Enhancing Crime Investigation: Attention-Based GAN for Sketch-to-Portrait Conversion
Sheng Yun, Haoran Xue, Xueqing Zhang, Junkai Zhang, Yijun Zhao · 2024
Generative Adversarial Networks (GANs), have sig-nificantly advanced deep learning by enabling the creation of realistic synthetic data through a dual network of generators and discriminators. This research paper introduces an innovative application of GANs in crime investigation, particularly in sketch-to-portrait conversion to aid suspect identification. Leveraging advancements in deep learning, our study proposes a novel attention-based GAN model, named Sketch2portrait, which sig-nificantly enhances the transformation of forensic sketches into photorealistic portraits. By integrating an attention mechanism with a pre-trained VGGface network, our model overcomes the limitations of existing systems that prioritize aesthetics over realism, thus providing a more reliable tool for law enforcement and civilians. Our contributions include the detailed development of the Sketch2portrait model, an exploration of its inner mechanisms, and an analysis of its attention-based approach. The model's efficacy is demonstrated through rigorous evaluation on a combined dataset of CUFSF, CelebA-HQ, and synthetic sketches, showcasing its potential to revolutionize suspect identification in criminal investigations.