Scrambled Domain Face Recognition Via Statistical Manifold And Information Bottleneck Method

With the development of the Internet of Things and computer vision technology,video surveillance systems have a wide range of applications in Public Safety.In many scenarios,scrambling algorithms can hide faces which can effectively prevent visual privacy leaking.At the same time,biometric verification of face needs to be performed in the scrambled domain.Scrambling images are characterized by chaos,diversity,and redundancy,which will make the scrambling face useless.In summary,face recognition in scrambled domain is full of challenges.This paper abstracts the scramble face recognition problem by statistical manifolds,and proposes a Cascading Information Bottleneck framework as computation model.Finally,the variational approximation method is used to solve the computation model,thus we access the Variational Cascading Information Bottleneck Network(VCIBNet).The main work includes:1.Using the statistical manifold hypothesis,we construct the cascading process of face recognition in scrambled domain.The cascading process is composed of parameter reduction and pattern adjustment.In parameter reduction,we start from the scrambling face manifold,using chaotic homomorphism and information monotonicity,design coarse-grained statistics to construct the parameter reduction manifold.In pattern adjustment,we start from the parameter reduction manifold,using supervision information and availability constraints to adjust manifold,while the feature pattern manifold is used as ideal target.The overall objective function of the cascade process is built on the basis of the divergence between probability distributions.2.Due to the asymmetry of divergence,it is difficult to solve the objective function uniformly,so divergence is equivalently replaced by mutual information.The information flow graph is used to construct the Cascade Information Bottleneck(CIB)model.The CIB model was split using the divide and conquer method.On the one hand,from the perspective of information theory,the compressed information flow graph is designed to access the Parameter Reduction Information Bottleneck(PRIB).On the other hand,the availability constraint is converted under rate distortion,and the Pattern Adjustment Information Bottleneck(PAIB)is constructed.The CIB objective function is constructed by mutual information.3.Applying variational inference to approximate CIB as a deep learning model.The PRIB is approximated to a -Variational Auto-Encoder(-VAE),and then PAIB is approximated to a Multi-Layer Perceptron(MLP)network.Under a certain variational error,the two can be collectively referred as the VCIBNet.This paper discusses the super-parameters of VCIBNet on the MNIST handwritten dataset,and analyzes the robustness of VCIBNet for multiple scrambling,and test recognition validity on the ORL,CMU-PIE and PubFig face datasets.Experiments have shown that VCIBNet can quickly achieve the recognition of scrambled faces and has a significant performance improvement over existing methods.