Part recognition is an important basis for the assembly and packing of mechanical components. The efficiency of manual recognition is low. The traditional machine vision inspection requires high and the scene is single. This paper proposes a machine part recognition method based on deep learning machine vision. The detection accuracy of the original Mask R-CNN instance segmentation model is improved by adding the PointRend module; the category subdivision method is designed for parts with high similarity, through size estimation and feature matching, It can better solve the problem of size feature loss caused by data-enhanced image scaling. Collecting 25 different parts for recognition experiments, the results show that the method in this paper can effectively improve the recognition accuracy of mechanical parts, and the algorithm can recognize similar parts with an accuracy of 100%, which is 11.51% higher than the original Mask R-CNN method. And the method in this paper can be extended to other recognition tasks with similar features.