基于YOLO v5-OBB与CT的浸种玉米胚乳裂纹检测  

Endosperm Crack Detection Method for Seed Dipping Maize Based on YOLO v5-OBB and CT Technology

作　　者：宋怀波[1,2] 焦义涛 华志新 李嵘 许兴时 SONG Huaibo;JIAO Yitao;HUA Zhixin;LI Rong;XU Xingshi(College of Mechanical and Electronic Engineering,Northwest A&F University,Yangling,Shaanxi 712100,China;Key Laboratory of Agricultural Internet of Things,Ministry of Agriculture and Rural Affairs,Yangling,Shaanxi 712100,China)

机构地区：[1]西北农林科技大学机械与电子工程学院,陕西杨凌712100 [2]农业农村部农业物联网重点实验室,陕西杨凌712100

出　　处：《农业机械学报》2023年第3期394-401,439,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划项目(2019YFD1002401)。

摘　　要：浸种是玉米生产中重要的播前增种技术,对浸种过程中裂纹的高效检测是分析玉米胚乳裂纹变化规律的基础,是优良品种性状选育的关键之一,尚存在内部胚乳裂纹不可见、自动化检测程度不高等困难。基于CT扫描技术,在YOLO v5n检测网络的基础上,设计了YOLO v5-OBB旋转目标检测网络,其中OBB为有向目标边框,该网络使用旋转矩形框代替普通矩形框,并在Backbone部分加入位置注意力模块(CA),同时采用倾斜非极大值抑制算法(Skew-NMS)进行非极大值抑制得到最终预测框,以此实现长宽比大、方向不一的玉米胚乳裂纹检测。经过300次迭代训练,模型在测试集上的精确率P为94.2%,召回率R为81.7%,平均精度(AP)为88.2%,模型内存占用量为4.21 MB,单幅图像平均检测时间为0.01 s,与SASM、S2A-Net和ReDet旋转目标检测网络相比,AP分别提高15.0、16.9、7.0个百分点,单幅图像平均检测时间分别减少0.19、0.22、0.46 s,同时YOLO v5-OBB模型内存占用量分别为SASM、S2A-Net和ReDet模型的1.50%、1.43%和1.73%,与采用水平矩形框标注的YOLO v5网络相比,AP提高0.6个百分点,模型大小减小0.19 MB,单幅图像平均检测时间不变,两者均为0.01 s。将YOLO v5-OBB网络获取裂纹目标框坐标信息后得到的裂纹长度与在DragonflyEZ软件中得到的裂纹真实长度相比,两者绝对误差为0.04 mm,相对误差为0.93%。对不同CT灰度分布情况下玉米胚乳裂纹检测结果表明,该模型对较小灰度、较大灰度、混合灰度3种玉米胚乳裂纹图像的P分别为100%、100%、93.3%,R分别为100%、82.4%和79.8%,AP分别为99.5%、91.2%和86.8%。结果表明,所设计模型能有效实现玉米胚乳裂纹的检测,同时模型鲁棒性高,内存占用量小,可为玉米浸种过程胚乳裂纹的自动监测提供借鉴。Seed immersion is an important pre-sowing seed enhancement technology in maize production,and efficient detection of cracks during seed immersion is the basis for analyzing the change pattern of endosperm cracks during seed immersion,which is one of the keys to the selection and breeding of good varieties of traits,and there are still difficulties such as internal endosperm cracks are not visible and the degree of automation is not high.Based on CT scanning technology,a rotating target detection network named YOLO v5-OBB was designed based on YOLO v5n detection network,where OBB used rotating rectangular box instead of normal rectangular box and added CA model in the Backbone part.The network used a rotating rectangular box instead of a normal rectangular box,and added CA model in the Backbone part,and also used Skew NMS for non-maximal suppression to obtain the final prediction box,so as to achieve the detection of corn endosperm cracks with relatively large length and width and different directions.After 300 iterations of training,the model had a precision of 94.2%,a recall of 81.7%,and an average precision of 88.2%on the test set,with model size of 4.21 MB and average detection time of 0.01 s for a single image,which improved the AP value by 15.0,16.9,and 7.0 percentage points compared with the SASM,S2A Net,and ReDet models,respectively,and the average detection time of single image was reduced by 0.19 s,0.22 s,and 0.46 s,respectively,while the YOLO v5-OBB model size was 1.50%,1.43%,and 1.73%of the SASM,S2A Net,and ReDet models,respectively,with an increase in AP value of 0.6 percentage points,a decrease in model size of 0.19 MB and an unchanged average detection time of 0.01 s for a single image compared with that of the YOLO v5 network with horizontal rectangular box labeling.Comparing the crack length information obtained from the YOLO v5-OBB network after obtaining the crack target frame coordinate information with the real length of the crack obtained in DragonflyEZ software,the absolute error of both was

关 键 词：玉米胚乳裂纹 YOLO v5 CT扫描 旋转目标检测 机器视觉 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]