Image Colorization with Generative Adversarial Network

Abstract

In the past decade, the process of automatic coloring of images has attracted attention in many application areas, including the restoration of aging or degraded images. In the process of color information distribution, this problem is highly ill-posed due to the large degree of freedom. Many recent developments in automatic colorization involve images that contain public topics or data that require high processing, such as semantic mapping, as input. In our approach, we try to use an improved deep convolution generation to combat the network to fully promote the coloring process. We developed a full-convergence generator with multi-layer noise to enhance diversity, maintain realism through multiple layers of conditions, and maintain spatial information through the first step. In the discriminator section, we use an auto-encoder to perform multiple tasks, making the discriminator role more accurate and learning more stable. We also added a GAN-based super-resolution process to create more realistic images. The network is trained through publicly available data sets such as places365. With such a new network structure, the model produces highly competitive performance on the data set. The results of the generated model and the traditional GAN-based image transform depth neural network are compared.