In deep learning, understanding the decision-making processes of complex models is essential for advancing interpretability and trust in artificial intelligence systems. We introduce Causal Relational Attribution Graph (C-RAG), designed to deliver comprehensive, multi-perspective explanations of convolutional neural networks (CNNs) via a graph representation. C-RAG integrates gradient-based local attribution with global feature importance by constructing a graph-based representation that captures hierarchical feature inter-dependencies. In this framework, feature clusters are represented as graph nodes, and their interactions are quantified through combined localized and global attribution metrics, ensuring interpretable insights into model behavior. We evaluate C-RAG across diverse benchmark datasets (ImageNet, CIFAR-10, MNIST) and CNN architectures (ResNet18, VGG19, DenseNet201, LeNet), demonstrating significant advancements over state-of-the-art explainability methods in faithfulness, robustness, and computational efficiency. The proposed approach facilitates accurate spatial feature localization, robust dependency mapping, and efficient explanation generation, making it a valuable tool for critical applications such as medical imaging and autonomous systems. We provide a novel graph-based explainability framework, which bridges the gap between local and global interpretability, C-RAG addresses key limitations in existing methods, establishing a robust foundation for explainable AI in computer vision.

Abstract Deep Neural Networks (DNNs) have achieved state-of-the-art performance across various domains, yet their widespread adoption remains constrained by substantial computational and memory demands. While model pruning has emerged as a compelling strategy to address these challenges, existing methods often suffer from critical shortcomings: (1) non-selective neuron pruning which overlooks neuron activation dynamics, leading to the removal of critical neurons; (2) an inability to account for task-specific neuron importance, which causes accuracy degradation; and (3) failure to mitigate redundancy in neuron activations, resulting in suboptimal compression. In this work, we propose a novel Neuron Efficiency Metric (NEM), which integrates three key components—Neuron Activation Rate (NAR), Class-specific Activation Strength (CAS), and Neuron Overlap Index (NOI)—to address these limitations and guide a more effective pruning process. By iteratively evaluating the relevance of each neuron along these axes, NEM ensures selective and structured pruning that minimizes the retention of redundant or irrelevant neurons while preserving task-critical activations. The proposed method is tested on modern architectures using benchmark datasets such as MNIST and CIFAR-10, demonstrating a significant reduction in computational complexity while maintaining or even improving model performance. The results reveal that NEM achieves a higher degree of compression with minimal accuracy loss compared to conventional techniques.

Abstract Convolutional neural networks have become prominent machine learning models, particularly in the realm of computer vision, due to their ability to predict and extract robust features from raw image data. CNNs, similar to other neural network models, undergo training via backpropagation, an iterative technique. However, the backpropagation algorithm has notable challenges, including slow convergence, susceptibility to local minima, and hypersensitivity to learning rates. These challenges not only impact the model’s accuracy but also make the training process computationally intensive. To address these limitations, We introduce a novel approach that trains the CNN classifier using a non-iterative learning method. The proposed approach involves automatic extraction of pertinent features from the raw-data, followed by the application of Gram–Schmidt process to decompose the feature matrix and determine classifier’s weights. The proposed method has shown enhanced predictive accuracy over state-of-the-art models when evaluated on two benchmark datasets, MNIST and CIFAR-10. The extensive experimentation using most cited pre-trained experiments validate the effectiveness of our proposed method.

Incorporating edge and cloud computing with robotics provides extended options for robots to perform real-time sensing and actuation operations in various cyber–physical systems (CPSs), including smart farms. Such systems are prone to uncertain failures triggered by mechanical disruptions. Consequently, the overall system performance degrades, primarily when location-specific tasks are already assigned to a faulty robot and require immediate recovery. Using edge and cloud computing resources is not always feasible due to communication and latency constraints. Therefore, this article exclusively focuses on harnessing the mobility of robots to support the computation tasks affected by uncertain failures of previously assigned robots and ensure faster resiliency management by relocating active robots near task sources. The proposed mobility-as-a-resilience-service (MaaRS) is formulated using a Markov decision process (MDP). Later, an edge server proximal to the robots is trained using deep reinforcement learning (DRL) to assign tasks among the robots. Specifically, a multiple deep$Q$-network (MDQN)-based dynamic task allocation mechanism is proposed to converge to a solution exploring reward uncertainties with the best exploitation. Numerical evaluation using Python and TensorFlow validates the effectiveness of the proposed approach compared to other benchmarks.

