Research On Face Hallucination Based On Sparse Representation Model

In recent years,learning based face hallucination method has drawn amount of attentions.Researchers have proposed a series of face hallucination algorithms from different perspectives.Especially,the sparse representation based face hallucination has obtained the best result as reported in literature.However,the performance of the sparse representation based method reduced greatly for real surveillance video face image superresolution.The result is influenced by noise,illumination,imaging process and other factors.So how to build the robust sparse representation model,learn an expressive dictionary and implement the effective high resolution face image reconstruction is the key to solve the face image superresolution issue in surveillance video.Under the support of related foundations(No.61172173,No.2008ZDXMHBST011),this paper foucus on the research on sparse representation based face hallucination technology.Specifically,the thesis focus on three aspects:(1)Face superresolution algorithm based on K-nearest neighbor sparse coding mean constraintAs in existing sparse coding based face superresolution methods,the dictionaries usually are overcompleted,which makes the spare representation coefficients changes greatly while the observed features have small changes.The standard sparse coding based face superresolution methods is sensitive to noise.In order to solve the problem,this paper presents a face superresolution algorithm based on K-nearest neighbor sparse coding mean constraint.In this method,the coefficient of a single block is constrained by the expectation of its K-nearest neighbor block coding coefficients under the standard sparse coding framework.In the case of zero mean,the expectation of multiple neighboring block codes in high and low resolution spaces is consistent with each other,also it is robust to noise.The experiments on CAS-PEAR-R1 face database[86]validate the proposed method.Compared to the state-of-the-art method[36],the proposed method improves the PSNR and SSIM values by 0.3364 dB and 0.0243,respectively.(2)Face superresolution algorithm based on high-dimensional graph constrained sparse codingMost of the existing sparse coding based face superresolution algorithm do not take into account the geometric structure information of the face.The neighborhood structure information of the local feature is lost in the coding process,which affects the expression ability of the redundant dictionary.In order to solve this problem,this paper presents a face superresolution algorithm based on high-dimensional graph constrained sparse coding.In the stage of sub-dictionary learning,the internal structure of the image block in the high resolution space is introduced into the sparse coding process.So that the sparse coding coefficient can maintain the similarity of the original image block.Also the local information in original image block can be preserved.By using the similarity and local information,it can improve the expression ability to learning dictionary,which lead to a better superresolution reconstruction capability.The effectiveness of the proposed algorithm is verified on CAS-PEAL-R1 face database[86]and real scene face image.Compared to the state-of-the-art method[76],the proposed method improves the PSNR and SSIM values by 0.8dB and 0.0262,respectively.(3)Face superresolution algorithm based on multi-morphology sparse representationFace is one of the most familiar contents to human beings.The subtle errors of reconstructed facial images will bring significant visual bias.The existing face superresolution methods are mainly reconstructed the high resolution face image from a single form of face component(or feature),such as the original pixel,gradient,and other features by using the same regular terms,which resulting in over smooth or jagged effects and other deficiencies.In order to solve this problem,this paper proposed a face superresolution method based on multi-morphology sparse representation.The method firstly decomposes the training samples into cartoon and texture components,and then learns the cartoon and exture component dictionary,respectively.Secondly,it decomposes the input face image into cartoon and texture component,which are reconstructed separetly via superresolution method.The total variational regular term is used to reconstruct the cartoon component,and the non-local similarity regular term is used to reconstruct the texture component.Finally,by fusing the reconstructed high resolution multi-component,we can get the final high resolution face image.The algorithm can effectively maintaining the details of face image,and can improve the reconstruction quality.Experiments on CAS-PEAL-R1 database[86]verifies the effectiveness of the proposed algorithm.Compared to the state-of-the-art method[26],the proposed method improves the PSNR by 1.23dB.In summary,this study foucus on sparse representation based face image superresolution methods from the aspects of robust sparse representation,expressive dictionary learning and multi-morphology sparse representation.It proposed three methods for face image superresolution:Face superresolution algorithm based on K-nearest neighbor sparse coding mean constraint,Face superresolution algorithm based on high-dimensional graph constrained sparse coding,Face superresolution algorithm based on multi-morphology sparse representation.The innovative research results may be used for video surveillance system.At the end of this paper,the future research work is forecasted,including the selection of efficient adaptive regularization parameters,and the construction of more efficient online dictionary learning method and so on.