Single-character jiandu images are essential for large-scale standardized databases and deep learning training. However, segmented images often suffer from character loss, uneven textures, mildew, and low stroke-background contrast. Existing spatial Fourier convolution-based restoration methods reduce detail loss and texture blurring but remain limited in efficiency, structural completeness, and spectral stability. To address these issues, we propose a restoration method based on cross-domain debiased Fourier fusion and parameterized coordinate querying. Adaptive parameters are generated only for damaged regions, avoiding simultaneous synthesis of intact and missing pixels. Coordinate-based pixel querying enables precise reconstruction. The spatial-frequency complementary module fuses spatial and frequency information, enhancing robustness to complex noise while preserving character semantics and structure. Debiased Fourier fractional convolution suppresses spectral offset through learnable range transformation and adaptive cropping, thereby reducing abnormal color blocks and texture contamination during restoration.Experiments on jiandu and Paris StreetView datasets demonstrate superior inpainting accuracy, structural fidelity, and adaptability.
Lu et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: