Forensic Automated Composite Evaluation System Using Deep Learning for Sketch-to-Photo Suspect Identification

Facial sketching in forensic science is an important area in criminal investigations, especially when there is no photographic evidence of the suspect available. Conventional facial sketching is a time-consuming and subjective process that requires expertise in forensic sketching. However, it is not amenable to automation with digital identification systems. In this paper, we propose a Forensic Automated Composite Evaluation System (FACES) that allows investigators to create facial composites using a drag-and-drop interface and automatically search the law enforcement image database for a match. Our proposed system uses a deep learning-based pipeline that includes facial preprocessing with MTCNN, feature extraction with a pre-trained InceptionResNetV1 model, and similarity comparison with cosine similarity. The proposed system provides a ranked list of possible suspects, which is much faster than the manual process. The experimental results show high accuracy for full facial sketches and reasonable results for partial facial sketches, which validate the proposed system as an effective forensic identification tool.