The Research Of RGB Image Based Face Anti-spoofing

With the popularization of face recognition systems,the challenges for the face antispoofing algorithms have become increasingly severe.And the challenges mainly comes from two aspects: 1)the face anti-spoofing algorithms should be robust in various situations,2)more and more new attacks are emerging.These issues requires that the face anti-spoofing algorithms should be well generalized to the data from the unseen scenes and the new attacks.However,the previous methods based on deep learning technology,whether they utilize the binary supervision or auxiliary information supervision,have the following two inherent shortcomings: 1)Feature entangled.The previous methods extract the features from the human face images and restricted them by various supervision,then utilized the whole features to make a prediction.while these features might contain the information that is not related to the face anti-spoofing task,which leads to the degradation of the generalization abilities.2)Strong relevance to facial structure.Unlike face recognition/detection systems,face anti-spoofing is inherently concentrated on detail exploration,which is independent of global facial structure.Former methods inevitably introduced structural information since the whole face serves as the input.This facial structure provides the network excessive context to well predict auxiliary depth,and hinders the representative capabilities of extracted features if tested on a non-structural input.Therefore,to address the above two issues,this paper proposes two algorithms from the perspective of the feature and the data,to solve those two limitations separately.This paper briefly introduces the contributions of these two algorithms:· Face Anti-Spoofing via Disentangled Representation Learning Algorithm.To sum up,the contributions of this algorithm are three-fold: 1)This paper addresses face antispoofing via disentangled representation learning,which separates latent representation into liveness features and content features.2)This paper combines low-level texture and high-level depth characteristics to regularize liveness space,which facilitates disentangled representation learning.3)Abundant experiments and visualizations are presented to reveal the properties of liveness features,which demonstrates the effectiveness of our method against the state-of-the-art competitors.· Structure Destruction and Content Combination for Face Anti-Spoofing Algorithm.The main contributions are summarized: 1)A novel Destruction and Combination Network(DCN)is proposed for face anti-spoofing.By simultaneously destroying the facial structure in the original image and excluding domain knowledge in the dataset,our framework is reliable for extracting discriminative features thoroughly.2)Among patch-based frameworks,this paper further models second-order relationship between patch pairs,promoting the capture of crucial information in each local region.3)This paper evaluates our method on extensive benchmarks.With superior performance and convincing visualization,this paper demonstrates the effectiveness of data recombination.