Cross-Modal Feature Learning For Heterogeneous Face Recognition

Face recognition is one of the most popular biometric identification technology.Owing to its natural,non-compulsive,non-contact and parallelism qualities and accurate,convenient,visualized characteristics,face recognition has attracted more and more attention in the field of computer vision and pattern recognition.In recent years,fast-growing face recognition provides robust technical support for the applications in many fields with its increasing recognition accuracy and recognition speed,especially in the field of public safety.However,with the diversification of image capturing equipment and the demand of certain specific application scenarios,more and more face images of non-visible light have gradually emerged.For example,portrait or facial sketch,near infrared image,thermal infrared image,three-dimensional face model,low resolution image,etc.We call these facial images heterogeneous faces,and due to their different imaging mechanisms,there is a huge appearance gap between the heterogeneous face images of same identity,which causes the performance of the traditional face recognition algorithms is not satisfactory when training on the heterogeneous face dataset.Therefore,different from the traditional face recognition algorithms,heterogeneous face recognition focus on the research of reducing the modal gap of heterogeneous face images to solve the recognition problem between the facial images of different modalities,it is also known as cross-modal face recognition.And in this paper,we proposed two solutions for the heterogeneous face recognition:(1)Coupled Discriminant Feature Learning(CDFL).CDFL eliminates the modal gap of heterogeneous face images by image filtering,CDFL constructs a filter-learning objective function by combining the idea of linear discriminant analysis with Pearson correlation coefficient.And the most discriminative features can be learned from the filtered heterogeneous face images.(2)Hierarchical Discriminant Feature Learning(HDFL).The learning of discriminative filter bank of CDFL is based on image blocks,which makes the structure of the learned filter bank is highly complex,and there are multiple filters in each image block of each sample from each modality.The complex structure of filter bank makes the recognition process of heterogeneous face recognition very cumbersome.Therefore,to simply the structure of the learned filter bank,this paper proposed a hierarchical AdaBoost network(HBN).HBN can fuse the filter bank into one optimal filter in a weighted voting manner by learning the weights of each filter.Extensive experiments on three different heterogeneous face databases fully demonstrate the effectiveness of HDFL,and its superiority also can be proved by comparing to the state-of-the-art heterogeneous face recognition algorithms.