Small object detection is a challenge for computer vision models due to a shortage of image details, textures, and varying distances from the camera, resulting in objects of different scales and partial occlusion issues. In this paper, we present a new method for enhancing the robustness of image detection models using AUGMIX. Our approach involves applying various augmentations to the input images in a stochastic manner, resulting in a single output image after all transformations have been applied. In addition, we used the Jensen-Shannon loss to maintain a more stable model. In our experiments, we observed a decrease in the number of “no-object” detections, which refers to the detection of unrelated or background objects. The new approach was evaluated using the Deformable DETR, a model known for detecting small objects accurately, and compared to DETR and EfficientDet. We verified an improvement of at least 4.15% using the proposed technique and a more stable loss error.

Deep learning models often require large datasets to perform well from scratch. Transfer learning methods solve this issue by using a pre-trained source network to improve a target network training. Recent approaches involve using feature maps from the source network to guide the target network training. The latest transfer learning methods use meta-networks to enhance the knowledge transfer process. These meta-networks bridge the source and target networks, deciding which pairs of feature map layers and channels should be matched for optimal knowledge transfer. This paper improves this approach by using pixel-level information, in addition to layers and channels, for better knowledge transfer. Our experiments on multiple datasets show that the proposed approach outperforms previous baselines in scenarios with limited labels per class. The source code is available at https://github.com/lucasdelimanogueira/L2T-www.

Based on the need to digitalize identification documents per several institutions and companies, the segmentation task of textual information acquired importance. Commonly, Convolutional Neural Networks are applied to solve such problems. Vanilla convolution in deep learning models can provide sub-optimal performance in some learning tasks, such as semantic segmentation. Due to this, more specific proposals may be applied to better fit this context, such as self-calibrated convolutions, which consider unique characteristics of different feature maps. Considering this context, we propose a new fully convolutional network architecture based on U-Net for segmentation associated with self-calibrated convolution. We consider changing all vanilla convolution layers in this new neural network by self-calibrating convolutions. In this work, we evaluate our proposal in three different tasks of document segmentation. All experiments demonstrate an increase in the proposed neural network's segmentation performance compared with traditional U-Net in the analyzed context, but the inference time performance decreases.

The robustness and resilience of the deep learning models offer consistent and competitive results in real-world applications. Despite its adaptability, the training and adjustment of the hyperparameters still demand knowledge and time from the designer. This paper proposes a simple and effective approach based on the Greedy Randomized Adaptive Search Procedure (GRASP) algorithm that we adapt to optimize deep neural networks models. We evaluated the performance of the proposed approach using the models Deep Feedforward Neural Network (DFNN) and TabNet, considering the Tabu Search algorithm as a baseline in five tabular datasets. Both optimization algorithms showed high performance regarding the (i) quality of the best solution, (ii) convergence, and (iii) local search. However, the adapted GRASP approach showed better results, optimizing the deep models in all datasets with statistical significance.

In this article, we argue that the unsatisfactory out-of-distribution (OOD) detection performance of neural networks is mainly due to the SoftMax loss anisotropy and propensity to produce low entropy probability distributions in disagreement with the principle of maximum entropy. On the one hand, current OOD detection approaches usually do not directly fix the SoftMax loss drawbacks, but rather build techniques to circumvent it. Unfortunately, those methods usually produce undesired side effects (e.g., classification accuracy drop, additional hyperparameters, slower inferences, and collecting extra data). On the other hand, we propose replacing SoftMax loss with a novel loss function that does not suffer from the mentioned weaknesses. The proposed IsoMax loss is isotropic (exclusively distance-based) and provides high entropy posterior probability distributions. Replacing the SoftMax loss by IsoMax loss requires no model or training changes. Additionally, the models trained with IsoMax loss produce as fast and energy-efficient inferences as those trained using SoftMax loss. Moreover, no classification accuracy drop is observed. The proposed method does not rely on outlier/background data, hyperparameter tuning, temperature calibration, feature extraction, metric learning, adversarial training, ensemble procedures, or generative models. Our experiments showed that IsoMax loss works as a seamless SoftMax loss drop-in replacement that significantly improves neural networks' OOD detection performance. Hence, it may be used as a baseline OOD detection approach to be combined with current or future OOD detection techniques to achieve even higher results.

