X-ray weld defect detection method based on dense connection and multi-scale pooling

In order to solve the problems of low segmentation accuracy and fuzzy boundary information of weld defects in X-ray films, this paper proposes an improved Dilated_Pooling_Unet (DP_Unet) network segmentation model. First of all, the codec information extraction module DP_block is added between up and down sampling, aiming to preserve the original defect semantic information to the greatest extent and reduce the loss caused by continuous convolution and pooling operations after down sampling. In addition, the GAM attention mechanism is added to the model to focus on welding. The seam defect part can effectively improve the learning ability of defect feature channels and reduce the influence of background noise. Finally, a hybrid loss function combining binary cross entropy and DiceLoss is proposed to solve the problems of unbalanced positive and negative data during network training. The experimental dataset is composed of the public dataset GDX-ray defect dataset. Experiments show that the method proposed in this paper has a good performance on the GDX-ray dataset, the Dice value reaches 93. 45%, which are significantly improved compared with the baseline algorithm. This method has good segmentation performance, is superior to traditional segmentation algorithms, and effectively improves the segmentation accuracy of negative weld defects.