基于多路卷积神经网络的大田小麦赤霉病图像识别

Image recognition of field wheat scab based on multi-way convolutional neural network

为了准确地识别小麦病害,及时采取防治措施,减少农药施用的成本,同时减少农业生态环境的污染,该研究以灌浆期感染赤霉病的小麦麦穗图像为研究对象,根据病变区域与健康区域的颜色分布特点,设计了一种多路卷积神经网络用于小麦赤霉病图像的识别。首先利用深度语义分割网络U-Net对大田环境下的小麦图像进行分割,去除小麦叶片及其他无关背景的影响,从而分割出麦穗图像。然后设计结构较为简单的多路卷积神经网络分别提取麦穗图像R、G、B 3个通道的特征,通过特征融合获得具有高辨识性的麦穗图像语义特征。最后,为了增大赤霉病和健康麦穗图像特征之间的可区分性,同时减小赤霉病麦穗图像类内特征的差异,采用联合损失函数进一步改善网络的性能。该研究对采集的大田环境下的510幅灌浆期小麦群体图像进行分割,选取2745幅完整单株麦穗图像利用所设计的多路卷积神经网络进行赤霉病识别试验,结果表明该研究所提算法对单株麦穗赤霉病识别精度达到100%,能够为小麦病害的智能识别提供帮助。

Various wheat diseases can seriously deteriorate the quality and yield to decline significantly,thereby to restrict the high-quality and sustainable development of modern agriculture in China.An accurate and efficient identification of wheat scab become urgent to control the spread of pests and diseases,and to guarantee for the wheat yield.In this study,a research object was taken as the image of wheat ears that infected with scab during grouting period.A multi-way convolutional neural network was designed to identify the wheat scab images based on the color distribution characteristics of the diseased and healthy areas on the research object.Firstly,a deep semantic segmentation network U-Net was used to segment the wheat images in the field environment to remove the influence of wheat leaves and other unrelated backgrounds,particularly on densely growing wheat ears,the cluttered backgrounds of wheat leaves and soil,and complex outdoor lighting.Since the segmentation can efficiently reduce the noise from complex backgrounds,the image of a single wheat ear was segmented for the subsequent image recognition of wheat scab.Then,a simple multi-way convolutional neural network was used to extract the feature information via three R,G,B channels of the wheat ear images.Three feature vectors can be output by the convolutional neural network,and then be merged by vector stitching at the end of the network.After the last pooling layer,the feature vectors were selected to be stitched in the channel dimension in order to form thicker features.This processing can enrich the features described by the wheat data sample in a high-level feature vector.Finally,a joint loss function was used to further improve the performance of the network,particularly on the image detection from the infected wheat ears with head blight,in order to reduce the detecting time for the difference between the features within the head of blight mildew images.After the segmentation,2745 images of complete single wheat scab were obtained via the comparison of 510 wheat images that collected in the field environment.A multi-way convolutional neural network combined color channels can increase the width of entire network,and enhance the utilization of each layer of channels,indicating that each layer can learn rich features,such as texture features in different directions and different frequencies.The experimental results showed that the joint loss function learning can increase the distance between different classes,whereas reduce the distance between the same class,thereby to make the network extract more robust features for the identification of wheat scab based on the multi-way convolutional neural network.The findings demonstrated that the proposed algorithm can achieve 100%recognition accuracy for wheat scab,and further provide a valuable support for the intelligent identification of wheat diseases.