计算机科学与探索 ›› 2021, Vol. 15 ›› Issue (9): 1762-1772.DOI: 10.3778/j.issn.1673-9418.2006067

• 图形图像 • 上一篇    

基于单种群蛙跳优化CNN的眼底图像多病变检测

任龙杰,孙颖,丁卫平,鞠恒荣,曹金鑫   

  1. 南通大学 信息科学技术学院,江苏 南通 226019
  • 出版日期:2021-09-01 发布日期:2021-09-06

Multiple Lesions Detection of Fundus Images Based on CNN Algorithm Optimized by Single Population Frog-Leaping Algorithm

REN Longjie, SUN Ying, DING Weiping, JU Hengrong, CAO Jinxin   

  1. School of Information Science and Technology, Nantong University, Nantong, Jiangsu 226019, China
  • Online:2021-09-01 Published:2021-09-06

摘要:

为了有效解决眼底图像病变处存在交织重叠,大小血管密布并且受光照影响严重等问题,实现眼底图像多标签分类,提出了采用单种群蛙跳优化的卷积神经网络算法(SFCNN)对眼底多种病变进行检测。该算法保留混合蛙跳算法(SFLA)的高效寻优能力,简化成单种群蛙跳算法,并与传统卷积神经网络(CNN)有效结合。在初始化网络时,通过蛙跳算法优化网络初始权值选择;在网络迭代过程中监听卷积神经网络前向传播损失值并利用蛙跳算法的寻优能力修正异常权值;在网络符合结束条件后对最终权值进行一次蛙跳寻优,使网络权值得到进一步的优化,从而实现对复杂的眼底图像多病变检测分类。该算法对眼底图像病变检测的实验表明,相对于传统CNN算法,无论是在单病变检测还是同时整体检测,正确率均有所提高。

关键词: 混合蛙跳算法(SFLA), 卷积神经网络(CNN), 眼底图像, 多病变检测, 权值优化

Abstract:

In order to effectively solve the problem of interlaced overlap in the fundus image lesions, large and small blood vessels and severely affected by light, and to achieve multi-label classification of fundus images, in this paper, a single population frog-leaping optimization convolutional neural network algorithm (SFCNN) is proposed to detect various fundus lesions. The algorithm retains the efficient searching ability of the shuffled frog leaping algorithm (SFLA). It is simplified into a single population frog-leaping algorithm and effectively combined with the traditional convolutional neural networks (CNN). When initializing the network, the initial weight of the network is optimized by the frog-leaping algorithm. In the process of network iteration, the forward propagation loss of convolutional neural network is monitored and the abnormal weight is corrected by using the optimization ability of frog-leaping algorithm. After the network meets the end conditions, the final weight is optimized by frog-leaping, which further optimizes the network weight, so as to realize the detection and classification of complex fundus image with multiple lesions. The experiment of the detection of fundus image lesions shows that compared with CNN algorithm, the accuracy of the proposed algorithm is improved in both single lesion detection and overall detection.

Key words: shuffled frog leaping algorithm (SFLA), convolutional neural networks (CNN), fundus image, detection of multiple lesions, weight optimization