A multiple scale spaces empowered approach for welding radiographic image defect segmentation

Abstract

Welding radiographic image defect segmentation (WRIDS) is a key technology to promote the automation and standardization of quality inspection. However, the complexity of scale variability, aggregation and contextual relationships presented by welding defects pose a great challenge to WRIDS, such as porosity, slag and lack of penetration. To address the issue, a multiple scale spaces (MSSs) empowered segmentation method of complex welding defects is proposed. First, a multi-scale feature space is constructed by dilated convolution with different dilated rates. Second, a multi-scale semantic space is constructed by max pooling in different windows. Third, a multi-scale relational space is constructed through a self-attention mechanism. Finally, the proposed MSSs method is validated on the basis of own collected weld inspection data from an aerospace structural component. The results show that the proposed method can effectively segment multiple complex defect types with an average Acc of 99.46% and an average Miou of 74.65% on the test set.