Defect segmentation algorithm for copper pipe soldering in X-ray images based on AM-TransUNet

In the X-ray digital images of soldered copper pipes with small diameters, the significant variations in size, shape, and contrast of soldering defect images make it difficult to achieve automatic defect recognition. To address this issue, an improved defect segmentation model, namely AM-TransUNet, based on TransUNet was proposed, which combined multi-scale feature representation with the Biformer attention mechanism. The Res2Net module, combined with the SimAM attention mechanism, replaced the ResNet module in the encoder of the baseline TransUNet model, enhancing the model’s multi-scale feature extraction capabilities and ability to capture local defect details in complex backgrounds. Additionally, the Biformer dynamic sparse attention mechanism was cascaded at the end of the encoder to improve detection speed and enhance the model’s ability to capture global information. Defect segmentation experiments conducted on a constructed X-ray image dataset of small-diameter copper pipes demonstrate that the AM-TransUNet model achieves 2.8% and 3.2% improvement in precision and recall, respectively, compared to the baseline model. Moreover, the HD95 metric is reduced by 13.2, and the model shows superior sensitivity in detecting unfused defects with a radial dimension of 0.5 mm. The results indicate that the AM-TransUNet model has higher reliability in defect segmentation. Meanwhile, the detection speed of the AM-TransUNet model is improved by 9 frame/s over the baseline model, and the number of model parameters is reduced by 66 M.