Recover Realistic Faces from Sketches

Abstract

Currently, Generative Adversarial Networks (GANs) is considered as the best method to solve the challenge of synthesizing realistic images from sketch images. However, the effectiveness of this method depends mainly on setting up a loss function to learn the mapping between sketches and realistic images. This leads to how to choose an optimal loss function to map them. In this paper, we investigate and propose a loss function that combines pixel-based error and context-based error on a proper ratio to obtain the best training result. The proposed loss function will be utilized to train the generator’s U-Net architecture in greater detail. To convert a drawing to an actual image, the trained architecture will be applied. Based on two metrics that are the Structural Similarity Index (SSIM) and visual observations, the assessment results on the CUHK Face Sketch Database (CUFS), AR database (AR), and the CUHK ColorFERET Sketch Database (CUFSF) prove that the suggested method is feasible.