Face Recognition for Embedded System Based on Optimized Triplet Loss Neural Network

Algorithms based on Convolution neural networks are significantly better than artificially designed features. Among these methods, the network trained with the triplet loss function can usually obtain better features than the direct classification, and thus get better recognition accuracy. The input of the triplet loss function consists of an anchor picture, a positive picture and a negative picture. The distance between the anchor picture and the positive picture is reduced by training, thereby achieving the purpose of face recognition. However, the parameters of the network are usually too much that it cannot be directly applied to embedded devices. In this regard, this paper proposes a solution for compression optimization based on triplet loss network. This scheme can reduce the resource overhead of the network, improve the processing speed, and realize high-precision real-time face recognition on the embedded device. This solution greatly improves the performance of the network on embedded devices, and achieves the effect of high recognition accuracy and low resource overhead.