基于改进YOLOv7⁃tiny的画钟测验识别  

Recognition of clock drawing test based on improved YOLOv7⁃tiny

作　　者：温远寒 曹娜 刘怡欣[3] 何小海[1] 滕奇志[1] Wen Yuanhan;Cao Na;Liu Yixin;He Xiaohai;Teng Qizhi(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China;The Affiliated Hospital of Southwest Jiaotong University&the Third People’s Hospital of Chengdu,Chengdu 610065,China;National Clinical Research Center for Geriatrics,West China Hospital,Sichuan University,Chengdu 610044,China)

机构地区：[1]四川大学电子信息学院,成都610065 [2]西南交通大学附属医院&成都市第三人民医院神经内科,成都610031 [3]四川大学华西医院老年医学中心,成都610044

出　　处：《现代计算机》2024年第3期18-25,共8页Modern Computer

基　　金：四川省科技厅重点研发计划(重大科技专项)(22GJHZ0044);成都市重大科技应用示范项目(2019⁃YF09⁃00120⁃SN)。

摘　　要：画钟测验是筛查认知障碍人群的一种重要医学手段。针对目前画钟测验中存在目标尺度不同和类别不平衡的问题,提出一种基于改进YOLOv7⁃tiny的画钟测验识别算法。为改善尺度变化和小尺度目标检测带来的错检漏检问题,引入BiFPN双向特征金字塔结构,双向信息传递机制可有效融合不同层级特征,捕捉不同尺度特征中更丰富的上下文和细节信息。为提升类别不平衡指标的识别准确度,采用WDLoss损失函数计算损失提高小目标识别敏感性。此外还创建了一个基于认知障碍群体的画钟测验数据集,在此数据集上实验表明,改进后YOLOv7⁃tiny算法对画钟测验数据集所有类别的mAP为94.28%,相比于原YOLOv7⁃tiny模型提高了1.13%,不均衡类别中指针的AP提高了12.2%。The clock-drawing test is an important medical method for screening people with cognitive impairment.Aiming at the problems of different target scales and category imbalances in the current clock-drawing test,this paper proposes a clock-draw-ing test recognition algorithm based on improved YOLOv7-tiny.In order to improve the problem of false detection and missed detec-tion caused by scale changes and small-scale target detection,BiFPN bidirectional feature pyramid structure is introduced.The bi-directional information transmission mechanism can effectively integrate different levels of features and capture richer context and detailed information in different scale features.In order to improve the recognition accuracy of the category imbalance index,the WDLoss loss function is used to calculate the loss to improve the sensitivity of small target recognition.This paper creates a clock-drawing test data set based on cognitively impaired groups.Experiments on this data set show that the improved YOLOv7-tiny algo-rithm has a mAP of 94.28%for all categories of the clock drawing test data set,which is 1.13%higher than the original YOLOv7-tiny model,and the AP of the pointer in the unbalanced category is increased by 12.2%.

关 键 词：画钟测验 YOLOv7⁃tiny BiFPN WDLoss 

分 类 号：R749.1[医药卫生—神经病学与精神病学] TP183[医药卫生—临床医学] TP391.41[自动化与计算机技术—控制理论与控制工程]