Enhanced YOLOv5 for Human Hand Recognition

The deployment of contemporary hand detection models in embedded and mobile robotic platforms for human-robot interaction is frequently constrained by prohibitive parameter volumes and computational intensity. To resolve these limitations, this study presents a comprehensive architectural innovation of the YOLOv5s framework through four synergistic optimizations. First, we implement Depthwise Separable Convolution within redesigned C3 modules to construct lightweight C3-DS blocks, substantially reducing network complexity. Second, Ghost convolution operations are systematically integrated to eliminate feature redundancy while further minimizing parametric overhead. Third, a Squeeze-and-Excitation attention mechanism is embedded to enhance discriminative feature learning through channel-wise recalibration. Fourth, the bounding box regression loss is reformulated using WiseIoU to accelerate convergence via dynamic gradient allocation. Experimental evaluations on the COCO-Hand and TV-Hand composite datasets demonstrate the optimized model achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$89.7 \% \text{mAP}$</tex> while reducing parameters by 35% relative to baseline YOLOv5s. The architecture attains an inference speed of 0.025 s per frame, satisfying stringent real-time operational requirements. Furthermore, we replace conventional upsampling with the CARAFE operator, improving feature map fidelity through context-aware reconstruction and enhancing semantic granularity in multi-scale representations. This optimized framework establishes a new paradigm for efficient hand detection in resource-constrained environments, demonstrating significant potential for real-time human-machine interaction systems in rehabilitation robotics and intelligent surveillance applications.