基于D-YOLOv3检测网络的温室叶菜幼苗图像检测

Image detection of leafy vegetable seedlings in greenhouse based on D-YOLOv3 detection network

为解决温室叶菜子叶期幼苗生长密集情况下的图像识别问题,提出一种密集连接型D-YOLOv3检测网络。该网络以YOLOv3为基础构建主干网络,改进检测结构和损失函数。以穴盘培育的油菜幼苗为例展开一系列试验。首先确定了YOLOv3和D-YOLOv3检测网络中损失函数的修正系数;其次通过构建的几种检测网络的对比试验验证了对YOLOv3主干网络、检测结构和损失函数改进的有效性,D-YOLOv3的幼苗检测精度高达93.44%,检测时间低至12.61 ms,与YOLOv3相比精度提升9.4个百分点,时间降低4.07 ms;最后进行不同密集程度和光照环境下幼苗图像的检测性能对比试验,结果表明D-YOLOv3的检测精度、检测时间及对小目标的特征提取能力均优于YOLOv3。D-YOLOv3能够对温室环境下的叶菜幼苗进行有效检测,可以为智能检测装备的作物识别提供依据。

In order to solve the problem of detecting densely planted leafy vegetable seedlings at cotyledon stage in greenhouse,a densely connected detection network named D-YOLOv3 was proposed.The backbone network of D-YOLOv3 was built based on that of YOLOv3.Simultaneously,the detection structure and loss function of D-YOLOv3 were constructed as the improvement of those of YOLOv3.A series of comparative tests were carried out with rape seedlings cultivated in pot as an example.Firstly,the correction coefficients of loss function in YOLOv3 and D-YOLOv3 detection networks were determined.Then,the effectiveness of the derived backbone network,detection structure and loss function of D-YOLOv3 was verified according to comparative tests with some networks mentioned in the paper.The test results showed that the detection accuracy of D-YOLOv3 was as high as 93.44%,and the detection time was as low as 12.61 ms.Compared with YOLOv3,the detection accuracy was improved by 9.4 percentage point and the detection time was reduced by 4.07 ms.Finally,the performance of YOLOv3 and D-YOLOv3 detection networks was obtained with images of rape seedlings growing under different density and illumination environments.The detection precision and speed of D-YOLOv3 for the proposed image were better than those of YOLOv3 at various densities.The ability of D-YOLOv3 was stronger than that of YOLOv3 to extract features of small targets under different illumination and density conditions.Consequently,the D-YOLOv3 detection network proposed in this paper can effectively identify rape seedlings in the greenhouse and serve as a powerful tool for crop recognition with intelligent detection equipment.