A Segmentation Network for Coastal Cegetation Guided by Category-Wighted Information from UAV Perspective Based on SAM
摘要
The segmentation challenge caused by the similarity of different semantic class features has always been a significant research topic in the field of semantic segmentation. Based on this, this paper proposes MEAFormer, which designs a fusion branch to address the segmentation difficulties caused by the similarity in vegetation height. In the fusion branch, we have designed a Mixed Convolution Attention module (MCA) composed of deep stripe convolution and deep dilated convolution to increase the receptive field, as well as a Dual Attention Fusion Module (DAFM) composed of spatial and channel attention to integrate features from different levels. Furthermore, we introduce the SAM branch, which utilizes the mask obtained from the MEAFormer branch as guidance information to perform fine-grained segmentation using SAM, and finally corresponds the mask with the semantic labels. Our model achieves excellent results on the UAV-MARINE-SEG dataset collected by us.