From Pixels to Insight: Enhancing Metallic Component Defect Detection with GLCM Features and AI Explainability

Steel surface defect detection is a practically important problem in manufacturing industries. Recently several efforts have been made to synergistically combine computer vision with artificial intelligence to achieve this goal. In this work, we have employed a methodology to employ an automated detection capable of handling large volumes of image data. We have used the NEU database which comprises 1440 Grey scale images containing six defective classes. From these images, we have extracted texture features facilitated by the Gray-Level Co-Occurrence Matrix (GLCM) analysis. GLCM enables quantification of image intensities by exploring spatial relationships between pixels in close proximities. The extracted features were sent as input to Random Forest Classifier with the aim of building a robust classification model. With these six informative attributes, we were able to get an overall test accuracy of 89%. As random forest is a complex approach to understand the overall working principle, we employed SHAP plots to unbox the model and explain the model outputs with easily interpretable visual illustrations.