PriorSAM-DBNet: A SAM-Prior-Enhanced Dual-Branch Network for Efficient Semantic Segmentation of High-Resolution...

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Semantic segmentation of high-resolution remote sensing imagery is vital for environmental and urban monitoring.
• Challenges include sensor occlusions, multi-scale targets, and the trade-off between global and local feature accuracy.
• Existing methods struggle with domain gaps and lack of semantic categorization when applying models like Segment Anything Model (SAM).

Purpose of the Study:

• To develop a high-precision semantic segmentation method for remote sensing data, overcoming limitations of current approaches.
• To leverage SAM's zero-shot capabilities for structural priors without domain-specific training.
• To create a computationally efficient and scalable solution for practical sensor applications.

Main Methods:

• A dual-branch cooperative network, PriorSAM-DBNet, featuring a Densely Connected Swin (DC-Swin) Transformer for global features.
• An auxiliary branch utilizing SAM for object-boundary masks as structural priors.
• Parameter-efficient Scaled Subsampling Projection (SSP) and Attentive Cross-Modal Fusion (ACMF) modules for feature alignment and ambiguity resolution.

Main Results:

• PriorSAM-DBNet significantly outperforms state-of-the-art methods on ISPRS Vaihingen, Potsdam, and LoveDA-Urban datasets.
• Achieved mean Intersection over Union (mIoU) scores of 82.50%, 85.59%, and 53.36% respectively, with minimal fine-tuning.
• Demonstrated effectiveness in handling dense urban scenarios and complex object boundaries.

Conclusions:

• PriorSAM-DBNet provides a scalable and accurate solution for remote sensing semantic segmentation.
• The method effectively integrates global context with local structural priors, enhancing performance.
• Offers significant potential for rapid feature recognition in time-critical applications such as disaster emergency response.