Applying a Novel Feature Set Fusion Technique to Facial Recognition

Abstract

An important use of facial recognition is the Take Me Home project. In this project, people with disabilities (PWD) are voluntarily registered so that law enforcement officers can identify them and bring them home safely when they are lost. In an application like Take me Home, optimization of person recognition is of prime importance. While facial recognition models have seen huge performance gains in recent years through improvements to the training process, we show that accuracy can be improved by combining models trained for different recognition objectives. Specifically, we find that the accuracy of facial recognition model is higher when its output is fused with the output of model trained to recognize specific attributes such as hair color, age, lighting, and picture quality. The fusion is performed with a linear regression that can be applied to countless other machine learning tasks. The main contribution of our methodology is the mathematical formulation and a neural network using the Inception Net architecture that enables the recognition of the person using up to 40 attributes. In addition, we designed a framework that uses a joint linear regression scheme to combine the facial feature vectors produced by the facial recognition module and the attribute vectors produced by the attribute recognition module. The result is an efficient solution in which a lost person is more accurately identified by police officers even under unideal conditions.