Research Of Face Recognition Based On Linear Encoding

In recent years,as a kind of important technology of biometric recognition,face recognition has attracted a lot of investigations in the field of pattern recognition,machine learning and computer vision.Among them,a method of classification based on linear representation model has been widely used and become one of the most popular research directions in the field.This method utilizes the linear representation to obtain codings of training samples corresponding to test samples,and then employs the codings for classification.However,due to the consistency of codings and the generalization of the model,there is still a lack of study on the sparsity and independence of the codings.In addition,current methods often have difficulty in taking advantage of the information of space structure of images,since the images are often processed into vectors by linearly transformation.To tackle the above problems,this thesis investigatives linear encoding based methods for face recognition.The innovations are listed as follows:1.This thesis is devoted to extracting high-order statistical information of the data from the aspect of feature extraction and proposes a method of representation of sparse factor coding(SFCR).The sparseness of linear sparse coding and the independence of codings obtained by independent component analysis are combined based on the principle of encoding consistency.And the influence of different distance metrics,methods of sparse optimization and classification on the performance of linear coding representation are analyzed and compared.This thesis presents a new objective function to optimize the sparse coding based on independent component analysis(ICA),solving the issue that codings of test samples are not as sparse as the training.Compared with coefficients of sparse coding by dictionary learning,the optimization of L1 norm sparsity constraintion is introduced in the SFCR,and the effect of coding sparsity on the recognition result is verified.2.This thesis proposes a method of similarity measure which is broadly consistent with the human visual experience.The method does not model noises,but can adaptively measure the similarity of samples based on the coding representation of training samples about a test sample.3.Combining the above method of similarity measure and nearest neighbor classification(NN),this thesis proposes a method for occluding face image recognition,face recognition based on human visual recognition.The method utilizes the unmasked part to identify the neighborhood of the pixels,which combines the gradient information and the neighborhood information of the image pixel to calculate neighborhood information on the residual image of samples by setting the pixel threshold.With the introduction of the pixel space information,linear transformation is employed,and training samples of each class are coded linearly,that is,the degree of aggregation.Then,the SFCR perform the classification based on the size of the degree.These three methods are all based on the idea of linear coding,which makes face images as much as possible to meet the requirements of subspace assumption and coding consistency,and finally achieved competitive experimental results.In addition,this thesis integrates above methods and exploits the extracted high-order information features of image data and the adaptive similarity metric to introduce the spatial neighborhood information of pixels for recognitizing the occlusion faces,which is also a bright spot.