论文部分内容阅读
穗瘟是一种严重影响水稻产量及品质的多发病害,有效地检测穗瘟是水稻病害防治的重要任务。该文提出基于深度卷积神经网络GoogLeNet模型的水稻穗瘟病检测方法,该方法利用Inception基本模块重复堆叠构建主体网络。Inception模块利用多尺度卷积核提取不同尺度穗瘟病斑分布式特征并进行级联融合。GoogLeNet利用其结构深度和宽度,学习复杂噪声高光谱图像的隐高维特征表达,并在统一框架中训练Softmax分类器,实现穗瘟病害预测建模。为验证该研究所提方法的有效性,以1 467株田间采集的穗株为试验对象,采用便携式户外高光谱成像仪Gaia Field-F-V10在自然光照条件下拍摄穗株高光谱图像,并由植保专家根据穗瘟病害描述确定其穗瘟标签。所有高光谱图像-标签数据对构成GoogLeNet模型训练和验证的原始数据集。该文采用随机梯度下降算法(SGD,stochastic gradient descent)优化GoogLeNet模型,提出随机扔弃1个波段图像和随机平移平均谱图像亮度的2种数据增强策略,增加训练数据规模,防止模型过拟合并改善其泛化性能。经测试,验证集上穗瘟病害预测最高准确率为92.0%。试验结果表明,利用GoogLeNet建立的深度卷积模型,可以很好地实现水稻穗瘟病害的精准检测,克服室外自然光条件下利用光谱图像进行病害预测面临的困难,将该类研究往实际生产应用推进一大步。
Head blast is a serious disease which affects the yield and quality of rice severely. It is an important task to detect head panicle blast disease effectively. This paper presents a method to detect rice panicle blast based on the GoogLeNet model of deep convolution neural network. This method uses Inception basic modules to repeatedly build the host network. The Inception module uses multiscale convolution kernels to extract the distributed features of panicle blast lesions at different scales and to perform cascade fusion. GoogLeNet uses its structure depth and width to learn hidden high-dimensional features of complex noise hyperspectral images, and Softmax classifier is trained in a unified framework to predict the frostbite disease. In order to verify the validity of the proposed method, panicle plants collected from 1 467 fields were used as experimental subjects. Hibiscus hyperspectral images were taken under natural light conditions using a portable outdoor hyperspectral imager Gaia Field-F-V10 Phytophthora pneumophila labels were identified by plant protection experts based on the description of panicle blast disease. All hyperspectral image-tag data pair the original dataset that forms the GoogLeNet model training and validation. In this paper, the GoogLeNet model is optimized by using the stochastic gradient descent (SGD) algorithm. Two data enhancement strategies are proposed, such as randomly discarding 1-band image and stochastic mean-shift image brightness, increasing the training data size and preventing model overfitting And improve its generalization performance. After testing, the highest predictive accuracy of ear blast on the validation set is 92.0%. The experimental results show that the deep convolution model established by GoogLeNet can accurately detect rice panicle disease and overcome the difficulties in forecasting the disease using spectral images under outdoor natural light conditions. This type of research can be applied to practical production and application A big step.