基于YOLOv5的静态手势识别检测模型

Detection Model for Static Gesture Recognition Based on YOLOv5

针对实时手势检测需求,提出一种基于YOLOv5的手势识别算法。通过采用轻量级主干网络MobileNetV3替代YOLOv5s中的CSPNet-53,优化后的主干网络整合了深度可分离卷积与SE注意力机制,形成模型M_YOLO_N(Mo⁃bileNet_YOLOv5_NewIou)。与原始模型相比,M_YOLO_N的参数量减少了33%,计算复杂度(GFLOPs)降低了54%,在自制手势数据集上的mAP@0.5提升了2.4%。该模型不仅实现了轻量化,而且有效解决了实时检测问题。针对多尺度手势检测,保留SPPF模块,并引入归一化高斯瓦伦汀距离(NWD)技术,提出新的边界框损失函数NewIoU。在不增加参数的前提下,改进后的模型在多尺度手势检测中的置信度提升了20%。

To meet the demand for real-time hand gesture detection,this paper presents a YOLOv5-based gesture recognition algorithm.By replacing CSPNet-53 in YOLOv5s with the lightweight MobileNetV3,the optimized backbone integrates depthwise separable convolutions and the SE attention mechanism,forming the M_YOLO_N model.Compared to the original,M_YOLO_N reduces parameters by 33%and decreas⁃es computational complexity by 54%.On a custom dataset,mAP@0.5 increased by 2.4%.This model achieves both lightweight design and re⁃al-time detection.For multi-scale detection,the SPPF module is retained,and the normalized Wasserstein distance(NWD)is introduced,proposes a new bounding box loss function NewIoU.Without increasing parameters,detection confidence improved by 20%.