Memory Optimized Deep Learning based Face Recognization

Abstract

Machine learning powerful deep learning model is difficult to apply on a low segment embedded system or low segment mobile due to memory-constrained and battery-constrained platforms. This is a major problem, deep learning models have not been yet deployed on very low segment embedded systems such as smartphones (low price), smart wearables, traffic signals, and embedded systems. We are going to present a new way of doing Face Recognition with CNN to solve this problem in this paper. With the help of this, we can easily deploy a trained CNN based model on memory constrained platform such as low segments embedded system or low memory mobile devices/smart wearables, etc. We present a novel and "very efficient" Network Architecture in this paper that includes MobileNetV2 with center loss and training tricks for "deep face recognition" for an embedded system that is memory constrained. In our tests, we ran our proposed Network Architecture on the network at some very memory constrained systems (as low as 2 MB). The result shows that when it is runs on a very low memory segment platform as compared to standard CNN based models, It produces very good results on various face verification datasets and model size less than 1MB for a memory-limited embedded device