Evaluation and Proposal of CNN for Defect Detection Trained Using Invariant Information Clustering and Only Non-defective Images

Abstract

This paper proposes CNN training using only invariant information clustering (IIC) and non-defective images for visual inspection of industrial products. In the field of visual inspection, there is a problem that there are fewer defective images than non- defective images. To solve this problem, CNN is trained as an encoder using invariant information clustering, and only non-defective images. One-Class Support Vector Machine (OCSVM) was utilized to classify the feature vectors obtained from the encoder. Defective images and non-defective images are classified by detecting values that deviate from the features of the non-defective images. A typical supervised learning model, a model using a variational autoencoder (VAE), and an IIC model are compared. The results confirm that the IIC model that does not use defective images is equivalent to a supervised learning model that uses a small number of defective images.