딥 러닝 알고리즘에 기반한 O- 링 결함 검사 프레임워크

Abstract

O-ring defect inspection is a challenging task because of their variable sizes as well as defect types. Traditional quality inspection systems use image processing to extract hand-crafted features and machine learning models to classify defect types. However, they are very sensitive to different sizes of O-ring objects and various defect types. Moreover, software maintenance of traditional systems is hard and requires much effort and time. To solve those challenges, we proposed a fully deep learning (DL) – based O-ring defect inspection framework. Proposed system, firstly, uses DL object localization model to find the location of O-ring objects in an image, then inspects it via using DL classification model. We used pre-trained Convolutional Neural Networks (CNN) for both object localization and defect classification and trained using transfer learning technique. The proposed system is evaluated by using our custom O-ring dataset. Object localization was achieved 98.5 % accuracy on test data. 97 % accuracy was achieved on a classification task. From the experimental results, we see that proposed DL-based system is capable of the inspection of O-ring objects accurately and robust for the various O-ring objects as well as defect types.