江汉大学学报（自然科学版）

目前主流图像分割算法在分割边界上对特征相似而类别不同的像素鉴别能力不佳，从而影响了分割精度。设计了一种基于曼哈顿距离自注意力机制的U-Net3+图像分割算法，通过关注不同特征点之间信息表征的差异程度来对大范围上下文信息关系进行建模，增强算法对特征相似而类别不同的像素的鉴别能力和对全局关系的学习能力；再通过U-Net3+的全尺度跳跃连接结构将不同尺度的特征相融合，为算法提供更多尺度的上下文信息，使分割算法兼顾细节信息和全局关系。使用COVID-19 CT数据集对该算法进行实验测试，结果表明，引入基于曼哈顿距离自注意力机制后U-Net3+的Dice和IoU指标分别提升了2.79%和3.17%，对比使用多头自注意力机制的U-Net3+分别提升了1.06%和1.02%，证明了该算法的有效性和优越性。

In response to the problem that the current mainstream image segmentation algorithms have poor discrimination ability of pixels with similar features but different categories on the segmentation boundary,which affects segmentation accuracy,this paper designed a U-Net3+ segmentation algorithm based on the Manhattan distance selfattention mechanism. Large-scale contextual information relationships were modeled by focusing on the degree of difference in information representation between different feature points,thereby the network′s ability was enhanced to distinguish pixels with similar features but different categories and learn global relationships. Then,different scale features are fused through the full-scale jump connection structure of U-Net3+,providing more scale contextual information for the network,making the segmentation network balance detailed information and global relationships,thereby improving the segmentation effect. Finally,this paper used the COVID-19 CT dataset to conduct experimental tests on the algorithm. The results showed that after the introduction of the Manhattan-distance-based self-attention mechanism,the Dice and IoU metrics of U-Net3+ were improved by 2. 79% and 3. 17%respectively,compared with the U-Net3+ using the multiple self-attention mechanism improved by 1. 06% and 1. 02%,Which proves the algorithm to be of certain effectiveness and superiority.