Generative Adversarial Multi-Task Learning for Face Sketch Synthesis and Recognition

Face sketch synthesis and recognition have wide range of applications in law enforcement. Despite the impressive progresses have been made in faces sketch and recognition, most existing researches regard them as two separate tasks. In this paper, we propose a generative adversarial multitask learning method in order to deal with face sketch synthesis and recognition simultaneously. Our framework is based on generative adversarial networks (GAN), in which an improved deep network named residual dense U-Net is used as generator to synthesize face sketch image and a multi-task discriminator is designed to not only guide the generator to produce more realistic sketch image, but also extract discriminative face feature. In addition, except the common adversarial loss, the perceptual loss and triplet loss are adopted for the learning of generator and discriminator, respectively. Compared with the state-of-the-art methods, the proposed method obtains better results in terms of face sketch synthesis and recognition.