Image Colorization using Generative Adversarial Networks and Transfer Learning
Paper ID : 1056-MVIP2020
leila kiani *1, Masoud Saeed2, Hossein Nezamabadi-pour3
1Computer Engineering Department Shahid Bahonar University of Kerman Kerman, Iran
2Computer Engineering Department Shahid Bahonar University of Kerman
3Intelligent Data Processing Laboratory (IDPL) Shahid Bahonar University of KermanŲŒIran
Automatic colorizing is one of the most interesting problems in computer graphics. During the colorization process, the gray one-dimensional images are converted to three-dimensional images with colored components. As a typical technique, Convolutional neural networks (CNNs) have been well studied and used for automatic coloring. In these networks, the information that is generalized over in the top layers is available in intermediate layers. Although the output of the last layer of CNNs is usually used in many applications, in this paper, we use a concept called Hypercolumn derived from neuroscience to exploit information at all levels to develop a fully automated image colorization system. There are not always millions of data available in the real world to train complex deep learning models. Therefore, the VGG19 model trained with the big data set of ImageNet is used as a pre-trained model in the generator network and the hypercolumn idea is implemented in it with DIV2K datasets. We train our model to predict each pixels color texture. The results obtained indicate that the proposed method is superior to competing models.
GAN, Hypercolumn, Image Colorization, Transfer Learning
Status : Paper Accepted (Oral Presentation)