Image Colorization Based on Deep Learning

Sammanfattning: With the development of artificial intelligence, there is a clear trend to combine computer technology with traditional industries. In recent years, with the development of digital media technology, many methods for coloring gray-scale images have been proposed. In the past, methods of image colorization were based on color transfer and expansion, requiring manual intervention, and the coloring effect was unsatisfactory. Therefore, the colorization research based on deep learning method has significance and broad application prospects. In recent years, the Generative Adversarial Networks (GAN) have outperformed in the fields of image generation, image denoising, image style conversion, etc., which fully proves the potential of GAN in image processing. Therefore, this thesis uses the GAN to colorize images. With image graying preprocessing in some public data sets such as CIFAR-10, CelebA, etc. Paired feature maps are obtained to train the neural network. The existing networks are difficult to extract global features on high-resolution images. Therefore Non-local Blocks are added to improve the global feature extraction effect. Then two loss functions are added to make the generated images more realistic. Exploring the influence of the color characteristics of the latent space and noise information on the generated images. The result is inspiring. In order to solve the problem that it is difficult to train against the network, the progressive growth training method and the two-time scale update rules are used for training. We verify our network coloring effects and quantify the coloring effects through various evaluation indicators.