A Lightweight and Robust Face Recognition Network on Noisy Condition

Abstract

Recently, deep learning has a significant breakthrough in face recognition research. Using the state-of-art convolutional neural network (CNN) model is continually improving the accuracy of recognition. However, it is difficult that the large CNN models deploy on mobile phones or embedded devices with limited computation resources and memory. At the same time, these face recognition networks show low performance in the complex environment, such as noise, shadow, illumination and so on. To address these problems, we propose a lightweight and robust face recognition network (LD-MobileFaceNet) to improve the traditional MobileFaceNet in noisy environment. In this paper, an efficient and flexible denoising block is proposed, which is an independent module to apply in MobileFaceNet. The proposed denoising block uses non-local means algorithm to denoise features that are extracted by convolutional layers. With the residual connection and the 1 × 1 convolution, it can remain more information and be combined with any layers in MobileFaceNet. Furthermore, we set fewer bottleneck layers, replace PReLU with swish nonlinearity to compensate for the loss accuracy. The experimental results demonstrate that LD-MobileFaceNet with swish is 21.35% more accurate on noisy LFW dataset while reducing parameters by 25 % compared to MobileFaceNet.