Binary Image Super-Resolution Based On Sparse Representation

With the development of image processing technology,the demand for high-resolution images is increasing.Binary image,that is,the image of the gray scale in two peak areas after shooting,is an important portion of images.In practice,there are a wide range of applications,such as typical text images,bar code image,license plates,which exist in people's lives in all aspects.However,low resolution will result in identification difficulties.Therefore,super-resolution reconstruction for binary images has a strong practical value.In this paper,binary image super-resolution based on sparse representation is studied deeply,and a super-resolution reconstruction algorithm for binary image is proposed.This paper contains the following aspects:1.This paper analyzes the super-resolution reconstruction model,and studies the basic ideas,classical algorithms of three different super-resolution methods,which are based on interpolation,reconstruction and learning respectively.Then the advantages and disadvantages of these approaches are compared.From the comparison,the super-resolution based on learning shows its superiority.Besides,sparse representation method is a prominent algorithm in the learning-based approaches.Therefore,sparse representation based method is adopted to conduct super-resolution reconstruct for binary images in this paper.2.This paper makes intensive studies of the basic theory of sparse representation model and several common used algorithms for solving the sparse coding problems.On the basis of fully understanding of the theory of super-resolution reconstruction and sparse representation,the framework of the super-resolution reconstruction for binary image based on sparse representation is proposed.3.The typical characteristics of binary images are analyzed deeply.It shows that there are obvious edge and texture features in binary images.Therefore,the most appropriate expression for the high frequency information of the binary images is sought from these two characteristics.By studying and comparing the extraction algorithms for edge and texture features,this paper chooses Kirsch and LBP operators as main feature extraction methods to extract the edge and texture features.Besides,another two second-order derivatives(horizontal and vertical directions)are also used to extract the features.These four operators act on the training samples for dictionary learning and the input low-resolution binary images.Extensive experiments demonstrate the positive effect of the proposed feature extraction approach for binary image super-resolution.4.There are different types of images in binary images,such as two-dimensional bar code and text images.Therefore,it is not accurate enough to adopt the same dictionary for these different images.This paper proposes a dictionary learning method for binary images.Firstly,content dependent samples are chosen randomly for the input of dictionary learning.Secondly,K-means algorithm is used to cluster the training samples.Thirdly,K-SVD method is used to train the dictionaries of the K sample sets,then the obtained K dictionaries are adopted to conduct super-resolution for binary images.The validity of the proposed dictionary learning approach for binary images is proved by experiments on the reconstruction quality.