Near-Infrared And Visible-Light Face Image Based Bimodal Recognition

Owing to high recognition accuracy,non-contact identification and nonmandatory.face recognition has great development and promising applications.With the development of market demand for algorithms,the unconstrained factors such as illumination,posture,expression and occlusion play an important role in the factors affecting the performance of face recognition systems.In order to weaken the influence of unconstrained factors and improve the robustness of face recognition systems in rear-world scenes,the fusion recognition task of bimodal or multi-modal face images has become one of the urgent problem to be solved.Although there is a lot of work in the field of fusion identification,the performance under unconstrained conditions has not reached the ideal result.In response to this problem,two novel algorithms are proposed by this paper.Firstly,starting from the idea of designing different feature extractors for complementary identification information.And the complementary features are merged by decision level and scoring level fusion techniques respectively.Secondly,from the perspective of classification,a classification algorithm based on sparse representation is introduced,and a fusion model with discriminative and low computational complexity is proposed.On the one hand,labeled face images on other modal are too small compared to visible face images.And intuitively,how to use existing large-scale visible data to construct data sets of other modal is a heterogeneous image or a bimodal image recognition problem.On the other hand,as the application scene updating continuously,the image to be recognized will not be limited in visible light images.Near-infrared face images,thermal infrared face images,sketch face images,and other bimodal or multimodal recognition problems require solutions.In view of the above problems,this paper uses the unsupervised learning method to reduce the modal difference between bimodal images from the perspective of adaptive domain discrimination.Firstly,this paper uses cross-entropy and central loss function to jointly pre-train a full-convolution network,which gives the network a strong discriminating ability and provides the prior knowledge to another network;Secondly,another structurally consistent full convolutional network is trained by the adversarial loss,so that the data distribution of features extracted by the two network is consistent which narrowing the gap between the modalities;And finally,using the prior knowledge provided by the previous network to output the posterior probability of another modal images.In general,we studied the latest developments in bimodal recognition and the existing problems from the perspective of fusion and cross-domain.At the same time,the accuracy of fusion recognition is improved from the perspective of traditional algorithms.From the perspective of deep learning,a new bimodal heterogeneity recognition algorithm based on adaptive adversarial discriminant domain adaptation is proposed.