EAM_DiffusionSeg: Research on Instance Segmentation Methods Based on Enhanced Aggregation Masks and Diffusion Models

In this paper, we propose an instance segmentation framework EAM_DiffusionSeg based on enhanced aggregation masks, to address the problems of incomplete instance detection and occlusion sensitivity in diffusion models for multi-target segmentation. To overcome the limitations of existing methods that rely on a single mask optimization path, this work innovatively constructs a dual-path mask generation mechanism: a dual-path mask generation mechanism is proposed: (1) a noise filter based on the diffusion model generates the structure-optimized mask M2(M_diffusion); (2) a mask prediction branch integrated into the local decoder outputsM1(M_local), which preserves finer details. and based on the spatial adaptive weighting, the augmented aggregation formula is designed to achieve the feature complementation. Experiments on the COCO dataset demonstrate that the proposed method significantly improves detection capability and boundary segmentation accuracy in occluded scenes, and provide a new solution for instance segmentation of complex scenes.