摘要
针对传统目标检测算法未考虑实际分拣场景目标物形态尺度的多样性,无法获取旋转角度信息的问题,提出基于YOLOv5的改进算法MR;-YOLOv5.通过添加角度预测分支,引入环形平滑标签(CSL)角度分类方法,完成旋转角度精准检测.增加目标检测层用于提升模型不同尺度检测能力,在主干网络末端利用Transformer注意力机制对各通道赋予不同的权重,强化特征提取.利用主干网络提取到的不同层次特征图输入BiFPN网络结构中,开展多尺度特征融合.实验结果表明,MR;-YOLOv5在自制数据集上的均值平均精度(mAP)为90.56%,较仅添加角度预测分支的YOLOv5s基础网络提升5.36%;对于遮挡、透明、变形等目标物,均可以识别类别和旋转角度,图像单帧检测时间为0.02~0.03 s,满足分拣场景对目标检测算法的性能需求.
An improved algorithm MR;-YOLOV5 based on YOLOv5 was proposed aiming at the problem that the traditional target detection algorithm did not consider the diversity of the target shape scale in the actual sorting scene and could not obtain the rotation angle information. Precise rotation angle detection was completed by adding angle prediction branches and introducing angle classification method of ring smooth label(CSL). The target detection layer was added to improve the detection ability of different scales of the model. Transformer attention mechanism was used at the end of the backbone network to give different weights to each channel and strengthen feature extraction. The feature graphs of different levels extracted from the backbone network were input into the BiFPN network structure to conduct multi-scale feature fusion. The experimental results showed that the mean average precision(mAP) of MR;-YOLOV5 on the self-made data set was 90.56%, which was 5.36% higher than that of YOLOv5 s with only angle prediction branch. Categories and rotation angles can be recognized for objects such as occlusion, transparent and deformation. The detection time of single frame is 0.02-0.03 s, which meets the performance requirements of target detection algorithm for sorting scenes.
作者
董红召
方浩杰
张楠
DONG Hong-zhao;FANG Hao-jie;ZHANG Nan(ITS Joint Research Institute,Zhejiang University of Technology,Hangzhou 310014,China)
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2022年第1期16-25,共10页
Journal of Zhejiang University：Engineering Science
基金
国家自然科学基金资助项目(61773347)
浙江公益技术研究项目(LGF19F030001)。
