Paper Title
EdgeSignNet: Real-Time Sign Language Recognition on Edge Devices via Hybrid Attention and Adaptive Inference

Abstract
Existing sign language recognition (SLR) systems achieve high accuracy in laboratory settings but struggle on resource-constrained edge devices, hindering real-world deployment. This paper introduces EdgeSignNet, a novel system addressing these challenges via three innovations: (1) a Hybrid Attention Module (HAM) that improves accuracy by 3.2% with only 0.3% computational overhead; (2) Adaptive Depth Inference (ADI), which dynamically adjusts network depth to reduce latency by 22%; and (3) Sign-Aware Mixed Precision Quantization (SAMPQ), achieving a 2.1× speedup by selectively protecting critical layers. Evaluated on ASL, LSA64, and GSL datasets, EdgeSignNet achieves 94.7% accuracy at 31.2 FPS on a Raspberry Pi 4, outperforming state-of-the-art lightweight models by 3.4% in accuracy while being 1.8× faster. A 42-participant user study further validates its efficacy, yielding 89.3% real-world accuracy and a 4.2/5 satisfaction rating. At only 67MB, EdgeSignNet enables high-performance assistive technology on commodity hardware, making widespread SLR adoption practically viable. Keywords - Sign Language Recognition, Edge Computing, Attention Mechanisms, Adaptive Inference, Mixed-Precision Quantization, Assistive Technology