视网膜眼底图像的快速鲁棒视盘定位方法研究

发布时间：2018-06-24 10:50

本文选题：视盘检测 + 血管　；参考：《湘潭大学》2016年硕士论文

【摘要】：视网膜眼底图像近年来广泛用于青光眼、白内障等眼底病变以及会对视觉造成严重影响的糖尿病、高血压、冠心病等疾病的诊断。由于病人规模庞大,眼科专家稀缺,因此基于计算机视觉的眼底图像自动病变检测系统得到广大研究人员的研究与关注。视盘是视网膜眼底图像的主要生理结构之一,准确的定位视盘对于眼底图像视觉检测系统具有以下重要作用:(1)由于视盘和黄斑之间的距离可以近似估算得到,所以定位视盘位置有助于定位黄斑;(2)视盘是视网膜血管的汇聚区,主血管从视盘向整个视网膜眼底图像其他区域延伸,所以确定视盘位置有助于分割血管;(3)视盘的准确定位能够将其和容易混扰的亮黄色病变(溢出物病变)区分开,从而有助于提高视网膜眼底图像病变检测的准确性和鲁棒性。由于成像条件的变化造成的成像质量问题,以及各类病变等因素对视盘外观的干扰或者破坏,使得视盘定位并不是一个容易解决的问题,目前还没有能在大规模测试中普遍适用的算法,现有算法在鲁棒性或快速性方面仍存在一定的缺陷。针对上述问题,本文致力于研究新的算法,提高视盘定位算法的鲁棒性和效率,论文主要工作包括以下几个方面:为了检验利用血管特性和外观特性的算法在检测不同成像质量和病变类型情况下的出错情况,选取了基于线性算子的视盘定位方法、基于血管分布和走向特性的视盘定位方法和基于投影法的视盘定位方法三种典型方法进行实验分析,在我们整理归类的一个较大规模的病变图像集进行了测试与算法鲁棒性分析。实验结果表明,利用视盘外观特性或只利用血管特性的方法对视盘外观质量和血管质量比较敏感,当质量变差时,准确率会有较大程度下降;此外,随着病变类型数目的增加,三种方法定位的准确率都会下降。提出一种快速的视盘定位算法。算法首先通过LoG斑点检测确定一定的视盘候选区域,然后综合利用候选区域的外观和血管特性辨别视盘真实位置。该方法由于可以在缩小后的眼底图像上进行,因此能够获得实时检测效果,同时由于具有多尺度检测能力,此方法能适应不同尺寸的图像数据集。该方法在五个公开的数据集1540幅图像进行了测试,定位准确率达到97.9%。在五个数据集中平均检测时间约0.2s,远快于目前绝大部分检测算法。提出一种基于多种特性和两级决策的鲁棒视盘定位方法。算法首先利用全局的血管分布和方向特性,以及局部外观特性找到若干视盘候选点,然后引入局部特征描述子HOG特征来描述视盘候选点的细节信息,进而通过SVM模型区分视盘和非视盘区域,最后采用基于相关度度量的非最大排除策略确定最终的视盘位置。实验结果表明,该算法具有良好的鲁棒性,在正常和病变的眼底图像中表现出良好的检测精度。在四个公开的图像集中,检测精度达到97.9%。
[Abstract]:Retina fundus images have been widely used in recent years for the diagnosis of diabetes, hypertension, and coronary heart disease, such as glaucoma, cataract, and other diseases such as diabetes, hypertension and coronary heart disease. Because of the large scale of the patients and the scarcity of ophthalmologists, the automatic detection system based on computer vision has been obtained by the researchers. Research and attention. Optic disc is one of the main physiological structures of retina fundus image. Accurate locating disc has the following important functions for the visual inspection system of the fundus image: (1) the location of the optic disc position helps to locate the macula, and (2) the optic disc is the retinal blood vessel. The main vessel extends from the optic disc to the other regions of the retinal image of the whole retina, so the location of the optic disc helps to divide the blood vessels; (3) the accurate positioning of the disc can separate the bright yellow lesions (spillover lesions) which are easily disturbed, thus helping to improve the accuracy and robustness of the detection of the lesion of the retina of the retina. Because of the imaging quality problems caused by the change of imaging conditions, and the interference or destruction of various factors on the appearance of the optic disc, the disc positioning is not an easy problem. There is no common algorithm which can be widely used in large-scale test at present. The existing algorithm still has a certain lack of robustness or rapidity. Aiming at the above problems, this paper aims to study the new algorithm and improve the robustness and efficiency of the video disc localization algorithm. The main work of this paper is as follows: in order to test the error of detecting the quality of imaging and the type of the disease, the algorithm based on the linear operator is selected to test the error in detecting the different imaging quality and the disease type condition using the blood vessel characteristics and appearance characteristics. The video disc location method, based on the three typical methods of visual disk location based on blood vessel distribution and direction characteristics and the projection method based disk positioning method, is tested and analyzed. We have done a test and algorithm robustness analysis on a large scale image set which is classified and classified. The method of vascular properties is sensitive to the visual quality of the visual disc and the quality of the blood vessel. When the quality is poor, the accuracy rate will be greatly reduced. In addition, with the increase of the type of disease, the accuracy of the three methods will be reduced. A fast video disc localization algorithm is proposed. First, the algorithm is determined by LoG speckle detection. This method can adapt to the image data set of different sizes. This method can be adapted to the image data sets of different sizes. Five open data sets are tested with 1540 images. The location accuracy reaches 97.9%. in five data sets with an average detection time of about 0.2S, far faster than most of the current detection algorithms. A robust video location method based on multiple characteristics and two level decision making is proposed. The algorithm first uses the global blood vessel distribution and direction characteristics. A number of visual disk candidates are found and the local feature is found. Then the local feature descriptor HOG features are introduced to describe the details of the video disk candidate points. Then the SVM model is used to distinguish the optic disk and the non optic disk region. Finally, the final video disc position is determined by the non maximum elimination strategy based on the correlation measure. The experimental results show that the algorithm is designed. It has good robustness and shows good detection accuracy in normal and pathological fundus images. In four public image sets, the detection accuracy is 97.9%.
【学位授予单位】：湘潭大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：R770.4;TP391.41

【相似文献】