To suppress irrelevant semantics and cross semantic gaps in object extraction using an encoder-decoder network structure， thereby achieving higher accuracy. U-Net is used as the backbone network for feature extraction. To reduce semantic differences between shallow and deep features， a multi-scale semantic pooling module （CSP， Channel-Spatial-Pyramid） integrates attention perception and replaces skip links in early layers. The CSP module emphasizes more meaningful semantic information from two levels corresponding to space and channel， extracts features at different scales through parallel branches of four different pooling cores， and aggregates all branch results to splice with the features of later layers. The experimental results show that the Dice index of CSP-Net in color fundus image disc segmentation reaches 99.6%， whereas that of cup segmentation reaches 92.1%. Both results represent improvements over existing algorithms. CSP-Net exhibits a high effectiveness and anti-interference ability for extracting small targets in fundus images， making it appropriate for clinical reference in glaucoma screening and diagnosis.