An evidence-prior collaborative image inpainting method based on counterfactual-inspired prototype contribution selection

The inpainting of complex damaged images relies not only on contextual inference from the visible regions of the current image, but also on the effective introduction of external reference priors when single-image evidence is insufficient. However, cross-image reference information and the image to be inpainted often suffer from structural, textural, and semantic mismatches, and direct prior injection may lead to erroneous reference propagation and unreliable content generation. To address this problem, this paper proposes an evidence-prior collaborative image inpainting method with counterfactual-inspired prototype contribution selection. The proposed method jointly models current-image evidence and cross-image multi-domain prototype priors, thereby enabling reliable prior screening and adaptive injection of external reference information. Specifically, a local topology-aware structural continuity enhancement module is introduced to enhance structural continuation modeling around damaged boundaries. Through missing-region-guided boundary-global collaborative retrieval and multi-domain prototype aggregation, external priors at the structural, textural, and latent levels are constructed. A counterfactual-inspired prototype contribution selection strategy is introduced to compare the response difference between candidate-prototype states and evidence-baseline states, suppressing low-contribution or misleading reference information. Furthermore, an uncertainty-guided geometric prototype calibration and fusion strategy is adopted to achieve adaptive collaboration between external priors and current-image evidence. Experiments on Places2, Paris, and DhMurals-Inpainting-dataset demonstrate the effectiveness of the proposed method. On DhMurals-Inpainting-dataset, the method achieves PSNR, SSIM, and LPIPS values of 30.251, 0.9638, and 0.0219, respectively. Ablation experiments and visual analyses further indicate its role and contribution in prototype prior injection, adaptive fusion, and structural-textural consistency recovery in complex damaged regions.