Development of Design and Evaluation Tool for CNN, CAE and SVM - Image Processing to Improve Visualization Function of Defect Areas Included in Mechanical Parts Using Grad-CAM -

Abstract

The authors have been developing a design, training and evaluation application with a userfriendly operation interface for CNN (Convolutional Neural Network), CAE (Convolutional Auto Encoder) and SVM (Support Vector Machine), which can be used for the defect detection of various kinds of industrial products even without deep skills and knowledges concerning information technology. When the Grad-CAM is applied to visualizing interested areas affecting the classification results, different areas not relating to target defects are sometimes mapped majestically. In this presentation, the visualization performance of defect areas using the Grad-CAM is tried to be improved. Before learning process, all images in training data set are preprocessed by a proposed masking method, in which not-interested areas in each image are replaced with randomly generated mask patterns. The effectiveness and promise are observed through visualization tests of defect areas using the Grad-CAM.