In this paper we present a novel application of detecting fruit picker activities based on time series data generated from wearable sensors. During harvesting, fruit pickers pick fruit into wearable bags and empty these bags into harvesting bins located in the orchard. Once full, these bins are quickly transported to a cooled pack house to improve the shelf life of picked fruits. For farmers and managers, the knowledge of when a picker bag is emptied is important for managing harvesting bins more effectively to minimise the time the picked fruit is left out in the heat (resulting in reduced shelf life). We propose a means to detect these bag-emptying events using human activity recognition with wearable sensors and machine learning methods. We develop a semi-supervised approach to labelling the data. A feature-based machine learning ensemble model and a deep recurrent convolutional neural network are developed and tested on a real-world dataset. When compared, the neural network achieves 86% detection accuracy.

Deep-learning models are popular machine learning models that have gained their popularity in various fields of computer vision, natural language processing etc, due to their excellent predictive accuracy and ability to automatically extract good features from raw input data. Though these models have these significant advantages, most deep learning models are notoriously black-box models. The features learned by these models are not explainable. The need to explain these predictions is important in many high stakes fields. In this work, we propose a method that uses convolutional layers like in Convolutional Neural Networks (CNNs) that automatically learns features from raw data. In our proposed method we extract feature representation that is unique to each target class in the given data. Visualizations of this unique representation explain the feature learned by the model for a specific target class. To classify an input data we use two approaches. The first approach uses the similarity scores to match the features learned for a particular instance with the extracted target class-specific features to make prediction. The second approach uses an intrinsically explainable model like logistic regression to make predictions. Thus our proposed method in addition to having good predictive accuracy, identifies features that are class-specific. The proposed method achieved an accuracy of 99.31 % on MNIST data set, 90.45% on CIFAR-10 data set and 96.45% accuracy on Assira data set.

To develop safe automated driving functions, knowing road-user’s lane change behaviour is critical. This detection problem may depend on multiple aspects such as road conditions, location, and weather. To understand the effect of these situational variables, this work introduces a lane change detection algorithm and assessed its performance under various light conditions, road types and weather conditions. The algorithm has developed in L3Pilot: a large-scale European pilot project on level 3 automation. In the current study, the algorithm was tested with data from a Dutch Field Operational Test on SAE Level 2 systems. The algorithm was assessed against manually annotated video recordings. New is that validation was executed with Dutch Field Operational Test data of different participants and vehicles, distinguishing three situational variables factors. These were day vs night, motorways vs trunk roads and dry vs rain. A bootstrap procedure was used to assess the statistical significance of differences among the conditions. The conclusion is that the algorithm in combination with the provided data is effective in detecting lane changes when data is collected on a sample of Dutch motorways, irrespective of light and precipitation conditions. However, the quality of the sensor signals was worse on trunk roads, yielding significantly worse lane change detection performance (for all light and precipitation conditions).

Reliable forecasting of power output from solar Photovoltaic (PV) systems is indispensable for successful penetration of increasingly high solar PV capacity into the structure of power systems. Forecasting is as equally important for efficient demand management as solar power output is variable in nature. Farm level forecasting is the primary focus of the current literature and various methods are available. However, the grid operators need to have the forecasts at aggregated regional level (e.g. state level) for decision making. Aggregated system level forecasting has received very limited attention in literature with only a very few methods so far. In this paper, we consider the task of forecasting PV power production at aggregated regional level (at state level in Australia) and present a data-driven approach. To develop the prediction models, we investigate three widely used machine learning algorithms for farm level solar power forecasting and two state-of-the-art deep learning algorithms. As the inputs to the prediction algorithms, we study two different feature sets based on the lagged power data and their descriptive statistics. An evaluation of the presented approach using PV power generation data from Australian Energy Market Operator (AEMO) demonstrates up to 6.30% improvement in forecast skill over the baseline models.

