基于改进SSD网络的西兰花叶片检测研究

Study on the detection of broccoli leaves based on the improved SSD network

实现对复杂环境下西兰花叶片的高精度检测,对西兰花病虫害的监测和防治具有重要意义。在田间复杂环境下叶片存在重叠或遮挡,增加了数据的获取成本,制约深度学习网络结构在田间复杂环境叶片检测上的应用。为此,提出基于改进SSD网络的西兰花叶片检测方法。通过特征辅助学习法更改训练集的组成,使SSD网络学习到较为完整的叶片边缘特征和叶片遮挡特征,采用数据增强和修改激活函数的方式构建模型。测试结果,单独叶片和植株叶片平均准确率为99.8%和89.9%,平均IOU为91.0%和84.2%,总体平均IOU和平均准确率分别为87.6%、94.9%。结果表明,使用小数据集进行西兰花叶片目标检测研究是可行的,可为农作物叶片检测研究提供参考。

It is significant that high precision detection of broccoli leaves in complex environments for monitor and prevention of broccoli diseases and pests.The frequent occurring of occlusion and overlapping condition not only increases the cost of agricultural data acquisition process,but also restricted deep learning network structure used in leaf detection.So this paper presents a method for detecting broccoli leaves based on improved SSD network.The feature-assisted learning method is used so as to facilitate the improved SSD network to study a complete set of blade edge and leaf occlusion features.The data enhancement and replaced activation function is used to construct the model.The test results showed that the average accuracy rates of individual and plant leaves were 99.8%and 89.9%respectively,the average IOU rates were 91.0%and 84.2%,the ensemble average IOU rate was 87.6%,and the average accuracy rate was 94.9%.The results show that it is feasible to use small data sets for the research of broccoli leaf target detection,which could provide a reference for crop leaf detection.