Paired Face Sketch-to-Photo Synthesis using Generative Adversarial Networks

Face Sketch-to-Photo Synthesis (FSPS) is a crucial problem to solve nowadays as it is used in many real-time applications like criminal case investigation, law enforcement, and digital entertainment. Earlier solving FSPS has been a challenging task due to the vast domain discrepancies between the face sketch images and photo images. However, with the invention of Generative Adversarial Networks (GANs), various image processing problems, such as image in-painting, image denoising, image colourization, and image draining, have been solved by treating them as Image-to-Image (12I) translation tasks. In this paper, the FSPS problem has been solved by treating it as a paired I2I translation task. Six State-Of-The-Art (SOTA) I2I translation methods named Pix2Pix, CycleGAN, PAN, AttentionGAN, CSA-GAN, and PCSGAN are implemented for FSPS over the four face sketch-photo datasets like AR, XM2VTS, CUHK, and CUFSF. Four famous image quality assessment metrics, Mean Squared Error (MSE) metric, Peak Signal-to-Noise Ratio (PSNR) metric, Structural Similarity Index Measure (SSIM) metric, and Learned Perceptual Image Patch Similarity (LPIPS) metric, have been utilized to evaluate the performance of the SOTA methods to select the best suitable I2I method to address the paired FSPS problem. In addition to the results analysis, a separate comparative analysis of the performance of the methods, along with their merits and demerits for all four datasets, are explained.