International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 10 | Oct 2025
p-ISSN: 2395-0072
www.irjet.net
Image Inpainting and Outpainting using Deep Learning: A Survey Kindipsingh Mallhi1, Aditi Chhajed2, Ojas Binakye3, Priyansh Katariya4 and Prof. Pramila M. Chawan5 1,2,3,4 B.Tech Student, Dept of Computer Engineering and IT, VJTI College, Mumbai, Maharashtra, India.
5 Associate Professor, Dept of Computer Engineering and IT, VJTI College, Mumbai, Maharashtra, India.
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Image inpainting and outpainting are crucial problems in computer vision, with applications ranging from restoration of historical artworks to modern-day photo editing, privacy preservation, and medical imaging. Traditional approaches such as diffusion and patch-based methods often fail to preserve semantic consistency for large missing regions. Recent advancements in deep learning, especially Convolutional Neural Networks (CNN), Generative Adversarial Networks (GAN), and Partial Convolutional Networks (PConv), have revolutionized this domain. This paper surveys existing approaches, discusses their merits and limitations, and positions our proposed work of applying PConv for image inpainting and outpainting as a robust and semantically consistent solution.
This survey consolidates the progress across these methods, evaluates their strengths and limitations, and establishes the research gap for our project focused on PConv for inpainting and outpainting.
1.2 Problem Image inpainting and outpainting deal with the reconstruction of missing or corrupted regions and the extension of image boundaries while ensuring that the generated content is visually plausible and semantically consistent. Traditional techniques, such as diffusion-based interpolation and patch-based texture synthesis, struggle with large missing regions and lack contextual awareness. Although deep learning has significantly improved results, challenges remain unresolved. Maintaining semantic consistency across diverse scenes, handling irregularly shaped holes, and preserving fine details without blurriness are still difficult. Moreover, models must generalize across different image types—from natural landscapes to medical scans—while balancing reconstruction quality with computational efficiency. These challenges highlight the need for advanced frameworks such as PConv that can dynamically adapt to missing regions and improve both accuracy and stability.
Key Words: Image Inpainting, Image Outpainting, Deep Learning, CNN, GAN, Partial Convolution, Image Restoration
1. INTRODUCTION 1.1 Brief overview of image inpainting and outpainting Image inpainting refers to the process of reconstructing lost or deteriorated parts of an image so that the restored regions are visually plausible and semantically coherent. Outpainting extends an image beyond its original boundaries while maintaining contextual consistency. Both problems are of immense significance in fields like digital heritage restoration, entertainment, e-commerce, and medical imaging.
1.3 Motivation The motivation for this study arises from the growing demand for reliable image completion across multiple domains. In cultural heritage, it is vital for restoring damaged artworks, wall paintings, and manuscripts. In healthcare, medical imaging often requires reconstructing incomplete scans for accurate diagnosis. In everyday applications such as photo editing, privacy preservation, and digital content creation, users expect seamless object removal, background completion, and boundary extension.
Early traditional methods relied on pixel diffusion and patch-based texture synthesis. While these approaches were widely used, they were computationally expensive and lacked semantic awareness, often producing unrealistic or repetitive patterns. The advent of deep learning introduced powerful alternatives such as Convolutional Neural Networks (CNNs), Generative Adversarial Networks (GANs), and Partial Convolutional Networks (PConv), which leverage large-scale datasets and semantic feature learning to produce realistic reconstructions.
© 2025, IRJET
|
Impact Factor value: 8.315
Similarly, industries like entertainment, e-commerce, and gaming rely on visually appealing and contextually accurate image modifications. While CNN and GAN-based models have shown remarkable progress, they still face limitations in stability, handling irregular regions, and computational cost. This necessitates exploration of more efficient and robust alternatives such as PConv.
|
ISO 9001:2008 Certified Journal
|
Page 133