Image recognition using artificial intelligence (old photo colourisation using deep neural network)

• Main Author: Lin, Jinmin
• Other Authors: Yap Kim Hui
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering; Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/149734
• Institution: Nanyang Technological University

This final year report focus on the design of fully automatic image colourisation using a Deep Neural Network. Before entering the era of colour photos, many stories and memories were recorded in black and white images. The vibrant and realistic colourisations of an image can significantly bring back memories and feelings from the old black and white photos. However, the general colourisation relies on heavy user interaction, such as requiring manual input or photoshop to edit pictures. Otherwise, it may result in desaturated colourisations. Thus, we propose a fully automatic image colourisation system to convert a grayscale image to a colour image that is vibrant and realistic. To achieve this, we have developed a new U-Net generator network based on deep learning techniques that can perform robust image recognition and image to image translation, and an ECA-Net (Efficient Channel Attention) module was added for enhancement of image colourisation. The whole proposed network consists of two neural networks, which are the new U-Net generator and PatchGAN discriminator. The PatchGAN discriminator is used to optimise the generator performance by distinguishing generator's image from the real image. The whole network is trained on 1.28million images from ImageNet and our collected dataset of 4500 old photos. The experimental results on the proposed new U-Net generator show a better performance for old photos than other methods. This report first introduced the knowledge of machine learning and deep learning method, which is the whole project's foundation. Related literature is reviewed, and different networks that can achieve the colourisation function are explored and compared. Following that, this report focuses on the system implementation, such as network design, environment setup, dataset preparation, network establishment and training method. Lastly, the network has demonstrated image colourisation with robust performance that can generate realistic colour on photos. We also discussed the further improvement for generating a higher resolution image by applying other's latest network.