Forensic Sketch-to-Face Recognition via Facial Attributes Synthesis using GAN

Abstract

Sketch-to-Face recognition is a critical problem in forensic investigation and law enforce- ment. Various methods have been implemented that identify sketches described by eye- witnesses, but their performance degrades to lower quality image generation when using unpaired sketch and face images dataset. Minimum research have been done to apply deep learning algorithms for Sketch-to-Face recognition, as it limits due to scanty im- ages available which is insufficient for training a network. As sketch and face images have different texture attributes, it is difficult to identify the sketch corresponding face. The proposed approach is designed to formulate this recent problem. This attempt in- cludes sketched face synthesis from attributes and reconstructing them for generating high resolution realistic facial images using Generative Adversarial Networks. Secondly, gen- erated face image/s are recognized from police mugshot database. Experiments related to Sketch-to-Face transformation and face recognition are performed on openly available CELEBA and LFW datasets.