Image Inpainting based on Visual-Neural-Inspired Specific Object-of-Interest Imaging Technology (2508.12808v1)

Abstract: This study addresses the bottlenecks of conventional holistic image inpainting methods -- susceptibility to informational redundancy and low computational efficiency under occlusions and complex backgrounds -- by proposing a novel three-stage framework: "Specific Object-of-Interest Imaging -- Coarse Structural imaging -- Fine Textural Refinement". First, the Network with Object-Imaging Module (NOIM) precisely delineates target structures from degraded inputs while actively suppressing background interference. Subsequently, the Structural Recovery Module (SRM) employs multi-scale dilated convolutions and feature pyramid fusion to complete shape reconstruction from preliminary imaging. Finally, the Global Detail Refinement Module (GDRM) utilizes sub-pixel convolution with skip connections to map structural priors into high-fidelity RGB outputs. We design a triple-branch hybrid loss (weighted reconstruction, perceptual, and stylistic components) alongside a phased training strategy to enforce structural coherence, semantic consistency, and visual stylization. Experimental validation demonstrates competitive performance against state-of-the-art inpainting models across metrics including SSIM (0.978), PSNR (33.86 dB), MAE (1.605), and LPIPS (0.018), while maintaining robustness in extreme scenarios (low illumination, high noise, multi-object occlusion, motion blur). Theoretical analysis integrated with cognitive neuroscience perspectives reveals profound correlations between the "object precedence perception" mechanism and dynamic feature modulation in visual cortices (V1--V4). This approach not only achieves efficient and precise target-centric imaging but also pioneers interdisciplinary pathways bridging brain-inspired computational frameworks with advanced image inpainting techniques.