Abstract: In recent years, image segmentation technology based on deep learning has made significant progress. However, existing methods suffer from insufficient differentiation between deep and shallow features during feature extraction, which affects the accuracy of segmentation results. To address this issue, an innovative segmentation network called ACB-Net is proposed, which improves segmentation performance through an attention compensation mechanism. Specifically, ACB-Net introduces ResPath to handle the lack of semantic information in low-level features, employs a convolution-compensated CC-Transformer to capture the global context information of deep features, and facilitates the effective fusion of shallow and deep features. This approach enhances local edge information through CNNs and compensates for global information via the attention mechanism, thereby improving the model's ability to represent features at different scales. Experimental results on public dermatology datasets (ISIC 2017 and ISIC 2018) and polyp datasets (Kvasir-SEG, CVC-ColonDB, CVC-ClinicDB, ETIS-LaribPolypDB) demonstrate that ACB-Net significantly outperforms existing methods across multiple metrics. This feature differentiation and complementary mechanism provides an accurate and efficient solution for automatic image segmentation and offers significant reference value for deep learning applications in biomedical image analysis.