Forensic Sketching using GANs

Facial sketching plays a crucial role in criminal investigations by converting eyewitness descriptions into visual representations of suspects. However, manual sketches are often inconsistent, subjective, and heavily dependent on artistic skill. Recent progress in deep learning, particularly Generative Adversarial Networks (GANs), has enabled automatic sketch generation with superior accuracy and reduced bias. This paper surveys the evolution of GANbased forensic sketch generation, emphasizing models such as Pix2Pix, CycleGAN, and StyleGAN. A quantitative comparison of these models using metrics like SSIM, MSE, and Feature Adjustability Score highlights StyleGAN’s superior realism and adaptability. The study further discusses challenges such as dataset bias, model interpretability, and fairness in sketch generation. Future work recommendations include dataset diversification, fairness-aware learning, and integration of diffusion-based models to enhance reliability in real-world law enforcement applications.