Small Defect Detection in Industrial X-Ray Using Convolutional Neural Network

摘要

It's crucial to ensure the complete reliability of each metallic component in vehicle industry. In the past few years, X-ray testing has been widely adopted in defect detection field. Due to huge production in industry, it's absolutely necessary for manufacturers to employ more intelligent and automated inspection scheme to detect defects efficiently. This study develops an accurate and fast detection method combined with X-ray images using computer vision and deep learning techniques to recognize small defects, mark theirs' area and divide them into different levels according to their sizes. This program modifies the original RetinaNet to adapt to tiny defects. We present a novel data augmentation method aiming to expand the number of defects. Then a multi-scale transform module is designed to generate scale-specific feature map which helps to grade defects better. Experiments show that the proposed method can achieve significant precision improvement over X-ray machine with similarly high recall rate. Both speed and accuracy of this scheme reach practical industrial-service demand.