Mobile malware detection is inherently an imbalanced data problem since the number of benign applications in the market is far greater than the number of malicious applications. Existing methods to handle imbalanced data, such as synthetic minority over-sampling, do not translate well into this domain since mobile malware detection generally deals with binary features and these methods are designed for continuous features. Also, methods adapted for categorical features cannot be applied here since random modifications of features can result in invalid sample generation. In this work, we propose a novel technique for generating synthetic samples for mobile malware detection with imbalanced data. Our proposed method adds new data points in the sample space by generating synthetic malware samples which also preserves the original functionality of the malicious apps. Experiments show that the proposed approach outperforms existing techniques in terms of precision, recall, F1score, and AUC. This study will be useful in building deep neural network-based systems to handle imbalanced data for mobile malware detection.

The fourth industrial revolution introduces an ideal opportunity for inclusion of cloud-enabled robots in a factory environment to improve productivity and reduce human intervention. For this novel paradigm, task offloading plays a critical role in leveraging computation support from resourceful cloud infrastructure. In particular, network connectivity and on-demand mobility of robot significantly influence the task offloading decision-making and vice-versa. While current studies in the literature separately consider mobility or communication aspects to accommodate offloading, ours is the first approach to integrate these three interdependent factors together in order to formulate a joint optimization problem for the proposed oil factory maintenance application. A modified genetic algorithm scheme is then developed to solve the problem with a novel 3-layer decision: task offloading, path planning, and access point selection. Simulation results and comparison with existing techniques suggest that communication-aware and mobility-driven offloading in industrial scenario leads to superior system performance and minimum consumption of resources.

A novel approach to deterministic modelling of diurnal water quality parameters in aquaculture prawn ponds is presented. The purpose is to provide assistance to prawn pond farmers in monitoring pond water quality with limited data. Obtaining sufficient water quality data is generally a challenge in commercial prawn farming applications. Farmers can sustain large losses in their crop if water quality is not well managed. The model presented provides a means for modelling and forecasting various water quality parameters. It is inspired by data dynamics and does not rely on physical ecosystem modelling. The model is constructed within the Bayesian filtering framework. The Kalman filter and the unscented Kalman filer are applied for inference. The results demonstrate generalisability to both variables and environments. The ability for short term forecasting with mean absolute percentage errors between 0.5% and 11% is demonstrated.

Task offloading opens a gateway for robotic applications to leverage computation support from the cloud infrastructure. It exploits a trade-off between the robot's and cloud's processing capabilities, and the communication between the two entities plays a critical role in making these decisions. Two major factors that significantly influence communication are network connectivity (bandwidth) and mobility of the robot. We integrate these two factors with the offloading decisions to formulate an optimization problem. Our objective in this paper is to improve the quality of service (QoS) for a 25 node application taskflow, known as direct acyclic graph. We propose a genetic algorithm based approach to solve the optimization problem which performs a novel three-layer decision making: (i) whether to offload a task or not, (ii) path planning to reach a desired location for offloading/local execution, (iii) select access point to associate with for offloading. We simulate for a smart city scenario consisting of 36-cell workspace with obstacles and compare the offloading results with a well- established fixed movement offloading method. The outcomes of our study suggest that motion and connectivity aware offloading leads to more efficient performance in terms of improved QoS and minimum consumption of resources, i.e., energy, time or distance.

The problem of learning long-term dependencies in sequences using Recurrent Neural Networks (RNNs) is still a major challenge. Recent methods have been suggested to solve this problem by constraining the transition matrix to be unitary during training which ensures that its norm is equal to one and prevents exploding gradients. These methods either have limited expressiveness or scale poorly with the size of the network when compared with the simple RNN case, especially when using stochastic gradient descent with a small mini-batch size. Our contributions are as follows; we first show that constraining the transition matrix to be unitary is a special case of an orthogonal constraint. Then we present a new parametrisation of the transition matrix which allows efficient training of an RNN while ensuring that the matrix is always orthogonal. Our results show that the orthogonal constraint on the transition matrix applied through our parametrisation gives similar benefits to the unitary constraint, without the time complexity limitations.