Out-of-distribution (OOD) detection approaches usually present special requirements (e.g., hyperparameter validation, collection of outlier data) and produce side effects (e.g., classification accuracy drop, slower energy-inefficient inferences). We argue that these issues are a consequence of the SoftMax loss anisotropy and disagreement with the maximum entropy principle. Thus, we propose the IsoMax loss and the entropic score. The seamless drop-in replacement of the SoftMax loss by IsoMax loss requires neither additional data collection nor hyperparameter validation. The trained models do not exhibit classification accuracy drop and produce fast energy-efficient inferences. Moreover, our experiments show that training neural networks with IsoMax loss significantly improves their OOD detection performance. The IsoMax loss exhibits state-of-the-art performance under the mentioned conditions (fast energy-efficient inference, no classification accuracy drop, no collection of outlier data, and no hyperparameter validation), which we call the seamless OOD detection task. In future work, current OOD detection methods may replace the SoftMax loss with the IsoMax loss to improve their performance on the commonly studied non-seamless OOD detection problem.

Given the actual volume and speed of financial transactions, financial fraud detection systems are constantly evolving based on new computational intelligence algorithms. Therefore, transaction monitoring and analysis prevent monetary losses caused by fraudsters. Since the fraud detection process is a labor-intensive task for human auditors given the huge amount of daily transactions processed by financial services information systems. Credit card is the financial product most explored in the financial fraud detection literature, while mobile money service is becoming a popular option for payments, fraud detection for such financial product has not yet been deeply explored. Therefore, it is interesting to optimize the auditing process and test new quantitative techniques, such as deep learning, to support human auditors before double-checking a suspicious transaction. Thus, we propose an integration of adversarial autoencoders and machine learning methods to perform an objective classification among three transaction types: regular, local, and global anomaly. The integration consists of using the autoencoder's generated latent vectors as features for the supervised learning algorithms. The experiments considered different latent vector space forms concerning their dimensionality and the clusters generated by a prior Gaussian mixture. The results show that some classifiers may accept latent characteristics well, getting better or similar performance when using all the original characteristics.

Microorganisms that cause infectious diseases are defined as pathogens, as they multiply and cause tissue damage. All microorganisms isolated in culture from a location on the body should be considered potential pathogens. The infectious processes demonstrate physiological responses to the multiplication invasion of the aggressor microorganism. The disease’s development is influenced by the patient’s general health, defense mechanisms, and previous contact with the offending agent. When an infectious disease is suspected, cultures should be performed. This article uses an electronic nose to collect and analyze volatile organic compounds VOCs expelled by colonies of microorganisms. We propose signature identification of these colony odors from microorganisms using InceptionTime. The InceptionTime model is a set of models of the deep convolutional neural network, inspired by the Inception-v4 architecture. The results were excellent, with an average accuracy in the test set above 98%. The aim of our research is to propose a faster, cheaper and more accurate method of detecting these pathogens and the encouraging results of this stage encourage further research.

Cyber-attacks cyber-physical systems (CPSs) can lead to sensing and actuation misbehavior, severe damages to physical objects, and safety risks. Machine learning algorithms have been proposed for hindering cyber-attacks on CPSs, but the absence of labeled data from novel attacks makes their detection quite challenging. In this context, generative adversarial networks (GANs) are a promising unsupervised approach to detect cyber-attacks by implicitly modeling the system. However, the detection of cyber-attacks on CPSs has strict latency requirements, since the attacks need to be stopped before the system is compromised. In this article, we propose FID-GAN, a novel fog-based, unsupervised intrusion detection system (IDS) for CPSs using GANs. The IDS is proposed for a fog architecture, which brings computation resources closer to the end nodes and thus contributes to meeting low-latency requirements. In order to achieve higher detection rates, the proposed architecture computes a reconstruction loss based on the reconstruction of data samples mapped to the latent space. Other works that follow a similar approach struggle with the time required to compute the reconstruction loss, which renders them impractical for latency constrained applications. We address this problem by training an encoder that accelerates the reconstruction loss computation. Experiments show that the proposed solution achieves higher detection rates and is at least 5.5 times faster than a baseline approach in the three studied data sets.

