Hyperspectral image classification (HSIC) is a challenging task due to the high dimensionality of hyperspectral data, the complex interplay of spatial and spectral features, and the scarcity of annotated samples. Existing approaches, mainly based on tokenization-based feature extraction, introduce artificial segmentation, increasing computational cost, and may lead to information loss. To address these issues, a novel framework, Byte Latent Mamba with Knowledge Distillation (BLM-KD), overcoming explicit tokenization by directly learning byte-level spectral-spatial representations from raw hyperspectral data, is proposed. The Byte Latent Mamba architecture learns compact and expressive byte-level features through an end-to-end convolutional encoder, preserving spectral continuity and spatial structure. A structured State Space Model (SSM) is integrated to model long-range spatial-spectral dependencies efficiently via learned dynamic state transitions. Additionally, an adaptive knowledge distillation (KD) strategy is adopted, where a high-capacity teacher model selectively transfers salient features to a lightweight student model, driven by a temperature-controlled weighting schedule. This ensures robust generalization with reduced model complexity. A patch-based preprocessing scheme also excludes irrelevant zero-labeled samples, refining the training process. Extensive experiments conducted on multiple real-world hyperspectral benchmarks demonstrate that BLM-KD outperforms existing state-of-the-art methods in both classification accuracy and computational efficiency.

Multihead self-attention and cross-attention mechanisms often suffer from computational inefficiencies, limited scalability, and suboptimal contextual understanding, particularly in hyperspectral image (HSI) classification. These mechanisms struggle to effectively capture long-range dependencies while maintaining computational feasibility due to the quadratic complexity of self-attention. To address these challenges, this work proposes PolicyMamba, a spectral-spatial mamba model enhanced with a localized policy attention mechanism. This mechanism reduces computational overhead by restricting attention to nonoverlapping localized regions and enforcing sparsity constraints, ensuring that only the most informative interactions are retained. A hierarchical aggregation strategy further integrates patch-wise attention outputs, preserving spectral-spatial correlations across scales. In addition, a sliding window patch process enhances local feature continuity while mitigating information loss. The PolicyMamba framework integrates spectral-spatial token generation, token enhancement, localized attention, and state transition modules, significantly improving HSI feature representation. Extensive experiments demonstrate that PolicyMamba achieves superior classification accuracy, outperforming conventional and state-of-the-art methods in land cover classification (LCC) by efficiently modeling intricate dependencies in HSI data.

In Transformer-based hyperspectral image classification (HSIC), predefined positional encodings (PEs) are crucial for capturing the order of each input token. However, their typical representation as fixed-dimensional learnable vectors makes it challenging to adapt to variable-length input sequences, thereby limiting the broader application of Transformers for HSIC. To address this issue, this study introduces an implicit conditional PEs (CPEs) scheme in a Transformer for HSIC, conditioned on the input token’s local neighborhood. The proposed spatial–spectral Transformer (SSFormer) integrates spatial–spectral information and enhances classification performance by incorporating a CPE mechanism, thereby increasing the Transformer layers’ capacity to preserve contextual relationships within the HSI data. Moreover, SSFormer ensembles the cross attention between patches and proposed learnable embeddings. This enables the model to capture global and local features simultaneously while addressing the constraint of limited training samples in a computationally efficient manner. Extensive experiments on publicly available HSI benchmarking datasets were conducted to validate the effectiveness of the proposed SSFormer model. The results demonstrated remarkable performance, achieving the classification accuracies of 97.7% on the Indian Pines dataset and 96.08% on the University of Houston dataset.

Hyperspectral images (HSIs) are used in a large number of real-world applications. HSI classification (HSIC) is a challenging task due to high interclass similarity, high intraclass variability, overlapping, and nested regions. The 2-D convolutional neural network (CNN) is a viable classification approach since HSIC depends on both spectral–spatial information. The 3-D CNN is a good alternative for improving the accuracy of HSIC, but it can be computationally intensive due to the volume and spectral dimensions of HSI. Furthermore, these models may fail to extract quality feature maps and underperform over the regions having similar textures. This work proposes a 3-D CNN model that utilizes both spatial–spectral feature maps to improve the performance of HSIC. For this purpose, the HSI cube is first divided into small overlapping 3-D patches, which are processed to generate 3-D feature maps using a 3-D kernel function over multiple contiguous bands of the spectral information in a computationally efficient way. In brief, our end-to-end trained model requires fewer parameters to significantly reduce the convergence time while providing better accuracy than existing models. The results are further compared with several state-of-the-art 2-D/3-D CNN models, demonstrating remarkable performance both in terms of accuracy and computational time.

