Defect Detection in Metal-Ceramic Substrate Based on Image Processing and Machine Learning

Abstract

Automatic detection of defects in a substrate is important in the manufacturing process. Defect detection in metal-ceramic substrates relies on manual assembly lines, which are laborious and time consuming. This study develops a novel defect detection method that automatically localizes the substrate area without a background and enhances the defect area by combining images acquired using a platform with different illumination patterns. In this study, we use a defect image dataset to train convolutional neural networks (CNNs) for defect discrimination. Seven evaluation metrics, i.e., accuracy, F1-score, area under the curve, and prediction time of all the patch images in the test set, are comprehensively considered to select the optimal parameter configuration. A developed ResNet-50-based model achieves the highest defect discrimination accuracy of 99.8%. Based on extensive experiments and their results, this study provides clear guidance for devising a feature extraction method based on multidirection filter processing and optimization of CNNs for classification tasks. Finally, a framework for defect detection in a metal-ceramic substrate for practical use in the power device industry is developed.