Most of the automatic fire alarm systems detect the fire presence through sensors like thermal, smoke, or flame. One of the new approaches to the problem is the use of images to perform the detection. The image approach is promising since it does not need specific sensors and can be easily embedded in different devices. However, besides the high performance, the computational cost of the used deep learning methods is a challenge to their deployment in portable devices. In this work, we propose a new deep learning architecture that requires fewer floating-point operations (flops) for fire recognition. Additionally, we propose a portable approach for fire recognition and the use of modern techniques such as inverted residual block, convolutions like depth-wise, and octave, to reduce the model's computational cost. The experiments show that our model keeps high accuracy while substantially reducing the number of parameters and flops. One of our models presents 71% fewer parameters than FireNet, while still presenting competitive accuracy and AUROC performance. The proposed methods are evaluated on FireNet and FiSmo datasets. The obtained results are promising for the implementation of the model in a mobile device, considering the reduced number of flops and parameters acquired.

The detection of objects considering a 6DoF pose is a common requirement to build virtual and augmented reality applications. It is usually a complex task which requires real-time processing and high precision results for adequate user experience. Recently, different deep learning techniques have been proposed to detect objects in 6DoF in RGB images. However, they rely on high complexity networks, requiring a computational power that prevents them from working on mobile devices. In this paper, we propose an approach to reduce the complexity of 6DoF detection networks while maintaining accuracy. We used Knowledge Distillation to teach portables Convolutional Neural Networks (CNN) to learn from a real-time 6DoF detection CNN. The proposed method allows real-time applications using only RGB images while decreasing the hardware requirements. We used the LINEMOD dataset to evaluate the proposed method, and the experimental results show that the proposed method reduces the memory requirement by almost 99% in comparison to the original architecture with the cost of reducing half the accuracy in one of the metrics. Code is available at https://github.com/heitorcfelix/singleshot6Dpose.

The Know Your Customer (KYC) and Anti Money Laundering (AML) are worldwide practices to online customer identification based on personal identification documents, similarity and liveness checking, and proof of address. To answer the basic regulation question: are you whom you say you are? The customer needs to upload valid identification documents (ID). This task imposes some computational challenges since these documents are diverse, may present different and complex backgrounds, some occlusion, partial rotation, poor quality, or damage. Advanced text and document segmentation algorithms were used to process the ID images. In this context, we investigated a method based on U-Net to detect the document edges and text regions in ID images. Besides the promising results on image segmentation, the U-Net based approach is computationally expensive for a real application, since the image segmentation is a customer device task. We propose a model optimization based on Octave Convolutions to qualify the method to situations where storage, processing, and time resources are limited, such as in mobile and robotic applications. We conducted the evaluation experiments in two new datasets CDPhotoDataset and DTDDataset, which are composed of real ID images of Brazilian documents. Our results showed that the proposed models are efficient to document segmentation tasks and portable.

Relation extraction (RE) consists in categorizing the relationship between entities in a sentence. A recent paradigm to develop relation extractors is Distant Supervision (DS), which allows the automatic creation of new datasets by taking an alignment between a text corpus and a Knowledge Base (KB). KBs can sometimes also provide additional information to the RE task. One of the methods that adopt this strategy is the RESIDE model, which proposes a distantly-supervised neural relation extraction using side information from KBs. Considering that this method outperformed state-of-the-art baselines, in this paper, we propose a related approach to RESIDE also using additional side information, but simplifying the sentence encoding with BERT embeddings. Through experiments, we show the effectiveness of the proposed method in Google Distant Supervision and Riedel datasets concerning the BGWA and RESIDE baseline methods. Although Area Under the Curve is decreased because of unbalanced datasets, P@N results have shown that the use of BERT as sentence encoding allows superior performance to baseline methods.

Speaker Recognition and Speaker Identification are challenging tasks with essential applications such as automation, authentication, and security. Deep learning approaches like SincNet and AM-SincNet presented great results on these tasks. The promising performance took these models to real-world applications that becoming fundamentally end-user driven and mostly mobile. The mobile computation requires applications with reduced storage size, non-processing and memory intensive and efficient energy-consuming. The deep learning approaches, in contrast, usually are energy expensive, demanding storage, processing power, and memory. To address this demand, we propose a portable model called Additive Margin MobileNet1D (AM-MobileNet1D) to Speaker Identification on mobile devices. We evaluated the proposed approach on TIMIT and MIT datasets obtaining equivalent or better performances concerning the baseline methods. Additionally, the proposed model takes only 11.6 megabytes on disk storage against 91.2 from SincNet and AM-SincNet architectures, making the model seven times faster, with eight times fewer parameters.