Convolutional Neural Networks (CNNs) have been extensively studied for Hyperspectral Image Classification (HSIC). However, CNNs are critically attributed to a large number of labeled training samples, which outlays high costs in terms of time and resources. Moreover, CNNs are trained on some samples and have been tested on the entire HSI. Perhaps, the entire HSI is taken into account at test time to appropriately generate the ground truth maps. In order to obtain a higher accuracy while considering the limited availability of training samples and disjoint validation and test samples, this work proposes a fast and compact 3D CNN-based Active Learning (AL) for HSIC that integrates both deep transfer learning and AL into a unified framework. In the proposed methodology, a 3D CNN model is trained with very few training samples (i.e., 5%, only) and in the next phase, the most informative and heterogeneous samples are queried from the validation set (candidate set) based on the fuzziness, mutual information and breaking ties of the trained model. The 3D CNN model is later fine-tuned (rather retraining from scratch) with the new training samples (i.e., 200 samples are selected in each iteration) to reduce the computational cost. The proposed method has been compared with the state-of-the-art traditional and deep models proposed for HSIC. Experimental results proved the superiority of our proposed method on several benchmark HSI datasets with significantly fewer labeled samples. Matlab demo can be accessed on GitHub: github.com/mahmad00.

Future generation vehicles equipped with modern technologies will impose unprecedented computational demand due to the wide adoption of compute-intensive services with stringent latency requirements. The computational capacity of the next generation vehicular networks can be enhanced by incorporating vehicular edge or fog computing paradigm. However, the growing popularity and massive adoption of novel services make the edge resources insufficient. A possible solution to overcome this challenge is to employ the onboard computation resources of close vicinity vehicles that are not resource-constrained along with the edge computing resources for enabling tasks offloading service. In this paper, we investigate the problem of task offloading in a practical vehicular environment considering the mobility of the electric vehicles (EVs). We propose a novel offloading paradigm that enables EVs to offload their resource hungry computational tasks to either a roadside unit (RSU) or the nearby mobile EVs, which have no resource restrictions. Hence, we formulate a non-linear problem (NLP) to minimize the energy consumption subject to the network resources. Then, in order to solve the problem and tackle the issue of high mobility of the EVs, we propose a deep reinforcement learning (DRL) based solution to enable task offloading in EVs by finding the best power level for communication, an optimal assisting EV for EV pairing, and the optimal amount of the computation resources required to execute the task. The proposed solution minimizes the overall energy for the system which is pinnacle for EVs while meeting the requirements posed by the offloaded task. Finally, through simulation results, we demonstrate the performance of the proposed approach, which outperforms the baselines in terms of energy per task consumption.

The proliferation of compute-intensive services in next-generation vehicular networks will impose an unprecedented computation demand to meet stringent latency and resource requirements. Vehicular edge or fog computing has been a widely adopted solution to enhance the computational capacity of vehicular networks; however, the computation requirements of these compute hungry applications will surpass the capabilities of such a solution. To address this challenge, the on-board resources of neighboring mobile vehicles can be utilized. However, such resource utilization requires an incentive mechanism to motivate privately owned neighboring vehicles to participate in sharing their resources. In this paper, we propose a contract theory-based incentive mechanism that maximizes the social welfare of the vehicular networks by motivating neighboring vehicles to participate in sharing their resources. The proposed approach enables the Road Side Units (RSUs) to provide appropriate rewards by offering a tailored contract to each resource sharing vehicle based on their contribution and unique characteristics. Moreover, we derive an optimal contract scheme for computational task offloading, taking into account the individual rationality and incentive-compatible constraints. Finally, we perform numerical evaluations to demonstrate the effectiveness of our proposed scheme. The proposed scheme achieves up to 28% higher computing resource utilization, 17.2% lower energy consumption per computing resource utilization, and 17.1% lesser energy consumption per task completed when compared to the linear pricing incentive baseline.

Anomaly detection suffers from unbalanced data since anomalies are quite rare. Synthetically generated anomalies are a solution to such ill or not fully defined data. However, synthesis requires an expressive representation to guarantee the quality of the generated data. In this article, we propose a two-level hierarchical latent space representation that distills inliers' feature descriptors [through autoencoders (AEs)] into more robust representations based on a variational family of distributions (through a variational AE) for zero-shot anomaly generation. From the learned latent distributions, we select those that lie on the outskirts of the training data as synthetic-outlier generators. Also, we synthesize from them, i.e., generate negative samples without seen them before, to train binary classifiers. We found that the use of the proposed hierarchical structure for feature distillation and fusion creates robust and general representations that allow us to synthesize pseudo outlier samples. Also, in turn, train robust binary classifiers for true outlier detection (without the need for actual outliers during training). We demonstrate the performance of our proposal on several benchmarks for anomaly detection.

