A comparative study on application of multi-task cascaded convolutional network for robust face recognition

Abstract

This paper proposes and compares performance of various methods for face recognition. The major steps in this proposal are detecting the faces, aligning the face from cropped images, representing the face features by FaceNet model, classifying the features with Support Vector Machine. Four different face detection methods which are Haar-cascades, Histogram of Oriented Gradients and Support Vector, Convolutional Neural Networks, Multi-task Cascaded Convolutional Networks have been tested with the UTK face dataset for examining negative impacts of intensity of light and head pose problems. After face aligning by the Facial Landmark method, the FaceNet based on Convolutional Neural Network was used to extract representative features. The Support Vector Machine is then applied to train a recognizer on LFW dataset. The obtained classification results show that the system using Multi-task Cascaded Convolutional Networks leads to highest F1-score of 96%. Further analysis among examined algorithms have been presented in this study.