基于改进YOLOv5算法的堆叠工件检测方法

Stacked Workpiece Detection Based on Improved YOLOv5 Algorithm

在当前复杂的装配场景下,各种工件堆叠在一起,给装配机器人的准确识别带来了挑战,为解决该问题,提出一种改进的YOLOv5模型用于堆叠工件的检测和识别。采用EIFM边缘信息融合模块对目标样本进行轮廓信息的增强;在特征提取网络末端添加MAM多尺度注意力模块来加强对复杂场景和较小目标的检测;将原YOLOv5的Neck网络中的PANet路径聚合网络替换为BiFPN双向特征金字塔融合结构,对高、低特征信息进行加权特征融合;最后,将传统非极大抑制算法改为DIOU_NMS,来减少因工件相互遮挡而产生的漏检。通过算法对比实验、堆叠程度对比实验表明,改进后的YOLOv5算法的mAP达到97.8%,比改进前提升了7.25%;在低、中、高堆叠工件数据集中,目标检测的mAP达到了98.76%、97.93%和94.96%,比改进前的YOLOv5算法分别提升了0.67%、1.56%、4.41%。相比较原YOLOv5算法,改进后的算法模型对堆叠程度较高的工件实现了更精确的识别与定位。

In the current complex assembly scenarios,various workpieces are stacked together,which brings challenges to the accurate recognition of assembly robots.To solve the problem,this paper proposes an improved YOLOv5 model for the detection and recognition of stacked workpieces.The EIFM edge information fusion module is used to enhance the contour information of the target samples;the MAM multi-scale attention module is added at the end of the feature extraction network to enhance the detection of complex scenes and smaller targets;the PANet path aggregation network in the original YOLOv5′s Neck network is replaced with the BiFPN bi-directional feature pyramid fusion structure,which performs weighted feature fusion on high and low feature information;finally,the traditional non-great feature fusion network is replaced with a BiFPN bi-directional feature pyramid fusion structure.fusion;finally,the traditional non-great suppression algorithm is changed to DIOU_NMS to reduce the leakage of detection due to mutual occlusion of artifacts.The algorithm comparison experiments and stacking degree comparison experiments show that:the mAP of the improved YOLOv5 algorithm reaches 97.8%,which is 7.25%higher than the pre-improvement;the mAP of the target detection in the low,medium and high stacked workpiece datasets reaches 98.76%,97.93%and 94.96%,which is 0.67%,1.56%,4.41%higher than the pre-improvement YOLOv5 algorithm,respectively.Compared with the original YOLOv5 algorithm,the improved algorithm model achieves more accurate identification and localization of workpieces with high stacking degree.