Fundus Image Quality Assessment

In clinical, a clear fundus image is a prerequisite for the right diagnosis of the disease.Fundus image quality assessment is a basic point in the development of fundus image aidedthe clinical diagnosis. Usually the image quality assessment is divided into two majorcategories: subjective quality assessment and objective quality assessment. Objective qualityassessment method is mainly divided into full-reference method, reduced reference methodand no reference method. The full reference and the reduced reference need to obtain theinformation of the reference image, which is often very difficult to practice. So, theno-reference quality evaluation method is chosen in this paper.The reasons of the fundus image degradation may be two: noise and blurred. Therefore,for quality assessment, the study is on the method of the noise estimation and the evaluationof the fundus image clarity in this thesis. The key problem of noise estimate is how toselection the block size in block algorithm, which is no general accepted theory to be applied.To estimate, the effectiveness of LMLSD algorithm based on the Gaussian waveformextracted and the effect of various block sizes to the algorithm performance are paid attention.The results show that a higher accuracy is get by the5×5sub-block size in the fundus image.This reduces the estimate’s blindness because of the fixed blocking size for the noiseestimation of the fundus image. To verify the above results further, applied to fundus imageswith different white Gaussian noise, the original LMLSD algorithm and the LMLSDalgorithm based on Gaussian waveform extracted are taken different sub-blocks to estimatenoise. In addition to the above, the other two noise estimation algorithms, which are the HHsub-band mean value and middle value of wavelet coefficients, had been researched. It isshown that the performance of the LMLSD method based on Gaussian waveform extracted ismore significant. In the fundus image clarity evaluation, after an overview of image clarityevaluation function, it is put forward the fundus image clarity evaluation function based onthe edge active measure. Various edge detection operators are adopted to select one for betterperformance of evaluation of edge active. On the end, the Roberts operator meets thedemands. When a image is called “clear”, it means that the gray value changes significantly in the image, at the same time, in the frequency domain, the high-frequency components is rich.So, it is exploited that the wavelet coefficient and the edge active measure is calculated inorder to build the image feature vector, and to design a linear classifier of image clarity. Thefisher criterion is used to determine the optimal projection direction and boundaries of theclassifier threshold, which is for the judgment of the clarity of fundus image. Eventually, therate is~92﹪, and the results is consistent with the subjective judgment. The method mayhave a certain reference value for the image with a similar feature.