Research On Face Liveness Detection Based On Illumination Consistency And Context Awareness

With the development of face recognition technology,face-based identity authen-tication system has been widely applied in various fields.Although the current face recognition technology can cope with the detection under different circumstances,it is still difficult to distinguish whether the face in front of the camera is a real person,a photo or a video.Therefore,the face liveness detection technology with both prac-ticality and reliability is the basis of face authentication system,which has important research value.Although the existing algorithm has achieved good detection results,it still faces many problems,such as the existing method is susceptible to the influence of light;The traditional algorithm takes the face region as input and loses the image context information.In view of the above problems,this thesis conducts an in-depth study on face liveness detection algorithm by combining illumination consistency and context awareness.The specific work and innovations are as follows:In order to improve the illumination robustness of the algorithm,this thesis pro-poses a illumination robust two-channel fusion face liveness detection algorithm.This algorithm proposes a dual channel convolutional neural network for feature extraction.In order to take the best advantage of images after illumination consistency,which have the ability to cope with different light conditions,and rich texture information of RGB images,this thesis proposes a attention-based generic fusion method to fuse these two features weighted by location and get the fusion feature seeing both light robustness and rich texture features.Experiments show that compared with existing algorithms,the method proposed in this thesis reduces the illumination sensitivity of the detection,and improves the performances of intra datasets and inter datasets.In view of the problem that existing algorithms need to take the face region as input and lose the image context information for detection,this thesis proposes a cas-cade face liveness detection algorithm based on multi-layer pooling feature fusion and iterative illumination estimation.The first-level detector expands the detection of face liveness detection into a three-way classification problem,which is used to detect real face,attack face and background.Multi-layer pooling feature fusion is used to en-rich the context information of features and improve the performance of the detector.The second detector uses iterative illumination estimation based Retinex algorithm to achieve illumination uniformity.The detector further cascade the enhanced brightness channel with multi-color channels and extract LBP features to obtain the secondary de-tector with illumination insensitivity,which is used to deal with the difficult cases of the upper detector.Compared with other algorithms,this method gives consideration to end-to-end detection characteristics and illumination robustness,and has better de-tection performances.This thesis further explores the task correlation between face detection and face liveness detection and proposes a multi-task liveness face detection model based on context awareness.In order to improve the detection performance of small-size faces,enrich the context information of features and expand the receptive field of features,shortcut feature pyramid fusion and context aggregation are carried out on the detec-tion layers of the model,so that the deep features can guide the shallow features for detection.In addition,the context supervision information of head and nose is added by semi-supervision.The depth characteristic of the nose region is used for face live-ness detection.The feature size of the small faces is increased by using the large re-ceptive field of the head region to improve the corresponding detection performances.Compared with the existing methods,the mAP of this algorithm reached 89.1%in the difficult subset of face detection task,and the EER and HTER of face liveness detection task reached the best level in the existing literature.