In this paper, we propose a new iterative hierarchical data augmentation (IHDA) method to fine-tune trained deep neural networks to improve their generalization performance. The IHDA is motivated by three key insights: (1) Deep networks (DNs) are good at learning multi-level representations from data. (2) Performing data augmentation (DA) in the learned feature spaces of DNs can significantly improve their performance. (3) Implementing DA in hard-to-learn regions of a feature space can effectively augment the dataset to improve generalization. Accordingly, the IHDA performs DA in a deep feature space, at level l, by transforming it into a distribution space and synthesizing new samples using the learned distributions for data points that lie in hard-to-classify regions, which is estimated by analyzing the neighborhood characteristics of each data point. The synthesized samples are used to fine-tune the parameters of the subsequent layers. The same procedure is then repeated for the feature space at level l+1. To avoid overfitting, the concept of dropout probability is employed, which is gradually relaxed as the IHDA works towards high-level feature spaces. IHDA provided a state-of-the-art performance on CIFAR-10, CIFAR-100, and ImageNet for several DNs, and beat the performance of existing state-of-the-art DA approaches for the same networks on these datasets. Finally, to demonstrate its domain-agnostic properties, we show the significant improvements that IHDA provided for a deep neural network on a non-image wearable sensor-based activity recognition benchmark.

Smart devices are becoming more universally adopted and can be used to track and model user activity and monitor for abnormalities. Deviations from what is expected may indicate that a fall is imminent or that an injury has been sustained. Healthcare practitioners can use descriptive models of human kinematics as a tool to monitor patient recovery. This work extends previous work which generated descriptive nonlinear symbolic models of human kinematics with genetic programming. Previously, linear models were developed and compared to the nonlinear models. Although the linear models fit the data well, they were significantly worse than the nonlinear models. In this phase of the project, ensembles of nonlinear models were created to more accurately fit and classify data. Different model selection strategies for the ensembles were investigated. As one would expect, ensembles of models were significantly better than a single model classifier. It was also observed that, although more models in the ensemble yielded better results, only 2 models were required to obtain significantly better results. It was also observed that a random model selection strategy for the ensembles produced competitive results when compared to a more rigorous model selection strategy.

In this paper we propose across-sensor representation learning framework for improving power-accuracy trade-off in multi-sensor human activity recognition (HAR). The goal of the study is to achieve the level of performance comparable to one of multi-sensor HAR systems by using fewer or even single sensor. Such performance is achieved by learning relations between these sensors at training time and utilizing them at test time. These relations are learned by supervised deep models which use multi-sensor data during training only. The absence of need for having multiple sensors during test time allows turning these sensors off and replacing them with a single sensor coupled with learned across-sensor relations. These across-sensor relations make up for the information lost from the turned-off sensors. Using fewer sensors reduces energy consumption of HAR systems deployed on a smartphone. Moreover, it allows building HAR systems for situations when collection of multi-sensor data is possible only during training. This work presents preliminary results achieved with the proposed approach on the SHL dataset. Obtained results show an improvement of up to 14% in classification accuracy of single-sensor HAR.

Smart devices are becoming an extension of ourselves that contain sensitive information and are often targeted for theft. The development of an intelligent and reliable means of user identification and authentication is critical. Not only can the development of user models performing tasks be used for user and task identification, but systems can also notify individuals if there is a potential health concern. The construction of an idealized model of human locomotion may give medical care providers a better understanding of individual differences and guide therapy and treatment. Data was gathered from a smartwatch worn by six subjects performing five different tasks and Genetic Programming was used to perform symbolic regression - a model free, nonlinear type of regression analysis. Symbolic regression was applied to smartwatch data and a collection of nonlinear closed form symbolic mathematical models were generated. Not only did these models fit the data well, but they provided insight into the underlying system. With only 5 seconds of unseen data, the models could classify which subjects were performing which task with 83.9% accuracy when chance was only 3.33%.

