Adaptive depth and receptive field selection network for defect semantic segmentation on castings X-rays

In order to ensure the safety of important castings, the ADR (Automatic-Defect-Recognition) system should recognize, locate, and count the area of internal defects that are undetectable to the naked eye. However, small differences between inter-classes, large defect scale change, and uncertainly annotation limit the achievement for ADR system. To solve these challenges, this paper presents an adaptive depth and receptive field selection network for defect semantic segmentation on castings X-rays. Firstly, the Resnet18 with ADSM (adaptive depth selection mechanism) is elaborately designed to extract and adaptively aggregate the different depth features, which is beneficial to distinguish the similar defect. Then the ARFB (adaptive receptive field block) is proposed to select the optimum receptive field in a data-driven manner to adapt to the scale change of defects. To overcome the problem of inaccurate labeling caused by the ambiguity of defect edges and the subjectivity of manual annotation, we propose a data augmentation method called lazy-label. Finally, we set up a castings defect segmentation dataset, called SRIF-CDS, to train and evaluate our method. Experiments on this dataset indicate that our method achieves 0.86 mIoU (mean intersection-over-union) and 0.92 mAcc (mean accuracy), which has better performance than the state-of-the-art semantic segmentation baseline.