基于图像和改进U-net模型的小麦赤霉病穗识别

Identification of Fusarium Head Blight in Wheat EarsBased on Image and Improved U-net Model

为了快速、有效地监测小麦赤霉病的发生情况,利用数码相机对人工接种赤霉病菌的小麦田进行RGB图像获取,在图像预处理基础上,利用深度学习网络U-net来对人工标记好的发病麦穗图像进行训练。通过深度学习数据集的构建与测试,建立了基于RGB图像和改进U-net网络模型的小麦赤霉病识别与监测模型,并对模型识别结果进行了验证。结果表明,U-net可以很好地提取图像波段信息,但对于比较复杂的麦穗图像,在使用Keras方法进行图像语义分割时,需要对U-net网络结构进行改进,即在下采样部分加入Dropout层。与人工标记结果相比,模型识别结果的一致性较好,具有较高的监测精度。该模型平均精度为0.9694,损失函数值为0.0759,平均交并比MIoU为0.799。上述结果说明改进的U-net模型可以很好地识别和监测小麦图像中的发病麦穗,并在发病麦穗的分割上具有很好的效果.

In order to quickly and effectively monitor the occurrence of wheat Fusarium head blight,the RGB image of wheat field inoculated with Fusarium head blight was acquired by digital camera.On the basis of image preprocessing,the deep learning network U-net was used to train the artificially labeled diseased wheat ear images.Through the construction and testing of deep learning data set,a wheat Fusarium head blight identification and detection model based on RGB image and improved u-net network model was established,and the recognition results were verified.The results showed that U-net can extract the band information of image well,but for complex wheat ear image,when using Keras method for image semantic segmentation,we need to improve the U-net network structure,that was,adding dropout layer in the down sampling part.Compared with the results of manual labeling,the model recognition results had good consistency and high detection accuracy.The average accuracy of the model was 0.9694;the loss function value was 0.0759,and the average intersection merging ratio(MIoU)was 0.799.The model could detect the diseased wheat ear in wheat images,and had a good effect on the segmentation of diseased wheat ears.