Research And Application Of Modular Sparse Representation For Face Recognition In Complex Situations

Since face recognition is widely used in identification,access control and security supervision,it is a very popular topic in computer vision and machine learning in recent years.At the same time,as the network environment is becoming more and more complex,thus the demand for information security is also incresingly strict.the traditional recognition algorithm is receiving the challenge from the real environment.So designing a face recognition algorithm which is fast,efficient,anti-interference with high recognition rate is our tireless pursuit.At present,the application of face recognition has been extensively appeared in our vision and life.For instance the current hot face recognition payment,and even the station's face security check,it shows the success of face recognition in practical application.Meanwhile due to the extensive application of face recognition,the challenges it faces are also omni-directional and diverse.To settle this problem,In the past few decades,faced with different challenges,numerous scholars have made outstanding contributions to the field of face recognition.Their contributions are mainly focused on how to improve the recognition rate and the cost of performance reduction in complex environments(such as illumination,pose,expression,occlusion and other interference information).Facing the new application and demand,it is necessary for scholars to gain a higher level of challenge in the field of theory and practice.1)In this paper,a modular joint feature algorithm is proposed based on sparse representation.The algorithm combines the advantages of Gabor operator and LGBP operator and dynamically assigns the weights to achieve the optimal classification stratege.In view of the shortage of training samples and the interference information such as illumination,expression change and even occlussion in the training set or test set.To mainly settle the occlussion problem,the training sample and the test sample are both modularized to give different weights for different modules.Specifically,each training sample is divided into several sub-modules of m * m,and then different modules are assigned different weights based on residual value and Fisher rate.In this way,it is possible to give lower weight for higher occlusion samples or give less occlusion samples with a higher or maximum weight,to reduce or shield the interference from the occlusion(wearing glasses,scarves,and masks).2)In reality,there may exist more serious interference than occlusion,such as lateral face.It means we cannot get the regularized face image,which greatly increases the difficulty of correct recognition,and reduces the recognition effect.For this reason,firstly,the FIP(face Identity-preserving)feature is extracted from the neural network,which is insensitive to the position angle change and the illumination change.Then we use the FIP feature to recover a frontal face image data through a fully connected layer,at this time,the recovered frontal face image has reduced the illumination change and the face tilt to the maximum extent.Finally the corrected face image and the original face image are used as training samples for classification and recognition respectively.The result of the experiment shows that the reconstructed face sample image is not only less interference in the light and pose,but also preserves the important features for classification,thus can effectively improve the recognition rate.