Edge guided image inpainting based on deep learning

Image inpainting is the process of reconstructing damaged or missing regions in an image by utilizing the available image information. Although deep learning techniques have enabled significant progress in this field, accurately reconstructing the structural and textural properties of severely corrupted images continues to pose a significant challenge. In this study presents enhancements to improve the image inpainting performance based on the Conditional Texture and Structure Dual Generation (CTSDG) architecture. The CTSDG architecture consists of two separate streams: a structural stream dedicated to capturing the overall layout and structure of the image, and a texture stream specialized in detailed texture synthesis. Various innovative modules have been utilized to strengthen the relationship between these two streams. One such module is the Gated Feature Fusion with Bidirectional Squeeze and Excitation (Bi-SE-GFF), which enhances the feature fusion process through bidirectional squeeze and excitation operations in both directions. Furthermore, the Enhanced Contextual Feature Aggregation (ECFA) module improves the outputs by aggregating multi-scale contextual features. Additionally, the Dual Attention Mechanism integrated with Selective Kernel Feature Fusion (SKFF) combines complementary features, allowing for more coherent structure and texture generation from corrupted inputs. Quantitative experiments demonstrate that the proposed approach achieves a 2.69% improvement in PSNR for 0-20% mask ranges and a 1.43% improvement in SSIM for 40-60% mask ranges compared to the baseline. Qualitative results indicate that the missing regions in severely corrupted images can be reconstructed with significantly higher visual quality.