面向深度学习识别高空农作物的方法

Method for deep learning to identify high altitude crops

为探讨深度学习技术运用于高空或卫星图像处理的问题,针对其高空层次特征丰富但细节信息不充分,以及背景多元化干扰等特点,研究一种有效的多分类方法。利用消费级无人机采集农作物高空图像,从图像数据规模有限,较难实现有效训练的问题出发,提出一种综合数据增强和迁移学习的方法克服数据集不足;结合高空图像的独特特征,改进高维空间的最优分类函数,对模型进行优化,使之更吻合高空拍摄农作物图像的识别与处理。通过构建多组数据、多种模型的对比实验,验证了该方法的有效性与性能,为农业智慧决策提供有益补充。

To explore the problem of deep learning technology applied to high-altitude or satellite image processing,an effective multi-classification method was studied for its high-altitude feature richness,insufficient detail information and background multi-dimensional interference.Using consumer-grade drones to collect high-altitude images of crops,starting from the problem that image data is limited in scale and difficult to achieve effective training,a comprehensive data enhancement and migration learning method was proposed to overcome the shortage of data sets.Combined with the unique characteristics of high-altitude images,the optimal classification function of high-dimensional space was improved,the model was optimized to make it more consistent with the recognition and processing of high-altitude crop images.The effectiveness of the method was verified by constructing multiple sets of data and comparison experiments of multiple models.Performance provides a useful complement to agricultural wisdom decision making.