Approaches for improvement of the X-ray image defect detection of automobile casting aluminum parts based on deep learning

Abstract

Nondestructive testing (NDT) for casting aluminum parts is an essential quality management procedure. In order to avoid the effects of human fatigue and improve detection accuracy, intelligent visual inspection systems are adopted on production lines. Conventional methods of defect detection can require heavy image pre-processing and feature extraction. This paper proposes a defect detection system based on X-ray oriented deep learning, which focuses on approaches that improve the detection accuracy at both the algorithm and data augmentation levels. Feature Pyramid Network (FPN) was primarily adopted for algorithm modification, which proved to be better suited for detecting small defects than Faster R-CNN, with a 40.9% improvement of the mean of Average Precision (mAP) value. In the final regression and classification stage, RoIAlign indicated apparent accuracy improvement in bounding boxes location compared with RoI pooling, which could increase accuracy by 23.6% under Faster R-CNN. Furthermore, different data augmentation methods compensated for the lack of datasets in X-ray image defect detection. Experiments found that an optimal mAP value existed, instead of it continuously increasing with the number of datasets rising for each data augmentation method. Research indicated that the three proposed improvement approaches performed better than baseline Faster R-CNN in X-ray image defect detection of automobile aluminum casting parts.