모노 카메라를 사용하는 머신 비전 시스템에서 비정형 결함을 검사하는 CNN을 훈련하기 위한 데이터 증식 방법

초록

Applying convolutional neural networks (CNN) to machine vision has recently exhibited excellent performance in amorphous defect inspection. However, collecting and annotating sufficient amounts of data to train CNNs in machine vision systems take considerable time. In this study, a data augmentation method is proposed that can be used to inspect amorphous defects using CNNs in machine vision systems using mono cameras. Class 1 subdata of the DAGM 2007 dataset produced to detect defects in arbitrary patterns were used as experimental data. We trained Mask R-CNN ResNet 50 for defect inspection. Through the proposed method, we found that the CNN trained with the augmented data exhibited an average accuracy of 61.38%, which is 16.73%p higher based on Mask mAP@0.5:0.95 compared to the CNN trained with the original data when validation was performed with original grayscale images. By using the data augmentation method applied in this study, CNNs in the machine vision systems using mono cameras can achieve higher inspection performance.

키워드