Image Restoration using Deep Learning Techniques

Abstract

In the modern era, due to the emergence of various technologies, most of the human work is now being performed by the computer system. The computer’s capacity to make everything possible is increasing as by the time. Photos are used to capture or freeze the moments in one’s life. We can embrace those moments at any time by looking at the pictures. It is natural that, as time passes by, these photos gets damaged due to environmental conditions that leads to loss of our important moments. Hence, preserving the photos is as important as taking them. The process of taking corrupt or noisy image and estimating the clean, original image is image restoration. Many forms of noise such as motion blur, camera misfocus etc., increases the complexity to restore the image. Image corruption comes in varying degrees of severity, the complexity of restoring photos in real-world applications will likewise vary greatly. Also, manual restoration is time consuming leading to lots of work to be piled up. To increase the capability of restoring old images from various defects, we must address several degradations intermingled in one old photo, such as structural defects like scratches and dust spots, and unstructured defects like sounds and blurriness. Furthermore, we may use a different face refinement network to restore small details of faces in ancient pictures, resulting in higher-quality photos. The aim of the work is to create a image restoration system that will be used to restore the images irrespective of the type of noise. In this paper, we present a model that would take image as an input and remove all the noises present in it to give a clean and restored image.

Introduction

The paper proposes a deep learning-based image restoration system for repairing old and damaged photographs. Old photographs often deteriorate due to environmental factors such as sunlight, humidity, temperature changes, chemical degradation, scratches, folds, water stains, fading, blur, and missing regions. Since these photographs preserve valuable memories and historical information, restoring them is important. The proposed approach aims to provide a simple, effective, and accessible solution using modern digital image processing techniques.

The restoration framework is based on Variational Autoencoders (VAEs), where two separate VAEs are trained to map old and clean photographs into different latent spaces. A latent space translation network learns to transform degraded images into clean ones using synthetic paired data, enabling the model to generalize well to real-world old photographs. The system is also designed to handle multiple types of degradation simultaneously, including structural defects (scratches, dust spots, missing regions) and unstructured defects (noise and blur). Additionally, a dedicated face refinement network enhances facial details, producing higher-quality restored images.

The literature survey reviews several image restoration techniques, including CNN-based denoising, Total Variation (TV) methods, ADMM optimization, shrinkage fields, and Noise2Noise learning. While these methods achieve good restoration performance for specific degradation types, they often struggle with complex real-world defects or require clean training data, highlighting the need for a more comprehensive restoration framework.

The proposed system is implemented using Python, Pillow (PIL), and PyTorch. Pillow is used for image preprocessing and manipulation, while PyTorch provides efficient deep learning capabilities with GPU acceleration and strong integration with scientific computing libraries such as NumPy.

The restoration process consists of four main stages:

1. Image Acquisition: The damaged photograph is provided as input.
2. Pre-processing: Standard image enhancement and preparation techniques are applied.
3. Denoising: Noise is reduced while preserving important image details and improving the signal-to-noise ratio (SNR).
4. Deep Latent Space Translation: Two VAEs encode old and clean images into latent spaces, where a translation network reconstructs a clean version of the degraded image. The encoder compresses the image while retaining essential information, and the decoder reconstructs it to its original size with reduced noise, blur, and scratches.

Experimental results demonstrate that the proposed method successfully restores degraded photographs by removing scratches, reducing noise, correcting blur, and reconstructing damaged regions, producing visually improved images. Overall, the system provides an effective and automated solution for restoring old photographs, helping preserve historical records and personal memories while reducing the time and effort required for manual restoration.

Conclusion

Image restoration is a difficult topic to solve. The primary goal of this project is to do a comparative analysis. Though each strategy has its own way of dealing with the situation and its own set of benefits and drawbacks. The use of the methodologies is governed by the comprehension, requirement, and standard of the output required, as shown by the previous explanations. The descriptor that results is compact, discriminative, and efficient. We have shown clear data that have drawn a reduction in complexity and an increase in the capacity to learn very complex aspects since the introduction of this method. This method, we hope, will be effective for future jobs involving the extraction of strong discriminative characteristics.

References

[1] Noise2Noise: Learning Image Restoration without Clean Data: https://arxiv.org/pdf/1803.04189.pdf . [2] https://analyticsindiamag.com/restore-old-photos-back-to-life-using-deep-latent-space-translation-pytorcg-python-demo/ . [3] Shrinkage Fields for Effective Image Restoration: https://openaccess.thecvf.com/content_cvpr_2014/papers/Schmidt_S hrinkage_Fields_for_20 14_CVPR_paper.pdf . [4] Image restoration segmentation using watershed method for basic medical applications: https://ph02.tcithaijo.org/index.php/past/article/view/244125/165992 [5] IMAGE RESTORATION FUNDAMENTALS AND ADVANCES [6] BY Bahadir k Gunturk and Xin Lee [7] On demand learning for deep image restoration: http://vision.cs.utexas.edu/projects/on_demand_learning/. [8] Image Restoration using Machine Learning: http://gpbib.cs.ucl.ac.uk/gp-html/Chaudhry_thesis.html. http://prr.hec.gov.pk/jspui/handle/123456789/4816 . [9] TV: A New Method for Image Restoration in the Presence of Impulse Noisehttps://ieeexplore.ieee.org/document/7299175 . [10] Poisson noisy image restoration via overlapping group sparse and nonconvex second- order total variation priors https://journals.plos.org/plosone/article?id=10.1371/journal.pone.025 0260 . [11] Convolutional Neural Network Combined with Half-Quadratic Splitting Method for Image Restoration https://www.hindawi.com/journals/js/2020/8813413/ . [12] Shrinkage Fields for Effective Image Restoration https://ieeexplore.ieee.org/document/6909751 . [13] Noise2Noise: Learning Image Restoration without Clean https://proceedings.mlr.press/v80/lehtinen18a/lehtinen18a.pdf . system and data science. She has done certifications in cisco python and AWS academy cloud foundation.

Copyright

Copyright © 2026 Akurathi Aravinda, Challagulla Yoshitha, Kakarla Meghana, Kandula Sreeja, B. Tejaswi. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.