Accurate prediction of wind power is of vital importance for demand management. In this paper, we adopt a cluster-based ensemble framework to predict wind power. Natural groups/clusters exist in datasets and learning algorithms benefit from group/cluster wise learning — a philosophy that is not well explored for wind power prediction. The research presented in this paper investigates this philosophy to predict wind power by using an ensemble of regression models on natural clusters within wind data. We have conducted a series of experiments on a large number of locations across Australia and analyzed the existence of clusters within wind data, suitability of linear and nonlinear regression models for the proposed framework, and how well the cluster-based ensemble performs against the situation when no clustering is done. Experimental results demonstrate prediction improvement as high as 17.94% through the usage of the cluster-based ensemble regression algorithm.

Cloud robotics is an emerging paradigm that enables autonomous robotic agents to communicate and collaborate with cloud computing infrastructures. It further complements Internet of Things (IoT) to improve the performance of smart city applications. By offloading heavy data- intensive computation to the ubiquitous cloud, quality of service (QoS) guarantee can be ensured. Unlike their mobile counterpart, the robots have unique characteristics of mobility, skill- learning, data collection and decision-making capabilities, which makes offloading decisions significantly complex. This paper proposes a generic cloud robotics framework to realize smart city vision while taking into consideration its various complexities. Specifically, task offloading is formulated as a constrained optimization problem capable of handling Direct Acyclic Graph (DAG) known as task flow. Given the constraints, a genetic algorithm (GA) based scheme is further developed to solve the problem. The performance of the algorithm is verified by evaluating the results via three benchmarks. To the best of our knowledge, this is one of the first attempts of task offloading approach for smart city applications of cloud robotics.

A study is presented comparing the effectiveness of unsupervised feature representations with handcrafted features for cattle behaviour classification. Precision management of cattle requires the interaction of individual animals to be continuously monitored on the farm. Consequently, classifiers are trained to infer the behaviour of the animals using the observations from the sensors that are fitted upon them. Historically, domain knowledge drives the generation of features for cattle behaviour classifiers. When new behaviours are introduced into the system, however, it is often necessary to modify the feature set; this requires additional design and more data. Autoencoders, on the other hand, can skip this design step by learning a common, unsupervised feature representation for training. Whilst stacked autoencoders successfully represent structured data including speech, language and images, deep networks have not been used to model cattle motion. Hence, we investigate using a stacked autoencoder to learn a feature representation for cattle behaviour classification. Experimental results demonstrate that the autoencoder features perform reasonably well in comparison to the statistical features that are selected using prior knowledge of behaviour motion.

BACKGROUND: Despite ongoing reduction in genotyping costs, genomic studies involving large numbers of species with low economic value (such as Black Tiger prawns) remain cost prohibitive. In this scenario DNA pooling is an attractive option to reduce genotyping costs. However, genotyping of pooled samples comprising DNA from many individuals is challenging due to the presence of errors that exceed the allele frequency quantisation size and therefore cannot be simply corrected by clustering techniques. The solution to the calibration problem is a correction to the allele frequency to mitigate errors incurred in the measurement process. We highlight the limitations of the existing calibration solutions such as the fact they impose assumptions on the variation between allele frequencies 0, 0.5, and 1.0, and address a limited set of error types. We propose a novel machine learning method to address the limitations identified. RESULTS: The approach is tested on SNPs genotyped with the Sequenom iPLEX platform and compared to existing state of the art calibration methods. The new method is capable of reducing the mean square error in allele frequency to half that achievable with existing approaches. Furthermore for the first time we demonstrate the importance of carefully considering the choice of training data when using calibration approaches built from pooled data. CONCLUSION: This paper demonstrates that improvements in pooled allele frequency estimates result if the genotyping platform is characterised at allele frequencies other than the homozygous and heterozygous cases. Techniques capable of incorporating such information are described along with aspects of implementation.

