基于改进YOLOv3算法的手势识别OA
Gesture recognition based on improved YOLOv3 algorithm
针对传统手势识别方法准确率低、速度慢的问题,提出了基于YOLOv3(the third version of you only look once)的手势目标检测算法——YOLOv3-FH(the third version of you only look once with fast speed and high accuracy).首先,提出用于聚类先验框的K-means++算法,以满足手势检测的特定需求.然后,删减部分残差单元并添加4条网络支路,以达到深度特征信息和浅层语义信息融合.最后,合并批标准化层和卷积层,以加速网络模型前向推理的速度.实验结果表明,在手势数据集上YOLOv3-FH算法的平均精度值达到95%以上,单张图片的识别时间大约是13 ms,交并比达到83.12%.
Aiming at the problems of low accuracy and slow speed in traditional gesture recognition methods,an improved gesture object detection algorithm based on YOLOv3(third version of you only look once)was proposed.The novel algorithm was called YOLOv3-FH(third version of you only look once with fast speed and high accuracy).Firstly,a K-means++algorithm for clustering prior boxes was proposed to meet the specific requirements of gesture detection.Then,some residual units were deleted and four network branches were added to achieve the fusion of deep feature information and shallow semantic information.Finally,the batch normalization layer and the convolutional layer were merged to accelerate the forward inference speed of network models.The experimental results showed that the average precision of YOLOv3-FH algorithm on the gesture dataset exceeded 95%.The recognition time for each picture was approximately 13 ms,and the average intersection over union reached 83.12%.
周根;储茂祥;杨永辉;巩荣芬
辽宁科技大学电子与信息工程学院,辽宁鞍山 114051辽宁科技大学电子与信息工程学院,辽宁鞍山 114051辽宁科技大学电子与信息工程学院,辽宁鞍山 114051辽宁科技大学电子与信息工程学院,辽宁鞍山 114051
信息技术与安全科学
手势识别目标检测YOLOv3K-means++残差单元
gesture recognitionobject detectionYOLOv3K-means++residual unit
《安徽大学学报(自然科学版)》 2026 (5)
16-23,8
国家自然科学基金资助项目(71771112)辽宁省自然科学基金资助项目(2022-MS-353)辽宁省高等学校基本科研项目(LJKMZ20220640)
评论