Tackling Rare False-Positives In Face Recognition: A Case Study

In this study, we take on one of the most common challenges in facial recognition, i.e. reducing the False Positives in the recognition phase, through studying performance of a standard Deep Learning Convolutional network in a real-life, real-time, and large-scale identity surveillance application. This application involved designing a queue management system that uses facial recognition, for an airport in the UK. Our approach was to capture the faces of passengers as they enter through Boarding Pass Gates (BPG) and as they exit Security Gates (SG). Thereafter, we compare the faces captured, within a fifteen minute window, from BPG against the ones from SG. When there is a match, we are able to calculate the time that someone has spent inside the security area, using the capture time of matched face. We call this the security queue time. Like any other facial recognition application, we have to deal with reducing the number of false positives, i.e. incorrectly matched faces. In this application false positives are statistically rare events. That is, the same or similar pair of images is unlikely to occur in a foreseeable time. To deal with this problem, we utilized several approaches including applying a second layer of detection using the Dlib library [3] to improve the quality of the detected faces. Specifically, by taking advantage of Dlibs Facial Landmarks, we created a scoring system similar to Dlibs, to choose the best frontal pose from amongst all faces attributed to a single person. Our large-scale trials show that this approach does measurably reduce the rate of false positives in such systems.