Dande Anil Kumar

Continuous sign language recognition (CSLR) faces challenges due to variations in motion on spatial features across consecutive frames, affecting the performance of visual feature extraction models. This paper proposes Motion Selective Sparse Spatial Features (MS3F) to represent motion-rich visual information across video frames. MS3F extracts contextual spatial features by optimizing cross-modal loss between visual and text features, computes frame differencing to construct motion feature vectors, and uses a gated recurrent unit (GRU) to learn motion-selective sparse spatial features. A bidirectional long-short term memory network (Bi-LSTM) is then trained to learn temporal dependencies in MS3F features for tokenized targeted glosses. Experiments on the novel Doordarshan Continuous Indian Sign Language (DC-ISL) dataset, along with established benchmarks RWTH-Phoenix-2014 and Chinese-CSL (CCSL), demonstrate the effectiveness of MS3F, achieving word error rates (WERs) of 20.1%, 19.8%, and 19.9%, respectively. Compared to dense baseline models, MS3F improves WER by 2.9% to 3.4% across datasets while reducing inference time by 30% to 34% and GPU memory consumption by 22% to 27%. The motion-selective sparsity mechanism achieves approximately 30% feature reduction, processing only motion-rich frames and enabling real-time performance on standard GPU hardware. This work demonstrates that MS3F effectively captures visual information containing significant motion content while maintaining computational efficiency, advancing practical CSLR technology for real-world deployment.

Rapid detection of Acute Respiratory Infections (ARI) is crucial to reduce breathing difficulties and severe life-threatening conditions. Automatic cough identification is being conducted using speech frequency analysis and machine learning models. Learning models trained on Mel frequency spectrum(MFCC) features of cough sounds represented as images have recorded an average binary classification accuracy of 68%. Variable cough sound vs silent intervals between samples of a class in MFCC spectral images has shown to influence training algorithms to learn meaningful patterns for classification. To learn all possible local patterns in the MFCC cough images using a vision transformer model (ViT), we propose an image patch overlapping vision transformer IPO−ViT. The patch overlapping factor k controls the quantity of common pixels between them. The IPO − ViT patch encoder computes all possible local pixel pattern relationships by breaking the image into overlapping patches and equating them across all classes making a balanced augmented dataset. The IPO − ViT is evaluated on our own 511 – sound cough dataset (IndiCough_2024) with 5 classes captured at AJ Institute of Medical Sciences, paediatric division along with benchmarks EPFL COUGH VID, Coswara for COVID-19 Diagnosis and Covid19-Cough. The IPO − ViT achieved higher accuracies of around 92.33% over the state-of-the-art cough sound-based disease identification networks.

Water is the most important role in day-to-day life. Water shortage is becoming a reality as water use and pollution continue to grow concern round the universe

Classifying the gestures for various human intelligent machines to work is a typical task. A deep learning based approach for classifying the hand gestures that are used to control the wheel chair is proposed in this work. The actual attributes of a human psyche, for example, hand signals is analyzed in this work and analyzed the developments that add to the improvement of a robotized wheelchair framework. Due to the unavailability of proper hand gesture dataset for controlling the wheel chair, an unique 2D hand gesture data set is created with five subjects in different view orientations. Gestures are created individually using both the hands. A two stream CNN architecture was proposed and the training is initiated on the developed data set. The novelty of the proposed work is tested by designing various experiments. Recognition rates were considered to analyze the efficiency of the proposed methodology. This approach allows people to interact with machines through hand postures. Further, the created dataset is trained and tested on various existing deep learning architectures to validate the novelty of the proposed architecture. An average of 85% recognition rate is achieved with the proposed two-stream architecture.

Action recognition (AR) has achieved outstanding performance in computer-vision techniques and machine learning in recent years. However, most of the other algo