. The applicability is tested on NIST SD19 handwritten numerals.

For handwritten pattern recognition, multiple classifier system has been shown to be useful in improving recognition rates. One of the most important issues to optimize a multiple classifier system is to select a group of adequate classifiers, known as ensemble of classifiers (EoC), from a pool of classifiers. Static selection schemes select an EoC for all test patterns, and dynamic selection schemes select different classifiers for different test patterns. Nevertheless, it has been shown that traditional dynamic selection does not give better performance than static selection. We propose four new dynamic selection schemes which explore the property of the oracle concept. The result suggests that the proposed schemes are apparently better than the static selection using the majority voting rule for combining classifiers.

This article presents a novel approach for detection of non-conventional events in videos scenes. This novel approach consists in analyzing in real-time video from a security camera to detect, segment and tracking objects in movement to further classify its movement as conventional or non-conventional. From each tracked object in the scene features such as position, speed, changes in directions and in the bounding box sizes are extracted. These features make up a feature vector. At the classification step, feature vectors generated from objects in movement in the scene are matched almost in real-time against reference feature vectors previously labeled which are stored in a database and an algorithm based on the instance-based learning paradigm is used to classify the object movement as conventional or non-conventional. Experimental results on video clips from two databases (Parking Lot and CAVIAR) have shown that the proposed approach is able to detect non-conventional events with accuracies between 77% and 82%.

Various methods for ensemble selection and classifier combination have been designed to optimize the results of ensembles of classifiers. Genetic algorithm (GA) which uses the diversity for the ensemble selection could be very time consuming. We propose compound diversity functions as objective functions for a faster and more effective GA searching. Classifiers selected by GA are combined by a proposed pairwise confusion matrix transformation, which offer strong performance boost for EoCs.

An ensemble of classifiers has been shown to be effective in improving classifier performance. Two elements are believed to be viable in constructing an ensemble: a) the classification accuracy of each individual classifier; and b) diversity among the classifiers. Nevertheless, most works based on diversity suggest that there exists only weak correlation between diversity and ensemble accuracy. We propose compound diversity functions which combine the diversities with the classification accuracy of each individual classifier, and show that with Random subspaces ensemble creation method, there is a strong correlation between the proposed functions and ensemble accuracy. The statistical result indicates that compound diversity functions perform better than traditional diversity measures.

In this paper we describe a synthetic database composed of 273,452 handwritten touching digits pairs to assess segmentation algorithms. It contains several different kinds of touching and it was generated by connecting 2,000 images of isolated digits extracted from the NIST SD19. In order to get a better insight on the proposed database and establish some parameters for further comparisons, we carried out experiments using four state-of-the-art segmentation algorithms.

This paper addresses the issue of class related reject thresholds for cascading classifier systems. It has been demonstrated in the literature that class related reject thresholds provide an error-reject tradeoff better than a single global threshold. In this work we argue that the error-reject tradeoff yielded by class-related reject thresholds can be further improved if a proper algorithm is used to find the thresholds. In light of this, we propose using a recently developed optimization algorithm called particle swarm optimization. It has been proved to be very effective in solving real valued global optimization problems. In order to show the benefits of such an algorithm, we have applied it to optimize the thresholds of a cascading classifier system devoted to recognize handwritten digits.

In this paper we address the issue of class-related reject thresholds for classification systems. It has been demonstrated in the literature that class related reject thresholds provide an error-reject tradeoff better than a single global threshold. In this work we argue that the error-reject tradeoff yielded by class related reject thresholds can be further improved if a proper algorithm is used to find the thresholds. In light of this, we propose using a recently developed optimization algorithm called particle swarm optimization. It has been proved to be very effective in solving real valued global optimization problems. In order to show the benefits of such an algorithm, we have applied it to optimize the thresholds of a cascading classifier system devoted to recognize handwritten digits.

In this paper we propose a parallel approach for the K-meansVector Quantization (VQ) algorithm used in a two-stageHidden Markov Model (HMM)-based system forrecognizing handwritten numeral strings. With thisparallel algorithm, based on the master/slave paradigm,we overcome two drawbacks of the sequential version: a)the time taken to create the codebook; and b) the amountof memory necessary to work with large trainingdatabases. Distributing the training samples over theslaves' local disks reduces the overhead associated withthe communication process. In addition, modelspredicting computation and communication time havebeen developed. These models are useful to predict theoptimal number of slaves taking into account the numberof training samples and codebook size.

This paper presents a segmentation algorithm based on feature selection in wavelet space. The aim is to automatically separate in postal envelopes the regions related to background, stamps, rubber stamps, and the address blocks. First, a typical image of a postal envelope is decomposed using Mallat algorithm and Haar basis. High frequency channel outputs are analyzed to locate salient points in order to separate the background. A statistical hypothesis test is taken to decide upon more consistent regions in order to clean out some noise left. The selected points are projected back to the original gray level image, where the evidence from the wavelet space is used to start a growing process to include the pixels more likely to belong to the regions of stamps, rubber stamps, and written area. Experiments are run using original postal envelopes from the Brazilian Post Office Agency, and here we report results on 440 images with many different layouts and backgrounds. 1.

The authors propose a handwritten numeral string recognition method composed of two HMM-based stages. The first stage uses an implicit segmentation strategy based on string contextual information to provide multiple segmentation-recognition paths. These paths are verified and re-ranked by using a verification stage based on a digit classifier. It allows the use of two sets of features and numeral models: one taking into account both segmentation and recognition aspects in an implicit segmentation based strategy, and another considering just recognition aspects of isolated digits. The two system stages are shown to be complementary in the sense that the verification stage is shown to be a promising idea to deal with the loss in terms of recognition performance brought about by the necessary tradeoff between segmentation and recognition carried out in the first system stage.