Due to frost and insufficient exposure to sunlight, some grape bunches remain undeveloped during harvesting. For automation of harvesting, it is required to automatically identify the mature grape bunches. This paper presents a sequence of image processing and computational intelligence methods to identify mature grape bunches. It's a two-step process where in the first step the grape bunches are separated from the background of an image and in the second step the grape bunch is classified into mature and undeveloped group. We achieved 96.88% accuracy on the images obtained from a strip of vineyard in Cambridge, Tasmania.

New sensor streams are being generated at a rapidly increasing rate. The sources of these streams are a diverse set of networked sensors, diverse both in sensing hardware and sensing modality. Machine learning algorithms are ideally placed to develop generalized methods for stream analysis. One exemplar problem is the detection and analysis of periodic structure within these streams. Our contribution is the proposal of a new machine learning framework that (i) classifies a signal as periodic or aperiodic, (ii) further analyses the signal to find periodic structure using a neural network, and (iii) groups the motifs in the periodic signals using a modified Self Organising Map algorithm. We also demonstrate the framework using data generated by an Oyster heart rate sensor. We find that the generalized approach our classifier improves the detection of signal periods by reducing the number of functions classified as periodic from 11% to 9%; however, most benefit occurs for period calculation with the number of erroneously calculated periods reducing from 14% to 4%.

The combination of low density SNP arrays and DNA pooling is a fast and cost effective approach to genotyping that opens up basic genomics to a range of new applications and studies. However we have identified significant limitations in the existing approach to calculating allele frequencies with DNA pooling. These limitations include a reduced ability to deal with SNP to SNP variation via the standard interpolation method. Our contribution is a new hierarchical learning framework which resolves these drawbacks. The framework involves a hierarchy of two greedily trained layers of learners. The first layer learns the bias of each SNP then applies a calibration to reduce SNP bias by mapping into a common coordinate system across all SNPs. The second layer learns an allele frequency function exploiting the global SNP data. A range of algorithms have been applied including linear regression, neural network and support vector regression. The framework has been tested on pooled samples of Black Tiger prawns that have been genotyped with low density Sequenom iPLEX panels. Analysis of pooled samples and the corresponding individually genotyped SNP samples indicate the pooling approach introduces an allele frequency RMS error of 0.12. The existing calibration approach corrects ~14% of the error. Our hierarchical approach is 4.5 times as effective by correcting for ~64% of the introduced error. This is a significant reduction and has the potential to enable genetic studies previously not possible due to allele frequency error. Although testing so far is limited to low density SNP arrays the approach was developed to generalize to other SNP genotyping technologies.

In this paper we present a novel application of genetic algorithm for obtaining mineral domains. An important step in mineral exploration is to identify potential underground areas for mining purposes. Cylindrical cores of rock are extracted from the subsurface using diamond drilling and data is logged on the cores leading to geo-data sets. These geo-datasets are analysed to obtain domains i.e. regions of unique characteristics. We have developed a genetic algorithm based approach to identify mineral domains from geo-data sets. A special feature of the proposed approach is that it can separate the domains by preserving their spatial continuity that conventional point-based and window-based clustering fail to capture. We have evaluated the proposed method on gold-assay and geochemical datasets obtained from a gold mine in Western Australia. Experimental results demonstrate that the proposed method can identify domains accurately and also outperform point-based and window-based clustering.

In this paper, we present an application of Multi-Objective Evolutionary Algorithm (MOEA) for generating cluster oriented ensemble classifier. In our recently developed Non-Uniform Layered Cluster Oriented Ensemble Classifier (NULCOEC), the data set is partitioned into a variable number of clusters at different layers. Base classifiers are then trained on the clusters at different layers. The performance of NULCOEC is a function of the vector (layers, clusters) and the research presented in this paper investigates the implication of applying MOEA to generate NULCOEC. Accuracy and diversity of the ensemble classifier is expressed as a function of layers and clusters. A MOEA then searches for the combination of layers and clusters to obtain the non-dominated set of (accuracy, diversity). We have also obtained the results of single objective optimization (i.e. optimizing either accuracy or diversity) and compared them with the results of MOEA. The results show that the MOEA can improve the performance of ensemble classifier.

