基于YOLOv8n的糖尿病视网膜病变检测算法  

Diabetic retinopathy detection based on YOLOv8n

作　　者：江祥奎 卢棋 董超 潘李冰 杨刚 苏耀恒[2] 赵明虎 JIANG Xiangkui;LU Qi;DONG Chao;PAN Libing;YANG Gang;SU Yaoheng;ZHAO Minghu(School of Automation,Xi’an University of Posts and Telecommunications,Xi’an 710121,China;School of Science,Xi’an Polytechnic University,Xi’an 710048,China)

机构地区：[1]西安邮电大学自动化学院,陕西西安710121 [2]西安工程大学理学院,陕西西安710048

出　　处：《西安邮电大学学报》2025年第2期85-92,共8页Journal of Xi’an University of Posts and Telecommunications

基　　金：陕西省重点研发计划项目(2022NY-087,2024GX-YBXM-300);陕西省社科联/陕西省应急管理厅项目(2021HZ1121);陕西省青年托举项目(20220129);陕西省教育厅科学研究计划项目(22JK0565);西安邮电大学研究生创新基金一般项目(CXJJYL2023071)。

摘　　要：为了提高糖尿病视网膜病变(Diabetic Retinopathy,DR)的检测精度,提出一种基于YOLOv(You Only Look Once Version)8n的DR检测算法。引入全维度动态卷积对YOLOv8n的骨干网络进行重构,以加强特征的适应性,提高特征提取能力。在颈部网络中增加小目标检测层,并引入轻量化双卷积以优化信息处理流程。最后对损失函数进行改进,优化梯度增益分配策略。实验结果表明,与YOLOv8n检测算法相比,所提算法的精确率提高5.8%,平均精度均值(Mean Average Precision,mAP)@0.5提高7.6%,有效改善了对DR小尺寸密集病灶的漏检问题。To enhance the detection accuracy of diabetic retinopathy(DR),a novel DR detection algorithm based on YOLOv(you only look once version)8n is proposed.The backbone network of YOLOv8n is reconstructed by introducing full-dimensional dynamic convolution,which strengthens the feature adaptability and improves the feature extraction capability.A small object detection layer is added to the neck network,while lightweight dual convolution is incorporated to optimize the information processing flow.Finally,the loss function is improved by refining the gradient gain allocation strategy.Experiment results indicate that compared to the YOLOv8n detection algorithm,the accuracy of the proposed algorithm is improved by 5.8%,and the mean average precision at a 0.5(mAP@0.5)is improved by 7.6%,which have effectively addressed the problem of oversight of small-sized,densely packed lesions in diabetic retinopathy detection.

关 键 词：糖尿病视网膜病变 视网膜图像 YOLOv8n 动态卷积 小目标检测层 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] R774.1[自动化与计算机技术—控制科学与工程]