Improved CycleGAN for Synthesis and Recognition of Face Photo Sketches

Facial photo-sketch synthesis and recognition is used in law enforcement to find key suspects. Nevertheless, the extant methodologies predominantly concentrate on the modal dissimilarities between photographs and sketches, and thus fail to fully consider identity information and detail preservation in the recognition task, which limits the recognition accuracy of synthesized images. For the purpose to deal with these issues, this paper puts forward a proposal for an identity-aware and detail-enhanced recurrent generative adversarial network. In this method, we optimize the generated sketches and photographs to retain more identity information and texture details by introducing perception loss and detail loss. Furthermore, by utilizing an attention strategy to boost the Visual Geometry Group(VGG) model's performance, we strengthen the recognition network and raise the accuracy of face picture sketch recognition. Finally, we perform experiments on the CUHK Face Sketch(CUFS) and CUHK Face Sketch FERET(CUFSF) datasets, and the results of the experiments are evidence of the effectiveness of our method.