Abstract: Accurate polyp segmentation can help doctors find and resect abnormal tissue, and decrease chances of polyps changing into colorectal cancer. Considering the characteristics of polyps with different sizes, shapes, colors, textures, high inter-class similarity, and intra-class variation, we proposed a polyp segmentation neural network combined with multi-attention to achieve accurate segmentation of polyp images. Firstly, Res2Net is used as a backbone network to extract image features, and features extracted by the backbone network are enhanced by channel group spatial enhanced module. Secondly, we use axial self-attention combined with receptive field block strategy to extract fine feature maps which both have low-level detail information and high-level global semantic information. Finally, we use reverse attention to mine boundary information to enhance the network’s segmentation performance for small polyps and its surrounding mucosa is not sharp. The experiments were conducted on five challenging polyp segmentation datasets and compared with five other benchmark methods. The results of experiments show that our method outperforms the compared methods in the mean Dice, mean IoU, MAE, segmentation accuracy, and performance of generalizability.