Face Feature Representation And Face Retrieval Based On Deep Learning

With the development of information technology,face recognition technology has been gradually applied to the security,surveillance,banking and other industries.Face recognition is playing an increasingly important role even in our daily lives.Face recognition technology usually consists of the following three components:face detection,face landmark detection and face feature extraction.Face detection and face landmark detection have been studied well,thus,the most important task is to extract more discriminative face feature.Traditionally,image scientists will design face feature descriptors based on their comprehension in image and digital image processing techniques.This approach will bring a lot of problems,first of all,it is hard to design an effective handcraft feature descriptor without understanding digital image processing deeply.And the performance of designed feature descriptors cannot be determined directly.Therefore,traditional face feature extraction is gradually replaced by deep learning techniques.Face recognition based on deep learning has many advantages:first,the face feature can be directly learned through the network model by end-to-end training;second,there are a large number of face images in the Internet,which can be easily collected;finally,the deep learning mathematical model itself has a strong expression ability and can extract discriminative face features.In this paper,convolutional neural network structure which is commonly used in image recognition will be introduced firstly.After analyzing and comparing the characteristics of different network structures,we will introduce an effective network structure which has a good performance in terms of calculation and parameter number.Subsequently,we will introduce some commonly used metric learning techniques in face recognition,and propose a supervised function of metric learning to optimize the feature angle and obtained the state-of-the-art results on the open dataset.Finally,we will introduce some feature hashing techniques which can learn the face binary features and greatly reduce the storage and expedites the retrieval efficiency.