Projects

Summary: Deep neural network-based techniques are effective for image classification, but the limitation of having to train classifiers for unseen concepts may increase the overall response-time for multimedia-based event processing models. This work focuses on foundational aspects of the problem of reducing response-time for online adaptive classifiers-based multimedia event processing which includes introducing object detection operators, standardization of the concept of response-time, identification, and proposed multiple IoMT based deep neural network models while using object detection specifically You Only Look Once (YOLO), Single Shot MultiBox Detector (SSD), and RetinaNet, and applying transfer learning. Lastly, I report the best possible performance of current object detection models for the online construction of classifiers. The major challenge in training deep neural network-based models is the need to collect many images with bounding box annotations, which is impossible for millions of unseen concepts. My final specific work is the design of first and fast detector for the training of unseen classes using only image-level labels with no bounding box annotations. It takes 10 min only to train an object detector.

Languages: Python, C, Shell Scripting (Linux Platform)
Libraries: TensforFlow, CUDA, cuDNN, Keras, OpenCV
Hardware Used: Nvidia Titan Xp GPU

Publication Outcomes: 3 Journals, 4 Conference papers (in CVPR 2022 selected for Oral Presentation, Journals of IEEE Access, Elsevier, and Springer; and other conferences including ICMR and DEBS)