Improvements of detection accuracy and its confidence of defective areas by YOLOv2 using a data set augmentation method

Recently, CNN (Convolutional Neural Network) and Grad-CAM (Gradient-weighted Class Activation Map) are being applied to various kinds of defect detection and position recognition for industrial products. However, in training process of a CNN model, a large amount of image data are required to acquire a desired generalization ability. In addition, it is not easy for Grad-CAM to clearly identify the defect area which is predicted as the basis of a classification result. Moreover, when they are deployed in an actual production line, two calculation processes for CNN and Grad-CAM have to be sequentially called for defect detection and position recognition, so that the processing time is concerned. In this paper, the authors try to apply YOLOv2 (You Only Look Once) to defect detection and its visualization to process them at once. In general, a YOLOv2 model can be built with less training images; however, a complicated labeling process is required to prepare ground truth data for training. A data set for training a YOLOv2 model has to be composed of image files and the corresponding ground truth data file named gTruth. The gTruth file has names of all the image files and their labeled information, such as label names and box dimensions. Therefore, YOLOv2 requires complex data set augmentation for not only images but also gTruth data. Actually, target products dealt with in this paper are produced with various kinds and small quantity, and also the frequency of occurrence of the defect is infrequent. Moreover, due to the fixed indoor production line, the valid image augmentation to be applied is limited to the horizontal flip. In this paper, a data set augmentation method is proposed to efficiently generate training data for YOLOv2 even in such a production situation and to consequently enhance the performance of defect detection and its visualization. The effectiveness is shown through experiments.