基于密集小目标识别算法的港口安全帽与工作服识别  

Safety Helmet and Working Clothes Recognition Based on Dense Small Target Recognition Algorithm

作　　者：祝绍嵩 韩卓成 ZHU Shaosong;HAN Zhuocheng(Research and Development Center of Transport Industry of Automated Terminal Technology,Shanghai 200080,China;COSCO SHIPPING Ports Limited,Shanghai 200080,China;COSCO SHIPPING Technology Co.,Ltd.,Shanghai 200135,China)

机构地区：[1]自动化码头技术交通运输行业研发中心,上海200080 [2]中远海运港口有限公司,上海200080 [3]中远海运科技股份有限公司,上海200135

出　　处：《上海船舶运输科学研究所学报》2024年第5期45-55,共11页Journal of Shanghai Ship and Shipping Research Institute

摘　　要：针对采用YOLOv5_6.0算法采集安全帽与工作服图像存在的目标密集、像素点小、像素色差小和检测难度大等问题,提出一种基于改进的YOLOv5_6.0算法的小目标检测算法。通过优化YOLOv5_6.0算法的边界框回归损失函数改善其对密集小目标特征信息的学习效果;通过增加1层特征提取层提升算法对小目标的检测效果;在算法主干部分添加全局注意力机制(Global Attention Mechanism,GAM)提升算法的整体检测效果;增加第二阶段复检算法对安全帽和工作服目标进行二次复验。在港口环境下采用该算法对包含安全帽和工作服的数据集进行训练,结果显示:改进算法相比原YOLOv5_6.0算法能使安全帽和工作服识别精度均值分别提升5.5%和5.3%;改进算法对安全帽和工作服的平均识别精确率分别达到97%和87%。研究表明,增加网络结构检测层和二阶段复验算法能提升密集场景下小目标识别的精确率和置信度,减少误检和漏检情况,有效满足港口环境下的安全帽和工作服检测需求。To achieve effective detection of safety helmet/working clothe targets from a work site image,an improved YOLOv5_6.0 small object detection algorithm is developed.Detecting helmet/working clothe targets has been challenging because they are small in size and similar in color,and yet,densely distributed.The bounding box regression loss function of the YOLOv5_6.0 algorithm is modified for optimizing the learning effect about dense small target feature information;One additional feature extraction layer is inserted to improve the small target detection performance;A Global Attention Mechanism(GAM)is introduced into the backbone to enhance the overall detection performance;A two-stage YOLOv5_6.0 small structure algorithm for confirming safety helmet/working clothe targets is integrated.The developed algorithm is used to process the images of shipping scenario for verification.The results show that the improved algorithm can achieve the average precision values of 97%for safety helmet detection and 87%for working clothe detection,an improvement of 5.5%and 5.3%respectively,compared to the original YOLOv5_6.0 algorithm.

关 键 词：YOLOv5_6.0算法 边界框回归损失函数 第二阶段复验算法 港口困难场景检测 安全帽 工作服 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TU714[自动化与计算机技术—计算机科学与技术]