Basic Research on Detection of Defective Products with Minute Defects Using Convolution Neural Network (CNN) and Support Vector Machine (SVM)

Abstract

In this paper, we propose visual inspection systems that detect defective products with minute defects seen in the production process of resin molded articles. Firstly, an application of 15-layered CNN is developed for visual inspection and is trained using a large number of images to perform effective generalization. Then, the trained CNN named sssNet is incorporated with a two-class learning based SVM to classify test images with high recognition rate into accept as OK category or reject as NG category, in which compressed feature vectors obtained from the CNN is used as the inputs for the SVM. Finally, the CNN and the SVM are compared and evaluated through classification experiments. It has been confirmed that the SVM has a higher classification ability than the CNN.