多尺度高分辨率保持特征融合的手势检测  

HAND DETECTION WITH MULTI-SCALE HIGH-RESOLUTIONPRESERVING FEATURE FUSION

作　　者：杨文姬[1,3] 郑隐馨 梅梦 赵应丁 王映龙[2] 殷华[1] Yang Wenji;Zheng Yinxin;Mei Meng;Zhao Yingding;Wang Yinglong;Yin Hua(College of Software,Jiangxi Agricultural University,Nanchang 330045,Jiangxi,China;College of Computer and Information Engineering,Jiangxi Agricultural University,Nanchang 330045,Jiangxi,China;State Key Laboratory of CAD&CG,Zhejiang University,Hangzhou 310058,Zhejiang,China;Jiangxi Business School,Nanchang 330045,Jiangxi,China)

机构地区：[1]江西农业大学软件学院,江西南昌330045 [2]江西农业大学计算机与信息工程学院,江西南昌330045 [3]浙江大学CAD&CG国家重点实验室,浙江杭州310058 [4]江西省商务学校,江西南昌330045

出　　处：《计算机应用与软件》2023年第11期176-185,共10页Computer Applications and Software

基　　金：江西省自然科学基金青年项目(20212BAB212005);江西省自然科学基金面上项目(20224BAB202015);国家自然科学基金项目(61462038)。

摘　　要：手势交互是人机交互系统的一个重要组成部分。针对现有SSD(Single Shot MultiBox Detector)网络中不同尺度特征间的独立性,无法充分利用各特征间的关联信息,导致对遮挡和不完整手的检测精度偏低等问题,提出一种改进的SSD算法,通过引进多尺度高分辨率保持特征融合模块,将来自不同层的不同分辨率的特征图进行融合形成新的特征图,其不仅保留了原有特征图的特征信息,还结合了不同层的细节信息和上下文较强的语义信息。利用原有SSD检测方法,生成候选预测框,利用非极大抑制(Non-maximum Suppression)得到最终检测结果。实验结果表明,该方法在EgoHands手势数据集上优于原始SSD方法和其他三种先进方法。Hand interaction is an important part of the human-computer interaction system.In view of the independence between features of different scales in the existing SSD(Single Shot MultiBox Detector)network,it is impossible to make full use of the associated information between each feature map,which leads to problems such as low detection accuracy for occluded and incomplete hands.Therefore,this paper proposes an improved SSD algorithm.By introducing a multi-scale high-resolution retention feature fusion module,feature maps of different resolutions from different layers were fused to form a new feature map,which not only retained the original feature information of the feature map but also combined the detailed information of different layers and the semantic information with strong context.The original SSD detection method was used to generate the candidate prediction box,and the final detection result was obtained by using non-maximum suppression(NMS).Experimental results show that this method is superior to the original SSD method and the other three advanced methods on the EgoHands hand dataset.

关 键 词：手势检测 SSD 多尺度 高分辨保持 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]