Colorectal polyp segmentation can effectively assist doctors in screening for colorectal adenomas， but polyp segmentation has problems such as more noise and insufficient boundary distinguishability. In response to these issues， this paper designed a multi-scale polyp segmentation network that adopts cascaded strategy to fuse boundary features. Firstly， this paper proposed an improved channel grouping spatial enhancement module to enhance the image features extracted by the backbone network， thereby improving the correlation between channels and spatial positions. Secondly， considering the insufficient boundary distinction， a cascaded feature fusion network was designed to better retain boundary information and improve boundary distinction， so as improve the segmentation accuracy. Finally， a dual-branch hybrid upsampling module was introduced to obtain more detail feature information， so as to realize the complementarity of features and capture more complete and effective features. Tested on the CVC-ClinicDB and Kvasir datasets， our algorithm achieves mean Dice coefficients of 0.944 and 0.920， and mean Intersection over Union of 0.900 and 0.869 respectively， compared to the M²SNet algorithm with average Dice coefficients of 0.922 and 0.912， and mean IoU of 0.880 and 0.861 respectively. Tested on the ETIS-LaribPolypDB， CVC-300， and CVC-ColonDB datasets， our algorithm achieves mean Dice coefficients of 0.776， 0.915， and 0.782 respectively， while the M²SNet algorithm achieves mean Dice coefficients of 0.749， 0.903， and 0.758 respectively. Experimental results show that the proposed algorithm has high segmentation accuracy and strong generalization ability.