Ph.D Defence of Sumira Rasool

Ms. Sumira Rasool, Ph.D. Research Scholar has submitted thesis on "An Efficient Deep Neural Network Architecture for Sign Language Recognition using Deep Compression" to the University of Peshawar, in partial fulfillment of the requirements for the award of degree of Doctor of Philosophy (Ph.D.) in Computer Science

The oral examination (Public Defence) is scheduled to be held on July 15, 2026 at 9.00 a.m in the Department of Computer Science, University of Peshawar. The abstract of the thesis is attached herewith. All those interested in the said research work may participate in the event. They may raise relevant questions during presentation by the scholar for further evaluation.

ABSTRACT

People with speech and hearing impairments face difficulty communicating with other people in the community. Sign language is the main source of communication used by hard-of-hearing people. Sign Language Recognition (SLR) is a challenging task. Deep learning models are powerful in solving today's complex classification tasks. Recent advances in deep learning and computer vision methods have shown remarkable progress in the area of sign language recognition. However, running these large models on resource-constrained devices for efficient inference is very challenging.

The focus of this research study is to develop computationally efficient and compact deep learning models. Compact machine learning models are revolutionizing industries by enabling complex artificial intelligence (AI) tasks to be performed on devices rather than on the cloud.

In this research study, a computationally efficient hybrid compression method is proposed that can compress deep neural networks for efficient sign language recognition. The benchmark dataset used is the American Sign Language (ASL) Alphabet dataset. Our proposed hybrid compression method combines pruning, knowledge distillation, and quantization techniques. Combining these approaches achieves significant compression of the existing deep learning models with minimal effect on accuracy.

The proposed hybrid model compression framework, which employs structured block wise pruning, confidence attention map based knowledge distillation, and low-bit INT8 quantization, systematically reconciled over parameterization and efficiency in deep learning. The proposed block wise pruning method with 50% pruning ratio achieves accuracy gains between 1% - 5% approx. after pruning. The proposed knowledge distillation method can effectively transfer useful features from the teacher to the student model by computing confidence scores, leading to improved accuracy (>99%) compared to the baseline and state-of-the-art approaches. The proposed low-bit INT8 quantization method effectively reduces model size and improves inference speed by reducing the bit width of weights from a higher 32-bit floating point to a lower 8-bit integer. The compressions range between 8× -11× approx, with <1% drop in accuracy compared with the baseline and state-of-the-art approaches, indicating the efficiency of 8-bit parameter representations without changing the architecture. A lightweight CNN is also proposed in this study that is computationally efficient in terms of parameter count, FLOPs, and model size while achieving accuracy close to the benchmark deep learning models. Hence, the findings presented in this study demonstrate that the proposed hybrid model compression method makes it easier to deploy deep learning models on devices with limited computing facilities.