Beyond Monochrome: Realistic Greyscale Image Colourisation Using Deep Learning

Chan, Edison Yoong Qi (2026) Beyond Monochrome: Realistic Greyscale Image Colourisation Using Deep Learning. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text RDS_Edison Chan Yoong Qi_Fulltext.pdf Restricted to Registered users only Download (2MB)

Abstract

Automated grayscale image colourisation is a challenging yet impactful task in computer vision, with applications spanning from historical restoration, medical imaging to digital media enhancement. While manual colourisation remains labor-intensive, recent advances in deep learning offer scalable solutions. This project develops a hybrid Convolutional Neural Network-Transformer model to predict realistic colour distributions from greyscale inputs, using the LAB colour space for decoupled luminance (L) and chrominance (A/B) processing. Convolutional Neural Networks (CNNs) are chosen for their proven ability to capture local texture patterns and spatial hierarchies efficiently. However, to address limitations in long-range colour consistency like sky gradients and ocean colour, we integrate Transformer layers for global context awareness. This hybrid approach balances computational efficiency with perceptual quality, which in theory will outperform CNN methods and other algorithms in the market. The proposed hybrid model aims to achieve realistic greyscale image colourisation while tackling the weaknesses of CNN model by implementing a transformer layer.