Deep Learning-Based Dimension Measurement System for Industrial Part Inspection

This paper develops a deep learning-based system for the dimension measurement of industrial parts. The system employs a general self-supervised image segmentation network to perform dichotomous image segmentation (DIS) on the parts, effectively extracting binary images of the target components from complex backgrounds without requiring preprocessing steps such as filtering or denoising. Following segmentation, contour detection, and dimension measurement are executed. The system utilizes a CMOS area scan camera and was tested on a computer equipped with a GTX1050 graphics card. The tests, conducted on two typical thrust ball bearing components (i.e., the ball retainer and the shaft ring) - yielded ten measurements per part. Comparison with a vernier caliper as the reference standard reveals that under normal operating conditions, the maximum standard deviation is 0.007 mm and the maximum absolute error is 0.017 mm.