Abstract:  In the current industrial bolt production process, the sorting of stacked bolts still needs to be completed manually, which not only has low work efficiency, but also leads to the waste of a large number of human resources. Aiming at this problem, the YOLOv5 network model was improved, and a SE_YOLOv5 network model was proposed. Firstly, the P′1 feature layer was deleted in the Neck part of the original network, which reduced the extraction of shallow information by the network, and improved the real-time performance of network detection without affecting the detection of large-size objects; Then, the Backbone module was improved to make the network focus on the important parts of the image more efficiently by adding the squeeze-and-excitation(SE) attention mechanism, and the accuracy of the network's detection of stacked bolts was enhanced; Finally, the minimum overlap method of the detection frame was proposed to reduce the collision between the gripper and the non-target bolt during grasping, and the grasping point pose of the bolt detection frame was optimized to improve the success rate of grasping. The research results show that the average accuracy of the SE_YOLOv5 network is 86.5% and the average detection speed is 13.02 FPS. Compared with the original YOLOv5s network model, the SE_YOLOv5 network has improved the detection accuracy by 1.2% and the detection speed by 2.71 FPS, and the SE_YOLOv5 also has higher detection accuracy and detection speed than other detection models. The grasping experiment shows that the model can effectively guide the robotic arm to carry out bolt grasping operation.
Key words:  sorting of stacked bolts; SE_YOLOv5 network model; squeeze-and-excitation (SE) attention mechanism; minimal overlap method; grasping operations; point-taking attitude optimization