Authors

Abstract

To solve the problem that existing agricultural unmanned plant protection equipment can't perceive crop growth on-site during corn seedling stage, this study proposes a crop row extraction method based on image processing. A semantic segmentation network was built with Unet framework, using VGG19 as encoder and transposed convolution for decoder with attention mechanism added. Traditional image processing was used to get key information. Data caching mechanism was introduced for crop row extraction. Automatic clustering algorithm was applied. Model training and testing show high accuracy and efficiency. This research can effectively segment crops from background and provide a reference for navigation and operation system of unmanned equipment.

Abstract in Chinese

为解决现有农用无人植保机具在玉米苗期作业时无法现场感知作物生长状况的问题，提出一种基于图像处理的玉米苗期作物行提取方法。基于Unet框架构建了作物语义分割网络，编码器为VGG19，解码器为转置卷积，通过添加注意力机制来增强主要特征信息的权重。采用传统图像处理技术获取作物的轮廓、最小外接圆的圆心与半径等关键信息，引入数据缓存机制，增加用于作物行提取的作物数据量。运用苗带自动聚类算法，在作物行数量与位置不确定的情况下，自动对作物行的数量及其所包含的作物对象进行聚类。模型测试结果表明：作物语义分割模型在训练集与验证集上的准确率分别达到0.9865和0.9864，损失率分别为0.0254和0.0270。在实际连续处理过程中，单张图片的语义分割时间平均为120毫秒，作物行提取时间为23毫秒。