Automatic understanding of videos is one of the complex problems in machine learning and computer vision. An important area in the field of video analysis is human action recognition (HAR). Though a large number of HAR systems have already been developed, there is plenty of daily life actions that are difficult to recognize, due to several reasons, such as recording on different devices, poor video quality and similarities among actions. Development in the field of deep learning, especially in convolutional neural networks (CNN), has provided us with methods that are well-suited for the tasks of image and video recognition. This work implements a CNN-based hierarchical recognition approach to recognize 20 most difficult-to-recognize actions from the Kinetics dataset. Experimental results have shown that the application of our method significantly improves the quality of recognition for these actions.

The unprecedented advancements in mobile phone technology on one hand offer plethora of applications to consumers, but on the other hand cause serious risk to users' privacy. Among other modules, camera is one of the most pervasive modules in smart phone used for taking pictures and videos. Furthermore, many applications use camera as an image capturing device that is used for different purposes such as entertainment or identification and authentication (e.g. biometric face authentication). In this paper, we aim at Android camera module and try to find vulnerabilities that could be exploited for camera spoofing. Particularly we aim at different techniques such as modifying the requesting application and creating virtual device at kernel level to use it as a camera. Our experiments revealed that it is still possible to spoof Android camera. Furthermore, based on our findings, we also suggest recommendations to avoid such exploit in Android applications.

Automatic analysis of the video is one of most complex problems in the fields of computer vision and machine learning. A significant part of this research deals with (human) activity recognition (HAR) since humans, and the activities that they perform, generate most of the video semantics. Video-based HAR has applications in various domains, but one of the most important and challenging is HAR in sports videos. Some of the major issues include high inter- and intra-class variations, large class imbalance, the presence of both group actions and single player actions, and recognizing simultaneous actions, i.e., the multi-label learning problem. Keeping in mind these challenges and the recent success of CNNs in solving various computer vision problems, in this work, we implement a 3D CNN based multi-label deep HAR system for multi-label class-imbalanced action recognition in hockey videos. We test our system for two different scenarios: an ensemble of k binary networks vs. a single k-output network, on a publicly available dataset. We also compare our results with the system that was originally designed for the chosen dataset. Experimental results show that the proposed approach performs better than the existing solution.

Pulmonary nodules detection play a significant role in the early detection and treatment of lung cancer. False positive reduction is the one of the major parts of pulmonary nodules detection systems. In this study a novel method aimed at recognizing real pulmonary nodule among a large group of candidates was proposed. The method consists of three steps: appropriate receptive field selection, feature extraction and a strategy for high level feature fusion and classification. The dataset consists of 888 patient's chest volume low dose computer tomography (LDCT) scans, selected from publicly available LIDC-IDRI dataset. This dataset was marked by LUNA16 challenge organizers resulting in 1186 nodules. Trivial data augmentation and dropout were applied in order to avoid overfitting. Our method achieved high competition performance metric (CPM) of 0.735 and sensitivities of 78.8% and 83.9% at 1 and 4 false positives per scan, respectively. This study is also accompanied by detailed descriptions and results overview in comparison with the state of the art solutions.

Bone suppression in lung radiographs is an important task, as it improves the results on other related tasks, such as nodule detection or pathologies classification. In this paper, we propose two architectures that suppress bones in radiographs by treating them as noise. In the proposed methods, we create end-to-end learning frameworks that minimize noise in the images while maintaining sharpness and detail in them. Our results show that our proposed noise-cancellation scheme is robust and does not introduce artifacts into the images.

Classifying hyperspectral data within high dimensionality is a challenging task. To cope with this issue, this study implements a semi‐supervised multi‐kernel class consistency regulariser graph‐based spatial–spectral feature learning framework. For feature learning process, establishing the neighbouring relationship between the distinct samples from the high‐dimensional space is the key to a favourable outcome for classification. The proposed method implements two kernels and a class consistency regulariser. The first kernel constructs simple edges where every single vertex represents one particular sample and the edge weight encodes the initial similarity between distinct samples. Later the obtained relation is fed into the second kernel to obtain the final features for classification where the semi‐supervised learning is conducted to estimate the grouping relations among different samples according to their similarity, class, and spatial information. To validate the performance of proposed framework, the authors conduct several experiments on three publically available hyperspectral datasets. The proposed work equates favourably with state‐of‐the‐art works with an overall classification accuracy of 98.54, 97.83, and 98.38% for Pavia University, Salinas‐A, and Indian Pines datasets, respectively.