Sensor faults or communication errors can cause certain sensor readings to become unavailable for prediction purposes. In this paper we evaluate the performance of imputation techniques and techniques that ignore the missing values, in scenarios: (i) when values are missing only during prediction phase, and (ii) when values are missing during both the induction and prediction phase. We also investigated the influence of different scales of missingness on the performance of these treatments. The results can be used as a guideline to facilitate the choice of different missing value treatments under different circumstances.

Abstract This paper presents cluster‐based ensemble classifier – an approach toward generating ensemble of classifiers using multiple clusters within classified data. Clustering is incorporated to partition data set into multiple clusters of highly correlated data that are difficult to separate otherwise and different base classifiers are used to learn class boundaries within the clusters. As the different base classifiers engage on different difficult‐to‐classify subsets of the data, the learning of the base classifiers is more focussed and accurate. A selection rather than fusion approach achieves the final verdict on patterns of unknown classes. The impact of clustering on the learning parameters and accuracy of a number of learning algorithms including neural network, support vector machine, decision tree and k‐NN classifier is investigated. A number of benchmark data sets from the UCI machine learning repository were used to evaluate the cluster‐based ensemble classifier and the experimental results demonstrate its superiority over bagging and boosting.

In this paper, we present an approach for predicting algae growth through the selection of influential environmental variables. Chlorophyll a is considered to be an indicator for algal biomass and we predict this as a proxy for algae growth. Environmental variables like water temperature, salinity, etc. have influence upon algae growth. Depending on the geographic location, the influence of these environmental variables will vary. Given a set of relevant environmental variables we perform feature selection using a number of algorithms to identify the variables relevant to the growth. We have developed an influence matrix-based approach to select the relevant features. The selected features are then used for predicting algae growth using different regression algorithms to identify their relative strength. The approach is tested on the algae data of Derwent estuary in Tasmania. The experimental results demonstrate that the accuracy of algae growth prediction with influence matrix-based feature selection is superior to using all the features.

This paper presents an algorithm to generate ensemble classifier by joint optimization of accuracy and diversity. It is expected that the base classifiers in an ensemble are accurate and diverse (i.e., complementary in terms of errors) among each other for the ensemble classifier to be more accurate. We adopt a multi-objective evolutionary algorithm (MOEA) for joint optimization of accuracy and diversity on our recently developed nonuniform layered cluster oriented ensemble classifier (NULCOEC). In NULCOEC, the data set is partitioned into a variable number of clusters at different layers. Base classifiers are then trained on the clusters at different layers. The performance of NULCOEC is a function of the vector of the number of layers and clusters. The research presented in this paper investigates the implication of applying MOEA to generate NULCOEC. Accuracy and diversity of the ensemble classifier is expressed as a function of layers and clusters. A MOEA then searches for the combination of layers and clusters to obtain the nondominated set of (accuracy, diversity). We have obtained the results of single objective optimization (i.e., optimizing either accuracy or diversity) and compared them with the results of MOEA on sixteen UCI data sets. The results show that the MOEA can improve the performance of ensemble classifier.

In this paper, we have investigated the influence of cluster instability on the performance of layered cluster oriented ensemble classifier. The final contents of clusters in some clustering algorithms like k-means depend on the initialization of clustering parameters like cluster centres. Layered cluster oriented ensemble classifier is based on this philosophy where the base classifiers are trained on clusters generated at multiple layers from random initialization of cluster centres. As the data is clustered into multiple layers some patterns move between clusters (unstable patterns). This instability of patterns brings in diversity among the base classifiers that in turn influences the accuracy of the ensemble classifier. There is thus a connection between the instability of the patterns and the accuracy of layered cluster oriented ensemble classifier. The research presented in this paper aims to find this connection by investigating the influence of unstable patterns on the overall ensemble classifier accuracy as well as diversity among the base classifiers. We have provided results from a number of experiments to quantify this influence.

The segmentation of road images from vehicle mounted video is a challenging and difficult problem. One of the problems is the presence of different types of objects and not all objects are present in the same frame. For example, road sign is not visible in all frames. In this paper, we propose a novel framework for segmenting road images in a hierarchical manner that can separate the following objects: sky, road, road signs, and vegetation from the video data. Each frame in the video is analysed separately. The hierarchical approach does not assume the presence of a certain number of objects in a single frame. We have also developed a segmentation framework based on SVM learning. The proposed framework has been tested on the Transport and Main Roads Queensland's video data. The experimental results indicate that the proposed framework can detect different objects with an accuracy of 95.65%.

