基于增强特征融合YOLOV5的视网膜病变检测  

Retinal disease detection based on enhanced feature fusion YOLOV5

作　　者：韩璐 毕晓君 HAN Lu;BI Xiaojun(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China;School of Information Engineering,Minzu University of China,Beijing 100081,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001 [2]中央民族大学信息工程学院,北京100081

出　　处：《应用科技》2022年第1期66-72,共7页Applied Science and Technology

摘　　要：针对视网膜黄斑病变数据集缺失以及视网膜图像冗余度过大问题,建立了包含3种视网膜病变的视网膜黄斑疾病检测数据集,并提出了一种基于改进YOLOV5的视网膜病变检测模型。该模型在特征提取网络中引入了改进的注意力机制模块,突出病变区域,降低视网膜图像中大量背景的影响。其次,改进加强特征提取网络,加权融合具有大量细节信息的浅层特征,增强网络对视网膜病变的定位能力。实验结果表明,本文模型具有良好的视网膜病变检测效果,检测精度达97.3%。In view of the lack of retinopathy dataset and the excessive redundancy of retinal images,a retinopathy detection dataset is established,including three kinds of retinopathy,and a retinopathy detection model is proposed based on improved YOLOV5.The model introduces an improved attention mechanism module into the feature extraction network to highlight the lesion area and reduce the influence of large numbers of background in the retinal image,and further,improves and enhances the feature extraction network,weights and fuses shallow features with a large amount of detail information,and enhances the ability of the network to locate retinopathy.The experimental results show that the model proposed in this paper has good effect for detecting retinopathy,and the detection accuracy is 97.3%.

关 键 词：深度学习 神经网络 目标检测 视网膜 光学相干断层扫描 黄斑病变 医学图像 注意力机制 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]