基于改进卷积神经网络的复杂背景下玉米病害识别

为解决田间环境复杂背景下病害识别困难、识别模型应用率低的问题,提出了一种基于改进卷积神经网络的玉米病害识别方法,探讨了数据集的品质对建立模型性能的影响。利用复杂背景下的玉米病害图像进行数据增强、背景去除、图像细分割和归一化等处理,设计了具有5层卷积、4层池化和2个全连接层的卷积神经网络结构,利用L2正则化和Dropout策略优化网络,对复杂背景下的玉米9种病害进行识别训练和测试,优化后的CNN模型平均识别精度为97.10%,比未优化的网络模型提高9.02个百分点。利用不同大小、不同品质的数据集对优选网络进行训练和测试,数据增强后比原始样本平均识别精度提高了28.17个百分点;将复杂背景去除后,模型性能进一步提升,识别精度达到97.96%;对数据集进行细分割处理后,平均识别精度为99.12%,表明卷积神经网络需要大量的训练数据,且数据集需有一定的代表性和品质。开发了基于移动端的玉米田间病害识别系统,系统测试结果表明,平均识别准确率为83.33%,系统能够实现田间复杂环境下的玉米病害识别。

Corn Disease Recognition under Complicated Background Based on Improved Convolutional Neural Network

Aiming to solve the problem of difficulty in disease recognition and low application rate of recognition model in complex field environment, a corn disease recognition method based on improved CNN was proposed. The influence of data set size and quality on the performance of disease recognition model was discussed. The complicated background images were used and preprocessed by using augmentation, background removal, local fine segmentation and normalized processing methods. Then the CNN structure was designed by using five convolutional layers, four pooling layers and two fully connected layers. The L2 regularization and Dropout strategy were utilized to optimize the network. The results showed that the optimized model achieved an average precision of 97.10% implemented in nine kinds of diseases images with complicated background, which was increased by 9.02 percentage points than that of unimproved CNN method. Compared with the model with original sample, the average precision of the model trained with augmented data was increased by 28.17 percentage points. The removal of complex background can eliminate the influence of environmental noise on the model, thus the performance can be further enhanced, which reached an average precision of 97.96%. After the local fine segmentation of data set, the average precision was raised up to 99.12%. These indicated that CNN needed a large number of representative and high quality training data to identify the target feature. On the basis of improved CNN, a corn field disease recognition system based on mobile was developed, which achieved 83.33% recognition accuracy when verified by experiment in the field. The research could provide guidance for disease recognition and precise prevention and control in corn field.