Super Resolution and Its Application Based on Sparse Representation
|School||Harbin Institute of Technology|
|Course||Control Science and Engineering|
|Keywords||sparse representation super resolution NSCT medical image|
Image super resolution based on sparse representation is a challenging problem insignal processing, which offers the possibility to recover the lost high frequencyinformation. The development and employment of this technology will help relieve thepressure of expensive costs for imaging system, and provide conveniences for peoplewhen they demand high resolution image.This paper presents an approach of image super resolution which appliesNonsubsampled Contourlet Transform (NSCT), which gets good results.Under the sparse representation framework, the presented algorithm appliesNSCT to the feature extraction which is needed for in training high and low resolutiondictionary. The NSCT is characterized by multiscale, multidirection and translationalconstancy, and the derivative of image is also used to extract the features. Thecharacteristics of these methods are benefit to extract useful high frequency informationfrom low resolution images. Consequently these methods provide strong guarantee forsuper resolution reconstruction.In the simulation，a series of images with rich details are selected as training setto obtain joined dictionary for high and low resolution. The dictionary is tested forimages of various structures and textures in test sets for different feature extractedmethods and sparse factors. The super resolution images are valued by the peak signalto noise ratio and structural similarity and the indexes are superior to the index oftraditional method. Except for the objective assessment, we also give the natural superresolution images generated by the proposed algorithm which show good visual effects.In addition, the proposed method is also applied to the detection of microcalcification ofmammography. The mammography should be enhanced by the contrast enhancementmethod of this paper first and then be enlarged by the proposed algorithm. Experimentsshow that the focus of patients are enlarged and separated from healthy body tissueswhich can be seen clearly by eyes. The algorithm has also been used to improve theresolution of magnetic resonance images with low resolution. Experiments results showthat the proposed method has a good performance on improving the resolution ofnatural and medical images and potential application value in computer-aided detection.