Comparative assessment of DCGANS and autoencoder-based models for image inpainting

Abstract Image inpainting aims to fill missing or corrupted regions in images with semantically coherent content. This paper presents an expanded comparative study evaluating deep generative adversarial networks (GANs) and autoencoder models for image inpainting. We conduct extensive experiments and analysis on Places365 dataset images with missing regions. The literature review is significantly expanded to cover the latest advancements in deep learning based inpainting. Methodology details are enhanced with additional architectural specifications, training procedures, evaluation metrics, and ablation studies. Comprehensive quantitative metrics, qualitative visual comparisons, user studies, and ablation studies provide novel insights into model capabilities and limitations. Results analysis is also broadened to provide more in-depth discussion of performance trade-offs. Our findings reveal complementary strengths of GANs in perceptual quality and autoencoders in reconstruction accuracy. Hybrid approaches can balance these properties for high fidelity inpainting. The expanded study offers rich insights to guide optimal selection and design of deep generative architectures for advancing image inpainting across diverse applications.