다수 조명의 채널별 융합을 이용한 CNN 기반 머신 비전 분류기

Abstract

In this paper, a method of combining three grayscale images into a color image and applying it to a recent CNN-based classifier is proposed. Each of the three color channels is assigned by three gray-scale images of a target object taken under different lighting conditions using the same optical system. Consequently, a single virtual color image is obtained. By creating a virtual color image of the classified ball, the generated images and the original gray-scale image used for the machine vision inspection are composed as a dataset, and the classification results are compared by learning with ResNet-50. In the case of bump ball pressing/protrusion, the accuracy of defect classification is computed by conducting experiments five times. Through a comparative evaluation, the accuracy of general grayscale image classification was determined as between 80.2% and 95.5% (depending on lighting conditions), and the average accuracy rate was 89.6%. According to the combination method, the accuracy rate of the virtual color image was distributed between 94.2% and 97.3% and the average accuracy rate was 96.1%. Accordingly, we can confirm a comparative advantage and an average improvement effect of 6.5% in terms of stability and accuracy.