Dual-Granularity Collaborative Progressive Reasoning Network for Referring Remote Sensing Image Segmentation

Referring Remote Sensing Image Segmentation (RRSIS) aims to recognize specific targets in remote sensing images based on natural language expression. However, synergy between coarse-granularity global alignment and fine-granularity local matching of image-text bimodal features is insufficient in existing methods. Furthermore, the cognitive reasoning logic of human visual perception is ignored during the processing of fine-granularity features. Consequently, the model's semantic understanding and target recognition capabilities in complex scenes are limited. To address these issues, we proposed the Dual-Granularity Collaborative Progressive Reasoning Network (DCPRNet) conforming to cognitive logic. First, dynamic extraction and alignment of multimodal features between the remote sensing image and the referring expression were performed. Multimodal representations containing global semantic information were established by the Global Context Perception Module (GCPM). Secondly, in the Fine-granularity Target Reasoning Module (FTRM), deep collaborative alignment of dual-granularity multimodal features was achieved with the injection of remote sensing priors. And the progressive reasoning on remote sensing targets was performed. Then, in the Asymptotic Feature Aggregation Module (AFAM), semantic depth conflation between features of different scales was balanced, and the feature-semantic consistency constraints were formed. Finally, specific targets were effectively recognized. We evaluated the performance of our proposed method on the public RRSIS-D and RefSegRS datasets. The performance of our method was superior to that of state-of-the-art methods in overall performance.