Research

• DUGO- A dual-scale temporal encoding framework combining depthwise convolutions and Transformer-based self-attention to capture fine-grained motion patterns and long-range temporal dependencies in streaming videos.
• DUGO's Gaussian Memory Bank (GMB) uses input-conditioned Gaussian prototypes and gated cross-attention to model diverse procedural histories and enrich encoder representations and action anchor queries.
• Integration of a 1D Distance-IoU (DIoU) regression loss within DUGO to improve temporal action boundary localization through joint optimization of segment overlap and center alignment.
• Evaluation of DUGO on EGTEA, EPIC-Kitchens-100, and MUSES, demonstrating strong online action localization performance across egocentric procedural and general video benchmarks.
• Ablation studies of DUGO highlighting the complementary benefits of dual-scale temporal encoding, stochastic Gaussian memory, and DIoU-based regression for online temporal action localization.

• Developed AGSA-Net, a dual-branch framework that integrates spectral unmixing with transformer-based hyperspectral image classification.
• Estimated physically meaningful subpixel abundance maps and constructed an abundance affinity prior to capture material-level relationships between pixels.
• Proposed an abundance-guided multi-head self-attention mechanism that injects material affinity directly into the transformer attention computation.
• Fused abundance-derived material descriptors with transformer-based spectral-spatial representations for final hyperspectral image classification.
• Evaluated the proposed approach on the Indian Pines, Augsburg, and Berlin hyperspectral datasets, demonstrating improved contextual modeling particularly in heterogeneous scenes.

DOI: 10.1109/TGRS.2026.3729910

• Utilized two data encoding techniques, Discrete Encoding and Ranged Encoding, to align the open-source CKD datasets (CKD15 and CKD21) with inconsistent feature types for model training.
• Investigated multiple feature selection techniques and a unified feature selection approach to identify the optimal number of features for training and evaluating various classifiers.
• Proposed a novel ensemble classifier, Tri-phase Ensemble, combining Voting, Bagging, and Stacking techniques, and compared its performance with base classifiers.
• Implemented two Multi-layer Ensemble Classifiers using blending and stacking approaches, with interchanged layers to monitor their performance and robustness.
• Applied explainable AI to understand and interpret the predictions made by the machine learning models.

DOI: 10.1016/j.cmpbup.2024.100173