In medicine courses, improve the skills of surgery students is an essential part of the program. For training the surgeon residents the institutions normally using a standard checklist to evaluate the student evolution. However, the checklist evaluation is susceptible to evaluator bias, inter-evaluator variability, besides being time-consuming. The automation of this process is an important evolution in medical training. An alternative to the instructor checklist is capturing and evaluation of kinematic data regarding the surgical motion. We propose a novel CNN architecture for automated robot-assisted skill assessment. We explore the use of the SELU activation function and a global mixed pooling approach based on the average and max-pooling layers. Finally, we examine two types of convolutional layers: real-value and quaternion-valued. The results suggest that our model presents a higher average accuracy across the three surgical subtasks of the JIGSAWS dataset.

Cardiovascular diseases are one of the most common causes of death in the world. Prevention, knowledge of previous cases in the family, and early detection is the best strategy to reduce this fact. Different machine learning approaches to automatic diagnostic are being proposed to this task. As in most health problems, the imbalance between examples and classes is predominant in this problem and affects the performance of the automated solution. In this paper, we address the classification of heartbeats images in different cardiovascular diseases. We propose a two-dimensional Convolutional Neural Network for classification after using a InfoGAN architecture for generating synthetic images to unbalanced classes. We call this proposal Adversarial Oversampling and compare it with the classical oversampling methods as SMOTE, ADASYN, and Random Oversampling. The results show that the proposed approach improves the classifier performance for the minority classes without harming the performance in the balanced classes.

Speaker Recognition is a challenging task with essential applications such as authentication, automation, and security. The SincNet is a new deep learning based model which has produced promising results to tackle the mentioned task. To train deep learning systems, the loss function is essential to the network performance. The Softmax loss function is a widely used function in deep learning methods, but it is not the best choice for all kind of problems. For distance-based problems, one new Softmax based loss function called Additive Margin Softmax (AM-Softmax) is proving to be a better choice than the traditional Softmax. The AM-Softmax introduces a margin of separation between the classes that forces the samples from the same class to be closer to each other and also maximizes the distance between classes. In this paper, we propose a new approach for speaker recognition systems called AM-SincNet, which is based on the SincNet but uses an improved AM-Softmax layer. The proposed method is evaluated in the TIMIT dataset and obtained an improvement of approximately 40% in the Frame Error Rate when compared to SincNet.

Inductive transfer learning is widespread in computer vision applications. However, in natural language processing (NLP) applications is still an under-explored area. The most common transfer learning method in NLP is the use of pre-trained word embeddings. The Universal Language Model Fine-Tuning (ULMFiT) is a recent approach which proposes to train a language model and transfer its knowledge to a final classifier. During the classification step, ULMFiT uses a max and average pooling layer to select the useful information of an embedding sequence. We propose to replace max and average pooling layers with a soft attention mechanism. The goal is to learn the most important information of the embedding sequence rather than assuming that they are max and average values. We evaluate the proposed approach in six datasets and achieve the best performance in all of them against literature approaches.

Semantic segmentation is an essential task in computer vision that aims to label each image pixel. Several of the actual best approaches in this context are based on deep neural networks. For example, SegNet is a deep encoder-decoder architecture approach whose results were disruptive because it is fast and performs well. However, this architecture fails to fine-delineating the edges between the objects of interest in the image. We propose some modifications in the SegNet-Basic architecture by using a post-processing segmentation layer (using Conditional Random Fields) and by transferring high resolution features combined to the decoder network. The proposed method was evaluated in the dataset CamVid. Moreover, it was compared with important variants of SegNet and showed to be able to improve the overall accuracy of SegNet-Basic by up to 17.5%.

Current research in the field of convolutional neural networks usually focuses on improving network accuracy, regardless of the network size and inference time. In this paper, we investigate the effects of storage space reduction in SqueezeNet as it relates to inference time when processing single test samples. In order to reduce the storage space, we suggest adjusting SqueezeNet's Fire Modules to include Depthwise Separable Convolutions (DSC). The resulting network, referred to as SqueezeNet-DSC, is compared to different convolutional neural networks such as MobileNet, AlexNet, VGG19, and the original SqueezeNet itself. When analyzing the models, we consider accuracy, the number of parameters, parameter storage size and processing time of a single test sample on CIFAR-10 and CIFAR-100 databases. The SqueezeNet-DSC exhibited a considerable size reduction (37% the size of SqueezeNet), while experiencing a loss in network accuracy of 1,07% in CIFAR-10 and 3,06% in top 1 CIFAR-100.