Improved FCOS network-based Instance Segmentation Method

Instance segmentation is an important task in the field of computer vision. Most of existing instance segmentation methods are less accurate and slower to detect, making it difficult to meet the needs of practical engineering applications. What’s more, the existing FCOS network model is too homogeneous in its network structure, which leads to a lack of feature representation capability and greatly limits the performance of target detection. To solve the above problems, a new instance segmentation network based on the FCOS framework is proposed in this paper. The network achieves higher accuracy by introducing bottom branches, top attention branches and fusion modules to eliminate redundant information and refine segmentation results, which could help to realize more accurate location and segmentation of target in image. Experiments on the MSCOCO dataset and qualitative experiments in real scenarios show that the proposed method is effective in improving the segmentation accuracy and can improve segmentation efficiency.