基于血管分割的视网膜图像无参考质量评价方法

No reference quality assessment method for retinal images based on blood vessel segmentation

为了研究不同失真类型和不同失真程度对血管分割的影响,本文将图像的失真类型和失真程度量化为图像血管分割精确度,由于现有公开库中包含血管分割标签的图像中均为低失真甚至无失真图像,因此本文构建了一个视网膜失真图像数据库,共包含2种失真类型,每种失真类型的图像均有8个等级的失真程度,共552幅视网膜失真图像,并将每幅失真图像对应的血管分割精确度作为该图像的标签。此外,本文提出了一种基于血管分割方法的视网膜图像无参考质量评价方法,通过提取视网膜图像的像素值统计特征、图像纹理特征以及血管形状特征得到最终视网膜图像的质量。在提出的数据库上测试结果显示,皮尔逊线性相关系数值高于0.96,斯皮尔曼等级相关系数值高于0.95。与现有评价方法相比,该方法优于传统的无参考评价方法,更能够客观的反映不同失真图像对血管分割这一应用的影响。

The distortion of retinal images is mostly caused by blurred and uneven illumination.In order to study the effects of different distortion types and different distortion levels on blood vessel segmentation,this paper quantifies the distortion type and distortion level of the image into the accuracy of image segmentation.Low distortion or uneven distortion images,constructs a database of retinal distortion images,including 2 distortion types,each distortion type has 8 levels of distortion,a total of 552 distorted retina images,and the accuracy of each distorted image as the label of the image.In addition,a no-reference quality evaluation method for retinal images based on blood vessel segmentation is proposed,and the quality of the final retinal image is obtained by extracting the statistical features of the pixel value of the retina image,the texture features of the image and the shape features of the blood vessel.The test results on the proposed database show that Pearson product-moment correlation coefficient is higher than 0.96 and the Spearman rank-order correlation coefficient is higher than 0.95.Compared with the existing evaluation methods,this method is superior to the traditional non-reference evaluation method and can objectively reflect the influence of different distorted images on the application of vascular segmentation.