Improvement of YOLOv3 Algorithm in Workpiece Detection

In order to solve the problem of high time complexity and low generalization of traditional methods in the human-machine collaboration scene, this paper applies the YOLOv3 deep learning network to the part of workpiece recognition and detection of the robot workpiece capture. According to the specific application scenarios, the corresponding data set is created to train the YOLOv3 model, and the anchor value suitable for the data set is obtained by the iterative self-organizing data analysis(ISODATA) clustering algorithm. A systematic and comprehensive data augmentation of the data set is carried out for the case where the self-made data set is small and the scene is single. Considering that the target to be detected is small and the background of the detection scene is simple, the YOLOv3 basic network architecture is appropriately pruned. Combining the shallow features with the deep features makes the detection time reduced 4ms while the accuracy of the model is basically unchanged. The comparison experiment on the self-made dataset shows that the improved YOLOv3 algorithm has a mean average precision(mAP) of 0.990 and an average detection time of 60ms. Compared with the original YOLOv3 algorithm, the accuracy of the improved YOLOv3 algorithm is improved by 6%, and the average detection time is reduced by 8ms.