Production of Color Images From the Sketches Using Generative Adversarial Networks

Abstract

This chapter discusses the application of contingent adversarial networks (CANs) as among the better image-to-image translation task approaches. Traditional models are interested in input image maps to output images with a specific focus on applying attention to this process; CANs involve the application of a loss function as a direction guide toward carrying out the mapping. This universal loss function enables the model to learn and translate better and applies to various tasks that otherwise would need its loss function to be built. The method demonstrates the generality of CANs in performing tasks such as image generation from label maps, object reconstruction from terrain maps, and colorizing grayscale images. One of the strengths of this method is that the network is free from hand-designed feature mapping. This eliminates the need for domain knowledge in designing features and allows the model to produce highly good results. The network can produce acceptable results with little human intervention by eliminating the need for hand-crafted loss functions. This ability provides CANs with wide generalizability to image translation tasks with efficiency and flexibility at no performance cost and with a heavy tailoring burden.