在肠道息肉分割任务中，由于息肉病变区域与周围正常组织颜色对比度不高，且边界模糊，这导致分割过程中容易丢失关键信息，并受到噪声干扰，影响分割质量。为了解决这些问题，提出一种改进的U-Net模型EEPSNet。首先，EEPSNet通过将普通卷积与空洞卷积相结合，替换了原有的双卷积块。这种组合能捕捉更广泛区域的空间特征，从而减少信息丢失。其次，在解码阶段的特征融合过程中，在U-Net的基础上引入了全局注意力机制特征融合，这种融合方式不仅能够在空间和通道维度上关注显著特征，还能增强模型对噪声干扰的抑制能力。EEPSNet在单卡NVIDIA Quadro RTX 5000 GPU上对四个公开的息肉分割数据集进行了实验，包括Kvasir-SEG和CVC-ClinicDB，用于评估模型的特征建模能力，以及CVC-ColonDB和ETIS-LaribPolypDB，用于评估模型的泛化能力。实验结果表明，EEPSNet模型在这些数据集上均取得了显著的性能提升，mdice分别提高至88.6%、91.1%、70.8%和66.0%，同时也证明了EEPSNet在保持建模能力的同时，也具有良好的泛化能力。In the intestinal polyp segmentation task, due to the low color contrast of the polyp lesion area and the surrounding normal tissue and the blurred boundary, this leads to the loss of key information in the segmentation process and the interference of noise, which affects the segmentation quality. To solve these problems, a modified U-Net model EEPSNet is proposed. First, EEPSNet replaces the original double convolution block by combining ordinary convolution with void convolution. This combination captures the spatial features of a wider region, thereby reducing information loss. Secondly, in the feature fusion process in the decoding stage, global attention mechanism feature fusion is introduced on the basis of U-Net, and this fusion mode can not only focus on salient features in spatial and channel dimensions, but also enhance the ability of the model to suppress noise interference. EEPSNet experiments on four publicly available polyp segmentation datasets, including Kvasir-SEG and CVC-ClinicDB, to evaluate the feature modeling ability of the model, and CVC-ColonDB and ETIS-LaribPolypDB to evaluate the generalization ability of the model. The experimental results show that the EEPSNet model achieved significant performance improvement in these datasets, and the mdice increased to 88.6%, 91.1%, 70.8%, and 66.0%, respectively, which also proved that EEPSNet has good generalization ability while maintaining the modeling ability.