This paper presents a novel cluster-oriented ensemble classifier. The proposed ensemble classifier is based on original concepts such as learning of cluster boundaries by the base classifiers and mapping of cluster confidences to class decision using a fusion classifier. The categorized data set is characterized into multiple clusters and fed to a number of distinctive base classifiers. The base classifiers learn cluster boundaries and produce cluster confidence vectors. A second level fusion classifier combines the cluster confidences and maps to class decisions. The proposed ensemble classifier modifies the learning domain for the base classifiers and facilitates efficient learning. The proposed approach is evaluated on benchmark data sets from UCI machine learning repository to identify the impact of multicluster boundaries on classifier learning and classification accuracy. The experimental results and two-tailed sign test demonstrate the superiority of the proposed cluster-oriented ensemble classifier over existing ensemble classifiers published in the literature.

In this paper we propose different ensemble classifier compositions and investigate their influence on offline cursive character recognition. Cursive characters are difficult to recognize due to different handwriting styles of different writers. The recognition accuracy can be improved by training an ensemble of classifiers on multiple feature sets focussing on different aspects of character images. Given the feature sets and base classifiers, we have developed multiple ensemble classifier compositions using three architectures. Type-1 architecture is based on homogeneous base classifiers and Type-2 architecture is composed of heterogeneous base classifiers. Type-3 architecture is based on hierarchical fusion of decisions. The experimental results demonstrate that the presented method with best composition of classifiers and feature sets performs better than existing methods for offline cursive character recognition.

This paper introduces a novel concept for creating an ensemble of classifiers. The concept is based on generating an ensemble of classifiers through clustering of data at multiple layers. The ensemble classifier model generates a set of alternative clustering of a dataset at different layers by randomly initializing the clustering parameters and trains a set of base classifiers on the patterns at different clusters in different layers. A test pattern is classified by first finding the appropriate cluster at each layer and then using the corresponding base classifier. The decisions obtained at different layers are fused into a final verdict using majority voting. As the base classifiers are trained on overlapping patterns at different layers, the proposed approach achieves diversity among the individual classifiers. Identification of difficult-to-classify patterns through clustering as well as achievement of diversity through layering leads to better classification results as evidenced from the experimental results.

This paper presents a novel approach for creating and training of an ensemble classifier. The approach is based on creating atomic and non-atomic clusters at different levels, training of weak classifiers on overlapping clusters and fusion of their decisions. The subsets of data are obtained by clustering of original training data sets into multiple partitions. As each partition represents highly correlated patterns from different classes, the proposed approach trains weak classifiers on difficult-to-classify patterns and combines the decision at various levels. The approach is tested on six benchmark datasets from UCI machine learning repository. The results show that the proposed approach achieves better classification accuracy than the existing approaches.

Image sequences of smoke, fire, etc. are known as dynamic textures. Research is mostly limited to characterization of single dynamic textures. In this paper we address the problem of detecting the presence of multiple dynamic textures in an image sequence by establishing a correspondence between the feature space of dynamic textures and that of their mixture in an image sequence. Accuracy of our proposed technique is both analytically and empirically established with detection experiments yielding 92.5% average accuracy on a diverse set of dynamic texture mixtures in synthetically generated as well as real-world image sequences.

Characterized by their distinctive motion patterns, temporal textures are natural phenomenon exhibiting spatio-temporal regularity with indeterminate spatial and temporal extent. This paper presents a real-time motion-based temporal texture characterization technique for the first time using block-based motion measures with very high classification accuracy against the popular opinion that such an accurate characterization is only possible using pixel-based measures. Finding an optimal weight ratio between space and time domain features where the accuracy of this block-based technique peaks has been the essence of this success. Computational complexity analyses and classification results clearly demonstrate the capability of the proposed technique in producing comprehensive classification results comparable to the best pixel-based technique with overwhelming reduction in computational complexity.