Realistic Sketch-based Face Photo Synthesis using Generative Adversarial Networks (GANs)

Abstract

Facial photo-image synthesis and sketch-based face recognition are highly advantageous, particularly in the fields of security forces and forensics. Furthermore, it makes it more feasible for law enforcement to reduce the number of possible suspects in criminal identification operations. However, since pencil drawings and photographs have different properties by nature, creating a synthesis of photographs based on sketches presents a difficult topic. In the last few decades, generative adversarial network-based systems have achieved enormous advances towards improving the performance of image synthesis. It can speed up identification times while improving matching outcomes by reducing gaps among sketch and photo representations. We perform investigations on the well-known photo-sketch pair database CUHK. First, we demonstrate how a generative adversarial network transforms hand-drawn sketches into realistic photos. Secondly, we employ suspect identification by using the pre-trained VGG16-based feature extractor network and KNN classifier. Our technique focuses on the use of deep learning-based networks, which are well-known for their capacity to process data and extract hierarchical features. The presented image-to-image translation framework minimizes the modality differences between hand-drawn face sketches and color images while improving visual quality. Tests on sketch-photo matching demonstrate significant improvements over current state-of-the-art methods on the challenging task of matching sketches with corresponding photos.