Sketch & Fetch: Draw a logo sketch to fetch the suspect from surveillance footage

Abstract

Logo Sketch-assisted Suspect Retrieval (LSSR) enables rapid identification of individuals based on eyewitness-drawn logo sketches, especially when facial details are obscured or unavailable in surveillance footage. The proposed Sketch-Fetch Generative Adversial Network (SF-GAN) translates sketches into realistic logo images, while the enhanced E-YOLOv7 (Elite-You Only Look Once version 7) detects logos on clothing in real-time. Local self-similarity descriptors with Euclidean matching are used to retrieve the query person. SF-GAN is trained on 100 sketch-image logo classes and shows adaptability to new designs. It achieves a low FID (Fréchet Inception Distance) score of 0.2, indicating high-quality generation. The system is tested on benchmark datasets under challenging conditions, including blur, occlusion, and low resolution. Achieving 95.6 % accuracy, the LSSR framework outperforms state-of-the-art approaches in logo-based suspect retrieval.