Machine vision-based defect detection system for watch back covers

Abstract This study presents a machine vision-based system for non-contact defect detection in watch back covers, comprising both hardware and software components. The system architecture integrates image acquisition devices (light sources, cameras, and lenses) with a multi-stage processing pipeline: 1) Image enhancement through bilateral filtering and gamma transformation; 2) Edge detection using an optimized Canny algorithm with four-directional convolution kernels; 3) Geometric feature extraction via least squares fitting; 4) ORB-based feature point detection with RANSAC optimization; 5) Defect identification through logical operations and image differencing; 6) SVM classification using aspect ratio and roundness features. Experimental validation on 110 samples demonstrated 97.6% detection accuracy, significantly outperforming traditional inspection methods in efficiency and reliability.