A Sparse Representation Method Based On Virtual Samples For Face Recognition

In the recent years,face recognition has been an important technology of modern biological information recognition with the rapid development of computer vision and pattern recognition technology.So far,the face recognition has made great progress in theoretical research and practical applied.However,it still can be greatly interfered by a lot of factors,such as lighting conditions,posture,obstruction,lack of samples and so on.Therefore,how to improve the accuracy of face recognition become the research priorities of scientists,and also the main direction of this paper.This paper analyses the mainly rational and advantage and disadvantage of sparse representation and virtual samples.On the basis of the feature of sparse representation and virtual samples,a novel sparse representation method based virtual test samples is proposed.Specifically,the main research works and contributions are summarized as follows.1)To solve the problem of insuffi cient capacity of training sample caused by variant face expression,a sparse representation method based on virtual test samples has been proposed.It exploited the symmetry of the face to generate virtual samples,and use the virtual samples to simulate the possible view of face image under the condition of variant face expression and posture change and increase the number of test samples.Then,the training samples are used to represent original test samples and the virtual test samples respectively,and use the representation to conduct weighted score level fusion.The test samples will be classified into the class which makes the largest contribution with the final results.A large number of face recognition experimental results on different face databa ses show that our method has better classification results than those traditional approaches.2)In order to further improve the recognition rate,a two-phase method and an improved two-phase method have been proposed.In the first phase,symmetry of the original samples is used to generate new virtual samples.Then training samples are used to represent original test samples and the virtual test samples respectively,and the deviation between the training samples and the each test samples is exploited to determine the appropriate training samples.In the second phase,it expresses the test samples and virtual test samples as a linear combination of the appropriate training samples respectively and received the contributions of each training samples,then it takes the advantage of the weighted sum of deviation between the training samples and the test samples to improve the accuracy of face recognition.Experimental results show that this method can realize lower face recognition error rate,but it would be improved.To further optimize the two-phase method,it exploits Manhattan distance to describe the distance between training samples and test samples and select the most suitable training samples to improve the method.Experimental results show that the improved method improves the accuracy of face recognition.