<p>Conventional instance segmentation models exhibit two significant limitations: inadequate capture of global features and insufficient refinement of segmentation boundaries, both adversely affect segmentation accuracy. To address these challenges, this study introduces a novel instance segmentation model based on Global-Local attention and Local Chan-Vese method (GLLCV), a box-supervised instance segmentation network. First, the SOLOv2 model is enhanced through the integration of a Bi-directional Feature Pyramid Network for feature extraction and the incorporation of a Global-Local attention module post-FPN, designed to improve the capture of global feature information and enrich global contextual representations. Second, a bounding box projection function is introduced, which maps instance masks to the initial level set of the improved Local Chan-Vese model, to integrate it with the SOLOv2 model to achieve box-supervised instance segmentation. Lastly, the evolution of the level set function in the improved Local Chan-Vese model facilitates the refinement of object contour boundaries, leading to more precise segmentation of object contours. Experimental evaluations demonstrate that the proposed GLLCV model achieves mAP scores of 40.3% and 33.1% on the Pascal VOC and COCO datasets, respectively, thus validating the superior edge segmentation performance of the GLLCV model.</p>