Design tool of deep convolutional neural network for visual inspection

Abstract

In this paper, a design tool for deep convolutional neural network (DCNN) is considered and developed. As a test trial, a DCNN designed by using the tool is applied to visual inspection system of resin molded articles. The defects to be inspected are crack, burr, protrusion and chipping phenomena that occur in the manufacturing process of resin molded articles. An image generator is also developed to systematically generate many similar images for training. Similar images are easily produced by rotating, translating, scaling and transforming an original image. The designed DCNN is trained using the produced images and is evaluated through classification experiments. The usefulness of the proposed design tool has been confirmed through the test trial.