Pipeline welding seam image segmentation method based on improved UNet

The invention provides a pipeline welding seam image segmentation method based on an improved UNet. According to the method, firstly, collected pipeline welding seam image data are preprocessed, a data set is established, then an improved UNet network pipeline welding seam image segmentation model is constructed, and the specific implementation method comprises the steps that firstly, dense connection blocks are combined with a parallel attention mechanism to construct a network down-sampling module, and the model feature extraction capacity is improved; the method comprises the following steps: firstly, calculating a pipeline welding line image, replacing common convolution with depth separable convolution, reducing the calculated amount of a model, and finally, relieving the condition that the proportion of a welding line area to a non-welding line area in the pipeline welding line image is unbalanced by using a new loss function combining weighted Dice and a cross entropy function; and finally, performing