Image inpainting refers to the process of restoring missing or corrupted regions of an image in a manner that is both visually plausible and semantically consistent. It serves as a foundational technique in a wide range of applications, including digi...
Image inpainting refers to the process of restoring missing or corrupted regions of an image in a manner that is both visually plausible and semantically consistent. It serves as a foundational technique in a wide range of applications, including digital photo restoration, object removal, augmented reality, and post-production in the film industry. Despite the remarkable performance of recent deep learning-based inpainting models, their excessive model complexity and computational demands pose significant challenges for practical deployment. This study addresses this limitation by proposing a lightweight image inpainting framework that maintains competitive performance while significantly reducing computational overhead. The proposed system is centered around the incorporation of guidance information, which alleviates the burden on the network to learn all relevant features solely through data. Specifically, guidance features such as edge maps, SIFT orientation maps, and Local Binary Pattern (LBP) descriptors, extracted via traditional computer vision techniques, are introduced into the model to provide explicit structural cues with low computational cost and high interpretability.
The research progresses in two main stages. The initial phase focuses on facial image inpainting, wherein a Laplacian edge map derived from the unmasked regions is integrated into a U-Net-based generator, resulting in improved reconstruction quality. Building upon this, the second phase extends the approach to general image inpainting. To address the greater structural and textural complexity of general images, additional guidance maps including SIFT and LBP are employed alongside edge maps. Architecturally, the generator adopts a lightweight U-Net structure, augmented with Dilated Residual Blocks in the bottleneck layer to expand the receptive field without increasing parameter count. A PatchGAN discriminator is used consistently across both facial and general models to enforce local realism. This hybrid approach, combining traditional feature descriptors with deep generative models, successfully balances efficiency and performance.
Experimental validation was conducted using benchmark datasets, including CelebA-HQ and FFHQ for facial images, and Paris Street View and Places2 for general images. Results demonstrate that the proposed method achieves substantial reductions in parameters and floating-point operations (FLOPs) compared to contemporary large-scale models, while delivering competitive performance in both visual fidelity and quantitative metrics. Crucially, the model achieves these results without relying on additional trainable guidance networks, thereby further reducing computational complexity.
The proposed approach significantly enhances the practicality of image inpainting in resource-constrained environments such as mobile devices, edge computing, and real-time systems, aligning with broader trends in efficient artificial intelligence. This work thus contributes a scalable and deployable solution for high-quality image restoration under real-world constraints.