Abstract: Accurately identifying the location of hard exudate in retinal fundus helps to reduce the risk of blindness caused by diabetic retinopathy. An approach based on IDA-RF is proposed for hard exudate detection. It needs to pretreat the fund image and then extract the exudate candidate region, which using the [k]-means clustering to initialize the population. Meanwhile, combining with the universal gravitation search algorithm, constantly updating the step formula, an Improved Dragonfly Algorithm（IDA） is proposed. IDA optimizes the parameters for the random forest algorithm in the optimization process, classifies the exudate region by using the optimized random forest algorithm, which aims to extract the position of the final hard exudate. This method is tested in a public fundus image database, compared with RF, DA-RF and GSA-RF, the accuracy of hard exudate is 97.28%. The experiments show that this method can accurately identify the hard exudates and has good robust performance.

Key words: fundus image, hard exudate, gravitational search, dragonfly algorithm, random forest algorithm

摘要： 为了准确检测眼底图像中的硬性渗出物，降低糖尿病性视网膜病变引起的失明，提出了一种基于IDA-RF的眼底硬性渗出物的检测方法。对眼底图像预处理，提取渗出物候选区域。利用[k]-means初始种群，与万有引力搜索算法相结合，改变步长更新公式，提出一种改进的蜻蜓算法（IDA）。IDA在寻优过程中对随机森林算法参数进行优化，并利用优化后的随机森林算法对渗出物候选区域分类，提取最终精确的硬性渗出物。该方法在公开的眼底图像数据库进行实验，与RF、DA-RF、GSA-RF相比，准确率达97.28%。实验表明，提出的方法能够准确检测硬性渗出物且鲁棒性能好。

关键词: 眼底图像, 硬性渗出物, 引力搜索, 蜻蜓算法, 随机森林算法