Research On The Face Super-Resolution Algorithm Based On Sparse Rep-Resentation

Super-resolution image reconstruction can overcome the limitation of low resolution image sensor and effectively improve the resolution of the images and increase the detail of the image.It is one of the important research field of computer vision.Face image super-resolut io n problem is one of the core issue in the field of super-resolution and has caught the attention of the police,security,etc..Face image super-resolution problem,especially the face of single image super-resolution problem has always been the hot spot of the research.In recent years,the single frame image super-resolution algorithm based on sparse representation has achieved good results in the ordinary image super-resolution reconstruction.But relative to the general image,face image structure is more complex,and with a specific prior information.It is necessary to study on a single frame of face image sparse representation super-resolution algorit hm to improve the efficiency of face super-resolution algorithms.According to the single frame image super-resolution algorithm based on sparse representation and the characteristics of face image,this paper proposes two novel single face image super-resolution algorithm in the framework of sparse representation.The main research results are as follows:(1)Due to face image structure is more complex,and with a specific prior informat io n,this paper proposes a face image super-resolution algorithm via multi-dictionary learning and sparse representation.First of al,classifying the face image patches.Different kinds of face image blocks represent different areas of the face image and have different distribution.Second,for every category face image patches,our method learn the sparse representation dictionar y,respectively.And different categories image patches have different dictionaries and mapping matrix.It can better reflect the relationship between the low resolution image blocks and the corresponding high resolution image blocks.At the same time,through the study of Bayesian nonparametric method to study the mapping relationship between image space,our approach can make the sparse representation of low resolution image blocks and high resolution image blocks have the same sparse structure increasing the consistency of sparse representation by the Bayesian nonparametric study.Experimental results show that the algorithm has the higher signal-to-noise ratio,structural similarity and improve the overall clarity of face image.At the same time,our method has less image reconstruction time and redundancy compared to other approaches.(2)Face images contain rich edge information.In the super-resolution reconstruction process,how to keep the edge details is need to attach importance to the problem.In the processes of learning dictionary and super-resolution face image reconstruction,if do not make up-sampling to the low resolution images,the further processing can not to do.Many super-resolut io n algorithms based on sparse representation,in the up-sampling process,adopt the Bilinear interpolation or Bicubic interpolation.These methods are easy to cause image edge blur.In this paper,the edge-guided nonlinear interpolation technique is introduced into the up-sampling process of low resolution image,which improves the ability that image keep edge details,at the same time combined with multi-dictionary learning improve the efficiency and precision of the super-resolution.Experimental results show that the algorithm has the higher signal-to-no ise ratio and less error compared to other approaches.And our method increase the face image edge details.