An Explainable Multi-Camera Vision System for Automated Bolt Inspection

Abstract

In this study, a classical image-processing-based, explainable, and multi-view inspection system for detecting bolt defects is presented. In the proposed system, geometric and morphological analyses are performed using images obtained from different viewpoints, without the need for deep learning. A trigger mechanism based on the region of interest (ROI) center captures a full-frame image at the appropriate moment, followed by defect classification using reference-based comparisons. In this context, features such as bolt head geometry, circularity, width-to-height consistency, visible thread length, thread structure, and profile continuity were evaluated. The experimental evaluation was conducted on a dataset consisting of 339 samples. According to the results, the system achieved an overall accuracy of 73.45%, precision of 88.13%, sensitivity of 75.10%, specificity of 68.29%, and an F1-score of 81.09%. The findings indicate that the proposed approach provides a feasible, interpretable, and low-cost prototype framework for multi-view bolt inspection.