The idea of using the gait of a walking person as a biometric identification method has been seen in a number of proposed authentication methods, yet previous works focus on the addition of other authentication methods along with the gait, or require a stationary sensor attached to the hip of the user. This paper uses Genetic Programming to model an identification gait fingerprint for two users, whose walking data was recorded from the accelerometer in a commercially available phone. With the phone freely placed within a pocket, users moved without a fixed protocol at a normal, nonuniform pace. This design of data collection more closely matches the real world applications of such a method. The highly specialized Genetic Programming system with multiple modular enhancements was implemented to perform symbolic regression. The system was demonstrated to be robust to noise and was able to effectively model each dataset with high accuracy. It was also determined that a model could be generated for a subject's whole dataset from only a single step's worth of data. Top models were applied to other subject's data in order to evaluate the uniqueness of these mathematical models.

Smartphones have ubiquitously integrated into our home and work environments. It is now a common practice for people to store their sensitive and confidential information on their phones. This has made it extremely important to authenticate legitimate users of a phone and block imposters. In this paper, we demonstrate that the motion dynamics of smartphones, captured using their built in accelerometers, can be used for accurate user identification. We call this mechanism gait fingerprinting. To this end, we first collected the acceleration data from multiple users as they walked with a smartphone placed freely in their pants pockets. Next, we studied the application of different feature extraction, feature selection and classification techniques from the machine learning literature on these data. Through extensive experimentation, demonstrated is that simple time domain features extracted from these data, which are further optimized using stepwise linear discrimination analysis, can be used to train artificial neural networks to identify legitimate user and block imposter with an average accuracy of 95%.

Wireless Sensor Networks (WSNs) are an emerging type of networks formed by a set of distributed sensor nodes that collaborate to monitor environmental and physical conditions. Mobile medical sensor devices are rapidly emerging as one promising way to monitor patient health and the quality of patient care while improving convenience to the patient and reducing the cost of care by allowing patients to spend more time out of the hospital. In the future, mobile sensors could keep track of everyday behaviors that are reflective of physical and physiological health states and predictive of future health problems. We expect that wearable, portable, and even embeddable sensors will overcome some of the challenges of existing approaches and enable long-term continuous medical monitoring for many purposes. Examples include: Outpatients with chronic medical conditions (such as diabetes); individuals seeking to change behavior (such as losing weight); physicians needing to quantify and detect behavioral aberrations for early diagnosis (such as depression); or athletes wishing to monitor their condition and performance. In this paper, we focus on adapting smartphones used by individuals for health monitoring and present a case study on the design and implementation of a context-aware wireless healthcare application that leverages the capabilities of phone sensor subsystem in tracking human health conditions. The application's detailed scenario and enhanced Android architecture for this solution is presented.

This paper introduces an accurate and robust facial expression recognition (FER) system. For feature extraction, the proposed FER system employs stepwise linear discriminant analysis (SWLDA). SWLDA focuses on selecting the localized features from the expression frames using the partial F-test values, thereby reducing the within class variance and increasing the low between variance among different expression classes. For recognition, the hidden conditional random fields (HCRFs) model is utilized. HCRF is capable of approximating a complex distribution using a mixture of Gaussian density functions. To achieve optimum results, the system employs a hierarchical recognition strategy. Under these settings, expressions are divided into three categories based on parts of the face that contribute most toward an expression. During recognition, at the first level, SWLDA and HCRF are employed to recognize the expression category; whereas, at the second level, the label for the expression within the recognized category is determined using a separate set of SWLDA and HCRF, trained just for that category. In order to validate the system, four publicly available data sets were used, and a total of four experiments were performed. The weighted average recognition rate for the proposed FER approach was 96.37% across the four different data sets, which is a significant improvement in contrast to the existing FER methods.

Physical-activity recognition via wearable sensors can provide valuable information regarding an individual's degree of functional ability and lifestyle. In this paper, we present an accelerometer sensor-based approach for human-activity recognition. Our proposed recognition method uses a hierarchical scheme. At the lower level, the state to which an activity belongs, i.e., static, transition, or dynamic, is recognized by means of statistical signal features and artificial-neural nets (ANNs). The upper level recognition uses the autoregressive (AR) modeling of the acceleration signals, thus, incorporating the derived AR-coefficients along with the signal-magnitude area and tilt angle to form an augmented-feature vector. The resulting feature vector is further processed by the linear-discriminant analysis and ANNs to recognize a particular human activity. Our proposed activity-recognition method recognizes three states and 15 activities with an average accuracy of 97.9% using only a single triaxial accelerometer attached to the subject's chest.