Efficient Detection of Disguised Faces Using Photos/Sketches from Low-Quality Surveillance Footage

Face recognition in surveillance footage plays a pivotal role in criminal investigations. However, challenges emerge when faces in surveillance videos are of low resolution, making criminal detection less feasible. While certain techniques enhance these low-resolution images to clearer, high-resolution versions, they still fail when confronted with disguised individuals. This creates a pressing need for a more adaptable solution capable of identifying faces that are both in disguise and of low resolution. Additionally, in some cases, only sketches of suspects are available instead of images. Our research introduces a comprehensive system that identifies criminals in disguise from low-resolution video streams and map it to either his/her photo or the sketch based on availability in the database. Notably, our system is designed for efficient deployment on edge nodes, demonstrated by its smooth operation on a server with just a 1GB RAM-enabled CPU, 30GB storage, and utilizing a t2.micro instance. This research represents a step forward in real-world deployable solutions for identifying disguised criminals from surveillance data.