Max Feature Map CNN with Support Vector Guided Softmax for Face Recognition

Abstract

Face recognition has made significant progress because of advances in deep convolutional neural networks (CNNs) in addressing face verification in large amounts of data variation. When image data comes from different sources and devices, the identifiability of other classes and the presence of profile face data can lead to inaccurate and ambiguous classification because other classes lack discriminatory power. Furthermore, using a complex architecture with many deep convolutional layers can become very slow in the training process due to a huge amount of Random Access Memory (RAM) usage during the reverse pass of backpropagation. In this paper, we design a light CNN architecture that addresses these challenges. Specifically, we implemented Max-feature-map (MFM) into each convolutional layer to improve the accuracy and efficiency of the CNN. The strength of the support vector-guided SoftMax (SV-SoftMax) is also used in the proposed method to emphasize misclassified points and adaptively guide feature learning. Experimental results show that the 9-Layers CNN with MFM layer and SV-SoftMax outperform VGG-19 with 96.22% validation accuracy and the second rank below FaceNet tested on the same dataset with fewer parameters. Moreover, the model performed well on data that is obtained from various capture devices such as webcam, CCTVs, phone cameras, and DSLR cameras. The implications of this research could extend to scenarios requiring face recognition technology implementation with light size, such as surveillance and authentication systems