Sketch Face Recognition Based On Deep Metric Learning

Since the birth of face recognition technology,it has penetrated into all aspects of our life.Face recognition technology has become an indispensable part of life.With the development of the Internet technology,social security has presented further challenges to face recognition technology.Sketch face recognition has a very important role in social security.Through sketch and face recognition technologies,police can find suspects quickly and efficiently.Sketch face recognition is to match the sketch face image hand-painted by the artist with the photo face image to find the identity information of the specific person.The main challenge of the sketch face recognition method is to reduce the difference between the sketch images and photo images,and to improve the recognition accuracy.This paper proposes three approaches for sketch face recognition based on deep metric learning technology,and three approaches can reduce the difference between different modalities,and the category discrimination analysis can be performed.Firstly,in order to reduce the differences between different modality photos in sketch face recognition,a cross-modality multi-task deep metric learning(CMTDML)for sketch face recognition method is proposed.CMTDML designs a two-channel neural network to extract non-linear features of photo modality and sketch modality,and the parameter sharing characterisitcs can reduce the differences of features between different modalities.Then,CMTDML develops the loss function to constrain the features in common space,where intra-class compactness and inter-class separability of features are promoted.Secondly,in order to further analyze the categories of the sketch and photo samples,a multi-category multi-task deep metric learning(MCDML)for sketch face recognition method is proposed.First,the triplet-margin-center loss is designed to enlarge the distance of inter-class samples and reduce intra-class samples variations simultaneously.Then,MCDML proposes the triplet-margin-center loss,which can effectively improve intra-class compactness.Moreover,the triplet-margin-center loss applies a hard triplet sample selection strategy.It aims to effectively select effective hard samples to avoid unstable training phase and slow converge.In this method,the samples from photos and from sketches taken from the same identity are closer,and samples from photos and sketches come from different identity are further.Finally,in order to simultaneously reduce the different modality differences between sketches and then perform category discrimination analysis of sketch and photo samples,multi-category cross-modality deep metric learning(MC~2MDML)for sketch face recognition method is proposed.First,MC~2MDML designs a quadruplet-channel neural network to extract features from sketch images and photo images.At the same time,this approach adds a residual module into the quadruplet neural network and designs a modality loss to reduce the differences between the photo modality and the sketch modality.Then,this approach proposes a quadruplet modality difference loss function based on the quadruplet-channel neural network,so that the samples projected into the common subspace are compact within classes and separable between classes.Compared with the traditional sketch face recognition method,the proposed approaches can obtain better recognition results on the CUFSF database and the IIIT-D database.