Heterogeneous Face Recognition Theories And Methods

Biometric based identity recognition plays an important role on security domain and all kinds of authentication systems.Owing to the non-contact,non-compulsive qualities and accurate,convenient,visualized characteristics,face recognition has great development and application prospects.With specific application scenarios appearing,the acquired face images may come from different modalities besides normal visible light imaging camera,such as sketches,near-infrared images,3D face data,thermal infrared images and low resolution images.However,the registry systems often collect visible light images from users.How to match the images from visible light and other modalities becomes the main objective of our heterogeneous face recognition.There are two bottlenecks in heterogeneous face recognition algorithms.One is the different data distributions in distinct modalities causing large gaps in their feature spaces,making the traditional face recognition algorithms invalid,and the other is the existing heterogeneous face image databases being small-scale,making the well-performed machine learning algorithm over fitting easily.Taking NIR-VIS heterogeneous face recognition as a main direction of research,we propose the following methods: To address the modality differences and over fitting problem in heterogeneous face recognition,we propose a deep representation transfer learning method.Firstly,we employ deep transfer convolutional neural networks with ordinal measures.The ordinal activation function boosts the robustness of CNN for heterogeneous variations.Secondly,pre-train the networks with large-scale visible light images to extract the general face features.The learned filters provide priors for NIR-VIS data.Finally,we come up with a cross-domain triplet strategy to enlarge the training data,and two kinds of triplet losses to impose restrictions on inter-class and intra-class distances,which make the learning process on deep model for small-scale data possible.To make the model have stronger discriminative ability,we select hard triplets and iterate the selection and training process.To summarize,we provide the latest developments for heterogeneous face recognition,including the related databases,research methods and results,and discuss the existing problems and future research directions.Specifically,we propose a deep transfer CNN model which can employ large-scale visible light images as well as small-scale heterogeneous data simultaneously.The proposed method achieves the state-of-art performance on the challenging NIR-VIS database.The presented approach can also be applied to other heterogeneous recognition problems.