基于改进YOLOv5s的手语识别算法  

Sign Language Recognition Algorithm Based on Improved YOLOv5s

作　　者：包书涵 孙慕伦 刘舒祺 邓琳 华俊宏 何向武 Bao Shuhan;Sun Mulun;Liu Shuqi;Deng Lin;Hua Junhong;He Xiangwu(College of Business,Jiaxing University,Jiaxing,Zhejiang 314001)

机构地区：[1]嘉兴大学商学院,浙江嘉兴314001

出　　处：《嘉兴大学学报》2024年第6期46-57,共12页Journal of Jiaxing University

摘　　要：为应对聋哑人群体在日常生活中的沟通障碍问题,提出了优化YOLOv5s网络的新算法.针对YOLOv5s模型体积庞大、参数过多导致移动端检测效率低下的问题,进行了两个方面的改进:一是轻量化处理,采用MobilenetV3轻量化结构取代其主干网络的骨干层;二是采用ECANet注意力机制替换颈干网络中的C3模块,增强模型的特征提取能力,提高模型的检测精度和对环境的适应能力.使用HAGRID数据集进行实验的结果显示:改进后的YOLOv5s-MobilenetV3-ECA模型平均精度mAP达到了99.6%,且在不同环境下,模型对于手语的识别置信度达到了0.9,与改进前的YOLOv5s模型相比,模型大小仅为原来的19.4%,帧率每秒提高了9.7帧,具有更好的泛化能力.改进后的模型可以在移动端上较好地部署,为手语识别和聋哑人士的交流提供了一定的参考价值.In order to address the communication challenges faced by the deaf and mute community in their daily lives,this paper proposes a new algorithm to optimize the YOLOv5s network.To tackle the issue of the large size and excessive parameters of the YOLOv5s model leading to low detection efficiency on mobile devices,two improvements are made in this study:first,lightweight processing is implemented by replacing the backbone layers of the YOLOv5s model with the lightweight structure of MobilenetV3;second,the ECANet attention mechanism is adopted to replace the C3 module in the neck network,thereby enhancing the model s feature extraction capability and improving its detection accuracy and adaptability to the environment.Experimental results using the HAGRID dataset show that the improved YOLOv5s-MobilenetV3-ECA model achieves a mean average precision(mAP)of 99.6%,and it reaches a confidence level of 0.9 for sign language recognition in different environments.Compared to the original YOLOv5s model,the improved model is only 19.4%of its size and has a frame rate increase of 9.7 fps,as well as better generalization ability.The improved model can be well deployed on mobile devices,providing a valuable reference for sign language recognition and communication for the deaf and mute community.

关 键 词：YOLOv5s 轻量化 手语识别 MobilenetV3 ECANet 通道注意力 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]