Gray to Glow: Colorize Grayscale Images Using Deep Learning

Abstract

The automatic colorization of grayscale images is a difficult computer vision problem because there are multiple possible colorizations for each grayscale image. To solve this issue, the Gray to Glow project aims to create an automatic deep learning-based system which transforms a grayscale image into a color image without the need for user scribbles or reference images, or manual intervention. The proposed system is based on the use of a CNN for color prediction in the CIE Lab color space, leaving the L channel unchanged and estimating the missing channels in a classification oriented way that minimizes the number of desaturated outputs while maximizing color vividness. It combines pretrained ECCV16 and SIGGRAPH17 colorization models, image preprocessing, output reconstruction, and it has a Flask web app to make it accessible for both command-line and browser. Experimental results presented in the thesis confirm that the system could produce visually meaningful colorized outputs, it was able to process outputs that were stable, and it was easy for users to use the system in enhancing grayscale images. The project illustrates how a deep learning model can be applied to develop a real-world and deployable system capable of image colourisation for image restoration, digital archiving and media enhancement applications. A key point is that computer vision applications can be extremely useful in colorizing images that are previously only in grayscale.One of the interesting aspects is that an image, when it is only in grayscale, can be very helpful in computer vision for image colorization, and it can be done very efficiently with deep learning techniques.

Introduction

This paper presents Gray to Glow, a fully automatic deep learning-based image colourisation system that restores realistic colours to grayscale images. Image colourisation is a challenging computer vision problem because multiple plausible colours can correspond to the same grayscale image, making the task inherently ill-posed. Traditional regression-based methods often produce dull or desaturated images by averaging possible colour values. To overcome this limitation, Gray to Glow employs pretrained deep convolutional neural networks that infer realistic colours from image structure, texture, and semantic context.

The proposed system is designed to enhance grayscale images for applications such as photo restoration, cultural heritage preservation, digital archiving, and visual media enhancement. By learning colour distributions from large collections of natural images, the model can assign visually convincing colours to objects such as skies, vegetation, and human skin without requiring any user input. The project contributes a complete end-to-end colourisation pipeline, including image preprocessing, model inference, reconstruction, and deployment through both command-line and web-based interfaces.

The literature review traces the evolution of image colourisation from user-assisted and regression-based methods to modern deep learning approaches. Recent techniques include Generative Adversarial Networks (GANs), diffusion models, and transformers, which improve realism, semantic consistency, and colour diversity. However, these advanced models often require greater computational resources and are less suitable for real-time deployment. Rather than developing a new architecture, Gray to Glow integrates established pretrained models from ECCV16 and SIGGRAPH17, providing a practical balance between image quality, inference speed, and ease of deployment.

The methodology is based on a modular system architecture. Images are first preprocessed by converting them into the CIE Lab colour space, where the luminance (L) channel is retained while the missing chrominance (a and b) channels are predicted by the neural network. The system performs image loading, resizing, normalization, colour prediction, image reconstruction, and result delivery through a Flask-based web interface. The implementation consists of eight functional modules that improve maintainability and allow future extensions such as batch processing and mobile deployment.

A key feature of the system is its classification-based colour prediction strategy, which predicts colour categories instead of directly regressing colour values. This approach reduces the washed-out appearance common in regression methods and produces more vibrant and semantically meaningful colourisations. The ECCV16 model contributes techniques such as class rebalancing, dilated convolutions, and annealed-mean decoding, while SIGGRAPH17 provides improved scene understanding and finer contextual reasoning.

Experimental results demonstrate that Gray to Glow generates realistic, visually appealing colour images with efficient inference suitable for real-time applications. The ECCV16 model produced more vivid colours than earlier methods and achieved a good balance between colour vibrancy and consistency through annealed-mean decoding, with an optimal temperature parameter of 0.38. Human evaluation further showed that approximately 32.3% of colourised images were perceived as real during a colourisation Turing Test, indicating the effectiveness of the system. Comparisons between ECCV16 and SIGGRAPH17 showed that ECCV16 offered higher colour vibrancy and excellent real-time performance, while SIGGRAPH17 provided stronger scene understanding.

