Image Inpainting Based on Patch-GANs

Abstract-Image inpainting is a crucial task in computer vision that focuses on reconstructing missing or corrupted parts of an image while maintaining structural consistency and visual realism. Traditional inpainting methods, such as diffusion-based and exemplar-based techniques, often struggle to restore fine textures and complex structures, leading to blurry and unrealistic results. The advent of Generative Adversarial Networks (GANs) has significantly enhanced image inpainting by learning to generate plausible image content. However, conventional GAN-based models emphasize global image coherence while neglecting finer local details, causing inconsistencies in high-texture regions. To overcome these limitations, PatchGAN-based inpainting evaluates image realism at the patch level rather than analyzing the entire image as a whole. This technique employs multi-scale discriminators that ensure improved texture synthesis and structural continuity at different spatial resolutions. Experimental studies reveal that PatchGAN-based models outperform conventional GAN-based methods in terms of perceptual quality, Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM), producing sharper and more realistic image restorations. This review explores the advancements in PatchGAN-based inpainting, highlighting its benefits, architectural components, and future research directions to further enhance image reconstruction quality.