Answers are grounded in this paper and cite their sources.
Sign language recognition serves as a critical communication aid for the deaf-mute community, enabling seamless interaction with the hearing population. While deep learning models have demonstrated remarkable recognition accuracy, their large model size makes deployment on resource-constrained edge devices challenging. Knowledge distillation is highly effective in compressing heavy models; however, the classic knowledge distillation approach relies on a static temperature setting, which limits the student model's ability to acquire rich inter-class knowledge from the teacher. This study validates the effectiveness of the Distribution-Aware Dynamic Temperature Knowledge Distillation framework for Indian Sign Language gesture recognition. MobileNetV2 is employed as the teacher model, and two student configurations - MobileNetV2 α=0.5 and α=0.35, representing moderate and aggressive compression levels respectively, are evaluated across multiple performance parameters. Experimental results demonstrate a consistent accuracy improvement of approximately 6% (α=0.5) and 5% (α=0.35) over classic knowledge distillation across both student configurations, with model size reduced from 33.93MB to 5.76MB and 4.64MB, respectively. The quantized student models achieve a maximum compression ratio of 12.85×, retaining 95.92% recognition accuracy with an inference time of 3.63ms, enabling real-time gesture recognition on Raspberry Pi. These results demonstrate the effectiveness of DAD-DTKD for sign language recognition and its potential for compressed model deployment on edge devices.
No takes yet. Share an insight, caveat, or question.
Bhalgaonkar et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: