A Novel Boundary-Guided Global Feature Fusion Module for Instance Segmentation

• DOI: 10.1007/s11063-024-11564-6
• Publication Date: 2024-03-06T19:02:36Z
• Authors: Linchun Gao; Shoujun Wang; Songgui Chen
• Keywords: Feature (linguistics)

Abstract The task of instance segmentation is widely acknowledged as being one the most formidable challenges in field computer vision. Current methods have low utilization boundary information, especially dense scenes with occlusion and complex shapes object instances, information may become ineffective. This results coarse masks that fail to cover entire object. To address this challenge, we are introducing a novel method called boundary-guided global feature fusion (BGF) which based on Mask R-CNN network. We designed branch includes Boundary Feature Extractor (BFE) module extract features at different stages. Additionally, constructed binary image dataset containing boundaries for training branch. also trained separately using dedicated before then input into where provide shape needed detection segmentation. Finally, use attention (GAM) further fuse features. Through extensive experiments, demonstrate our approach outperforms state-of-the-art algorithms, producing finer more complete while improving model capability.