EFFECT OF CONVOLUTIONAL NEURAL NETWORK ARCHITECTURES ON DEFECT DETECTION OF PRODUCTS WITH DIFFERENT SURFACE TYPES

Fast and accurate detection of product defects in manufacturing systems is crucial for cost, quality and customer satisfaction. Surface defects comprise a large portion of production defects and can be successfully detected using deep learning methods. This study employed Convolutional Neural Network (CNN) architectures to classify different types of surface defects and examine the effects of different architectures on various surface types. Four datasets (Metal, Wood, Food, and Fabric) were classified using three CNN architectures: GoogleNet, SqueezeNet, and VGG 19. Validation accuracy was considered in the classification process. According to the results, for the grayscale Metal and Fabric datasets, the GoogleNet architecture achieved higher classification accuracies than the other architectures, with 99,89% and 96,86%, respectively. In contrast, for the color datasets, namely Wood and Food, the VGG19 architecture demonstrated higher classification performance, achieving 94,39% and 99,96%, respectively. As a result of this study, the classification accuracies obtained for the Wood and Fabric datasets were above the average reported in the literature, while the accuracies for the Metal and Food datasets surpassed those reported in previous studies.