Crowding Sourcing and Pattern Recognition for Identification and Sharing of Thief’s Information

Abstract

Security solutions based on facial recognition are widely employed. However, due to the lack of information and appropriate structures, this technology does not benefit small and mediumsized Brazilian businesses (the victim of most criminal occurrences). This research demonstrates a solution focused on this audience and based on: affordable hardware device; Multilayer Perceptrons (MLP); Cloud Computing (CC) and Crowdsourcing (CS). The main structure was developed with: Python programming language; MySQL as database management system; and OpenCV (Open Source Computer Vision Library) to perform the real-time detection of the faces of the people who are entering the establishment. The images are recorded and later analysed by a second algorithm responsible for returning an image vector. This vector is compared with the others vectors already registered in a database with the list of suspects, returning the percentage of similarity between them. A similarity higher than 70% enable an alert to the manager of the establishment, discreetly, for later consultations of the owners, or by judicial order. The structure works in CC and the filling of the database is done via CS by the establishments. Experiments with MLPs were performed to optimize the recognition process, considering 5 types of MLPs (Backpropagation Standard, Momentum, Weight Decay, Quick propagation; Resilient Propagation); 50 to 500 epochs, 7 to 10 neurons, learning rate of 0.01%, 25-35% validation and 75-65% training. It was performed 155 training processes (total: 19 hours and 31 minutes of test execution time). A maximum accuracy of 94% was reached. The solution can be implemented and integrated to the available Google cloud services.