基于卷积神经网络的温室黄瓜病害识别系统

基于图像处理和深度学习技术,该研究构建了一个基于卷积神经网络的温室黄瓜病害识别系统。针对温室现场采集的黄瓜病害图像中含有较多光照不均匀和复杂背景等噪声的情况,采用了一种复合颜色特征(combinations of color features,CCF)及其检测方法,通过将该颜色特征与传统区域生长算法结合,实现了温室黄瓜病斑图像的准确分割。基于温室黄瓜病斑图像,构建了温室黄瓜病害识别分类器的输入数据集,并采用数据增强方法将输入数据集的数据量扩充了12倍。基于扩充后的数据集,构建了基于卷积神经网络的病害识别分类器并利用梯度下降算法进行模型训练、验证与测试。系统试验结果表明,针对含有光照不均匀和复杂背景等噪声的黄瓜病害图像,该系统能够快速、准确的实现温室黄瓜病斑图像分割,分割准确率为97.29%;基于分割后的温室黄瓜病斑图像,该系统能够实现准确的病害识别,识别准确率为95.7%,其中,霜霉病识别准确率为93.1%,白粉病识别准确率为98.4%。

Disease recognition system for greenhouse cucumbers based on deep convolutional neural network

Abstract: Cucumber is one of the most common vegetables in China, which is severely affected by various diseases, such as downy mildew and powdery mildew. The process of recognizing diseases is often time consuming, laborious and subjective. Most of disease damage evaluation and treatment are done by farmers in the field with guidance of plant pathologists. Incorrect diagnosis and pesticide over usage are very common. Therefore, a timely and accurate recognition method of cucumber diseases is in great demand. Convolutional neural network is one of the most popular and best performing methods for image recognition. Because convolutional neural network has been extensively applied to agriculture applications, it is feasible to use convolutional neural network as the pattern recognition method for plant disease recognition. Convolutional neural network can automatically learn appropriate features from training datasets instead of manual feature extraction. The efforts on feature extraction and optimization can be saved. This not only reduces the computation cost, but also increases the accuracy and efficiency of the recognition. In this study, the state of the art convolutional neural network and deep learning techniques were applied to the recognition of cucumber diseases using visible leaf symptoms. A disease recognition system for greenhouse cucumbers based on convolutional neural network was presented in this paper based on deep learning and image processing. The key point of effective identification and diagnosis of diseases was to acquire the disease information accurately. With the development of computer vision technology, segmenting the disease symptom images from leaf images was presently considered as the main route of disease information acquisition. Color was the most direct information to discriminate disease symptoms from the other parts in a single image captured under real field conditions. Disease images captured under real field conditions were suffering from uneven illumination and complicated background, which was a big challenge to achieve robust disease symptom image segmentation. The symptom images were segmented by a novel image processing method using color information and region growing. Firstly, combinations of color features (CCF) and its detection method were presented. The combinations of color features consisted of three color components, excess red index (ExR), H component of HSV color space and B component of CIELAB color space, which implemented powerful discrimination of disease symptoms from clutter background. Then an interactive region growing method based on the comprehensive color feature map was used to achieve disease symptom image segmentation from clutter background. Input datasets was built from the symptom images. In order to decrease the chance of overfitting, data augmentation method that was to rotate the original datasets by 90, 160, 180 and 270 degrees and flip horizontally and vertically was utilized to enlarge the input datasets, which produced 12 augmented datasets. With the augmented input datasets, the system achieved good classification performance. Experiments were conducted to test the performance of the system. Results showed that the symptom image segmentation method can achieve an overall accuracy of 97.29%, which indicated that the method was capable of obtaining accurate and robust segmentation under real field conditions. The system achieved an overall accuracy of 95.7%, 93.1% for downy mildew and 98.4% for powdery mildew respectively, which indicated that the disease recognition system was capable of recognizing cucumber downy mildew and powdery mildew.