From Words to Faces: Automated Suspect Image Synthesis and Retrieval from Criminal Databases

Forensic technology has greatly changed the nature of suspect identification, which plays a critical role in public safety and the delivery of justice. Traditional artist-drawn sketches are highly skill-based, time-consuming, and tend to result in different renderings that may lead to inaccurate depictions of the suspect. Existing approaches also often lack the integration of modern database capabilities, limiting the ability to rapidly cross-reference criminal records. This paper proposes a novel architecture to directly generate suspect images from eyewitness descriptions based on the Stable Diffusion Model. Through an intuitive form, the eyewitness would be able to generate a preliminary image of the suspect and iteratively improve the resemblance by adjusting the salient facial features. A unique feedback mechanism in place allows for iterative improvement that increases the reliability of the generated image. Once finalized, the suspect images are matched against the criminal databases using an optimized FAISS-implemented HNSW indexing for rapid and efficient search. Additionally, our framework provides a “Wanted” poster generation feature that applies disguises to suspect images, assisting law enforcement in the distribution and use of suspect images more effectively in investigations. This comprehensive approach, combining generative models with powerful database search and distribution mechanisms, marks a significant advancement in forensic image synthesis and retrieval.