【作者】 汪刘洋； 后兴海； 李智多； 盛智勇； 刘文楷；

【Author】 WANG Liuyang;HOU Xinghai;LI Zhiduo;SHENG Zhiyong;LIU Wenkai;College of Information,North China University of Technology;School of Electrical and Control Engineering,North China University of Technology;

【机构】 北方工业大学信息学院； 北方工业大学电气与控制工程学院；

【摘要】 一种基于渐进融合与位置强化驱动的YOLOv8(you only look once v8)模型被用于书法握笔姿势检测技术中,针对握笔姿势细节丰富,握笔姿态多样,关键特征难以提取等问题,该技术在模型主干网络中融入了CA(coordinate attention)坐标注意力机制,增强模型对手部关键特征的学习能力,提高检测精度。同时采用AFPN(advanced feature pyramid network)渐进式特征金字塔网络以代替原有的PAN(path aggregation network)结构,提高了模型在不同背景与光照条件下的握笔姿势检测稳健性和泛化能力,同时保证了运行效率。该方法依靠握笔手势特征自建了包含12种握笔姿势类型的20 651张图像的握笔姿势数据集。通过实验验证,改进后的YOLOv8模型在握笔姿势检测任务上比原有的模型m AP值高了3.4个百分点,达到了94.3%,展现出较高的识别准确率、召回率及良好的实时性。更多还原

【Abstract】 An improved YOLOv8 model,driven by progressive fusion and position enhancement,targets the detection of intricate calligraphy brush holding postures.Addressing issues of rich grip details,diverse postures,and the difficulty in extracting key features.The model integrates CA(coordinate attention) into YOLOv8’s backbone,enhancing its capability to learn critical hand features and improve detection accuracy.Replacing the original PAN(path aggregation network) with an AFPN(advanced feature pyramid network) boosts robustness and generalization across varied backgrounds and lighting conditions while maintaining operational efficiency.A custom dataset of 20 651 images,encompassing 12 distinct brush holding styles,supports the development.Experimental validation shows that the modified YOLOv8 surpasses its predecessor by 3.4 percentage points m AP,reaching 94.3%,demonstrating high identification accuracy,recall,and real-time performance.更多还原