Toward U-Net-based GANs for Diverse Facial Image Synthesis from Sketch

Abstract

Face physical changes may result from aging, surgery, or disguise. The criminal suspects conceal their identity with false appearances such as wearing a wig, glasses, beard and mustache. This research benefits the generation of various fictitious appearances of the suspects or facial changes of lost persons. The technique of GANs is applied for synthesizing a color image from a sketch. The output image can be varied in five facial attributes with a toggle: bald, makeup, straight hair, wearing glasses, beard and mustache. The approach enhances the Generator of StarGan2 with the U-Net architecture. The experiments were carried out to evaluate the performance of the proposed model compared to that of StarGan2. FID scores are used for measuring the quality of the generated images. The FID scores measured on the test data reported about 40% less than that of the baseline model and the synthesized images with varied facial attributes look natural and realistic.