Everyone is A Forensic Artist: Sketch-to-Photo Transformation for Human Face

Abstract

There are a lot of crimes happening all over the world every day. Among the criminal acts, homicide is the type of crime with a large number of victims. Occassionally there are witnesses who see the incident and remember the face of the criminal. Thus the police ask them to sketch to find out the suspect. Since the human face is the most significant and informative part of the human body, the sketch of the face is used to identify the suspect with high certainty. However, the suspects may change their facial features by makeup, such as putting on glasses or dyeing hair. If a sketch is converted into photographic images with modified facial features flexibly, the investigation of crime might accelerate effectively. Recent research has shown the techniques that transform a sketch of a human face into a photographic image or change the style of a human face according to the designated facial features. However there has not yet been an integrated architecture to transform pencil sketches to photographic images directly with desired facial features. In addition there is no fast and automatic evaluation approach that consider multiple metrics of images jointly. It is hence difficult to optimize and select the setting among a few alternatives efficiently. With the limitation we propose the Forensic GAN a network of performing the sketch-to-photo transformation and image manipulation according to the designated attributes. A voting mechanism with multiple metrics including PSNR SSIM SCC and ERGAS for fast evaluating the image quality jointly was proposed. Accordingly the training settings such as the loss function and the number of epochs can be optimized and selected based on the quality of the synthesized images. The performance of the Forensic GAN was tested and its potential for the real forensic task has been verified with synthesized pencil images and real hand-drawing sketches