Dual Branch Graph Attention Mechanism Algorithm for Pelvic Image Restoration

Pelvic X-ray images often suffer from local defects or blurriness caused by surgical tool occlusion, screw implantation, and other factors, which can impair visualization quality and diagnostic accuracy. Current existing restoration methods usually struggle to balance global information and local details in complex occluded scenes, limiting the effectiveness of restoration. To address this challenge, we propose a novel pelvic X-ray image restoration algorithm based on a dual-branch graph attention mechanism (DBGA-PIR). The proposed method follows a two-stage framework: coarse and refined restorations. In the coarse restoration stage, a local feature branch (CNN) and a global feature branch (Vision Transformer) extract detailed local features and global contextual information, which are fused using a global-local feature coupling module (GLFCM) to generate an initial restoration. The refined restoration employs a shallow encoder-decoder network combined with a graph attention mechanism (GAM) to further enhance image quality by modeling spatial relationships and recovering intricate details. The proposed algorithm is validated on a private dataset of pelvic X-ray images with occlusions, demonstrating superior performance across multiple metrics. In the coarse restoration stage, the algorithm achieves MSE, PSNR, SSIM, and UQI values of 12.2156, 38.056, 0.9439, and 0.9256, respectively. In the refined restoration stage, these metrics improve significantly to 10.6548, 47.764, 0.9825, and 0.9889, outperforming existing image restoration methods. These results highlight the efficiency of DBGA-PIR in restoring image integrity under complex occlusion conditions, offering valuable support for clinical pelvic image analysis.