Research And Application Of Retina Image Super Resolution Algorithm Based On Dictionary Learning

Vision is an important medium for obtaining external information.With the aging of the population and the lack of awareness of eye protection,retinal lesions such as age-related macular disease,pathological myopia and retinitis pigmentosa cause vision loss.Ophthalmologists primarily use retinal images acquired by Spectral-Domain Optical Coherence Tomography(SD-OCT)to diagnose and predict disease.SD-OCT imaging technology produces 128 low-resolution(LR)images in one scan to form three-dimensional volume data.At the same time,five high-resolution(HR)images are generated in the foveal macular fossa,the organizational structure is clearly visible.Due to the small number of high-resolution images,it is impossible to perform three-dimensional quantitative analysis of diseased tissues.Low-resolution images contain a large amount of speckle noise,and low-quality retinal images present difficulties for the clinician's diagnosis and quantitative prediction of diseased tissue.With the wide application of Super Resolution(SR)technology in medical images,this paper performs super-resolution reconstruction on low-quality retinal images.The quality of the reconstructed retinal image is improved,which is conducive to the precise treatment of retinal diseases.In this paper,based on the super-resolution algorithm framework of dictionary learning,the algorithm of suppressing speckle noise and maintaining the boundary is studied,and the reconstructed image quality is improved.The main work of the paper is summarized as follows:(1)A retinal image super-resolution algorithm with non-local similarity is proposed.As the sparse coefficients of similar image patches are similar.The constraint is constructed by the similarity between the noise image patch and its neighbor image patches,and the non-local similarity constraint is used as the regular term of the objective function of the super-resolution algorithm.Experiments show that the proposed algorithm can effectively suppress the effects of speckle noise while super reconstruction.(2)A edge-preserved retinal image super-resolution algorithm is proposed.The boundary information of the retinal image is crucial for the detection and segmentation of the lesion area.Therefore,the image should maintain the boundary information as much as possible after the super-resolution reconstruction.This paper uses the boundary prior information of high-resolution images to guide the reconstruction of low-resolution image boundaries.The experimental results show that the proposed algorithm achieves the preservation of boundary information in the super-resolution process.(3)A retinal image super-resolution algorithm with edge preservation and non-local similarity constraint is proposed.In this paper,the non-local similarity constraint and the edge preservation constraint are used as the regular term of the objective function of the super-resolution algorithm.Experiments verify that by solving the objective function,it is possible to achieve edge preservation and noise suppression.(4)This paper designs and implements a super-resolution system for retinal images.The system mainly implements four super-resolution algorithms for retinal image reconstruction and image processing basic operations(browsing,type conversion,geometric transformation,noise addition,denoising,enhancement and edge detection).In addition,the system also implements a hand animation line for the retinal layer structure.