Conclusion

This paper revisits the Gray to Glow study for automatic image colorization of gray images with deep learning. The biggest advantage of the project is the practical end-to-end system that integrates the work of CIE Lab preprocessing, the pretrained models for the colorization with the ECCV16 and SIGGRAPH17 competition, the modular inference stages and the deployment for the user via command-line and web interfaces. The reported results demonstrate that the system is working correctly, stable, and can produce meaningful results in the process of enhancement of grayscale images in the form of colorized output results. The literature analyzed in this, reveals that with CNN-based image colorization, the study has evolved from regression-based models to the GANs and diffusion models, which present some pros and cons regarding the realism, diversity, and computational complexity of colorization.The literature reviewed here demonstrates the shift from regression-based CNNs towards GANs, diffusion models, and transformer-based approaches, each with its own set of trade-offs in terms of realism, diversity, and computational efficiency. As such, Gray to Glow should be interpreted as a contribution to research in implementation, a search that is substantive and geared toward practice rather than a new learning architecture. There are a number of areas where future work is possible. This system can be further improved by quantitative assessment on benchmark images, incorporation of more recent transformer- or diffusion-based colourisation models, better processing of less crisp historical photos, and larger perceptual user studies. Other topics that could be covered include mobile deployment, edge optimization, and other domain-specific colorization settings, including archives, medical imaging, or remote sensing.

References

[1] Shivaprasad Satla, “A Deep Learning Framework for Semantic Colorization of Synthetic Aperture Radar Images” , in Springer Conference, Volume:84, 2026. [2] Jinyi Luo, Xi Li, “Analysis and Comparison of Image Colorization Based on Machine Learning with PyTorch and ChromaGAN”, science direct conference, 2025. [3] Muhammad Shafiq, Muhammad Asad, Asra Aslam, \"A Comprehensive Review of Modern Image Colorization\", in Springer Volume:84, 2025. [4] Richard Zhang, Jun-Yan Zhu, Phillip Isola, Alexei A. Efros, \"Transforming Color: A Novel Image Colorization Method\", in Science Direct Volume:248, 2024. [5] “DDColor: Towards Photo-Realistic Image Colorization via Dual Decoders”, Sungmin Kang, Myeongsu Kang, Donghyeon Cho, Sang Min Yoon, In So Kweon, 2023. [6] Arash Vahdat, David J. Fleet, \"BigColor: Colorization using a Generative Color Prior for Natural Images\", in springer nature link, 2022 . [7] Haojun Xu, Bing Li, Jianmin Jiang, \"ColorFormer: Image Colorization via Color Memory Assisted Hybrid-Attention Transformer\", International conference paper link, in 2022. [8] Aditya Kumar, Richard Zhang, Jun-Yan Zhu, Colorization Transformer, in 2021. [9] Miguel Perelló Nieto, Javier Ventura, Vitoria Pereira, \"ChromaGAN: Adversarial Picture Colorization with Semantic Class Distribution\", in 2020. [10] Chaoqun Su, Weidi Xie, Tao Kong, \"Instance-Aware Image Colorization\", IEEE Conference, in 2020. [11] Sukrut Oak, Young Chol, Song Da Sun, “Improved Recoloring of SAR Satellite Images, ” Stanford University, vol. 141, p. 112963, 2020, doi: 10.1/16.eswa.2019.112963. [12] Kaushal Kishor, Chirag Sharma, Himanshu Sharma, Manmohan Fulara & Aditya Yadav, “SAR Image Colorization for Comprehensive Insight Using Deep Learning”, in 2025, springer, chapter/10.1007/978-981-96-3725-6_22 [13] Cheng, Z., Yang, Q., Sheng, B.: Deep colorization. In: Proceedings of the IEEE International Conference on Computer Vision, 2015, 415–423. [14] Zhang, R., Isola, P., and Efros, A. A. Colorful Image Colorization. European Conference on Computer Vision (ECCV), 2016. [15] Wang, X., Gupta, A.: Unsupervised learning of visual representations using videos. In: Proceedings of the IEEE International Conference on Computer Vision. 2015, 2794–2802. [16] Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected arXiv preprint arXiv:1606.00915, 2016. [17] Patterson, G., Hays, J.: Sun attribute database: Discovering, annotating, and recognizing scene attributes. In: Computer Vision and Pattern Recognition (CVPR), 2022 IEEE Conference on, IEEE 2751–2758.

Copyright

Copyright © 2026 Mukashifa Fatima, Razia Begum . 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.