Abstract: Aiming at the defect that U-Net image segmentation loses too much information in the down-sampling process and the recovery effect is poor in the up-sampling process，which leads to the reduction of image segmentation accuracy，this paper proposes a U-Net image segmentation algorithm based on the multi-level self-attention mechanism. The multi-level self-attention mechanism embedded a self-attention module in front of the upper sampling layer of each layer，which processed the input and scaled the original image of the upper sampling layer into a template image after splicing. Then the upper sampling layer after fusing it with the original input information was output. The U-Net image segmentation algorithm based on the multi-level self-attention mechanism proposed in this paper can not only provide more detailed information through the self-attention module of the original image mosaic but also reduce the impact of background noise caused by the original image mosaic by using the feature selection function of the upper sampling layer and improve the segmentation accuracy of the model. Finally， human body and garment segmentation experiments based on the PASCAL VOC and DeepFashion2 data set were carried out. The experimental results show that the improved method proposed in this paper can improve the image segmentation performance， thus proving the correctness and effectiveness of this method.

Key words: image segmentation, U-Net, self-attention module, multi-level