A Consideration on Image Composition of Defects and Background in Appearance Inspection of Plastic Products Based on Machine Learning

Abstract

This paper describes a synthesis method for the images with defects used in training a Region-based Convolutional Neural Network (R-CNN) applied to an appearance inspection of plastic products. The proposed heuristic method is that, firstly, some typical defect patterns are cut from the actual images and then, these defect patterns are pasted to the background by changing the size, rotations, colors, or making them into black-and-white, etc. In the experiments, firstly, 81 defect patterns are cut in 50 actual images and 500 images are synthesized by applying the suggested method. Then, R-CNN is trained by using these 500 images and it is applied to the detection problem of 53 defective patterns in 20 images. As a result, the detection ratio and the hit ratio are 81% and 86%, respectively and it is likely that the suggested training method is promising for practical uses with some improvements.