基于微调卷积神经网络迁移学习模式下被害棉叶图像的识别

【目的】研究一种基于卷积神经网络的危害棉叶症状识别技术,提高棉花病虫害的识别准确率。【方法】基于caffe深度学习框架,在CaffeNet网络结构基础上增加一层全连接层(记为CaffeNet+1),并结合迁移学习方法对网络进行训练。采集健康、红叶茎枯、红蜘蛛、枯萎、黄萎、双斑萤叶甲、蚜虫、褐斑棉叶图像各975张作为样本集。随机选取验本集中80%的图像样本作为训练集,剩余20%作为测试集。【结果】迁移学习方式下学习率取0.005时的CaffeNet+1模型最优,在测试集上其识别准确率可达98.9%。【结论】在与全新学习模式下的CaffeNet模型相比,该方法可加速网络模型收敛,且具有更高的识别准确率,该技术方法在准确识别田间病虫害棉叶后表现症状的图像写出来具体方面具有重要的应用价值。

Preliminary Study on Image Recognition of Damaged Cotton Leaf Based on Fine-tuning Convolution Neural Network Transfer Learning Model

【Objective】 In order to improve the recognition accuracy of cotton diseases and insect pests, a method based on convolution neural network was proposed to identify the symptoms of harmful cotton leaves. The occurrence of cotton diseases and insect pests can cause severe yield loss in cotton. Identification of diseases and pests is key to the prevention and control of diseases and pest. 【Method】To improve the identification accuracy of cotton diseases and insect pests, a method for identification cotton leaf disease and pest based on improve convolutional neural network and transfer learning were proposed. 【Result】The experimental results showed that in the dataset containing 7,800 images of healthy cotton leaves and leaves of seven kinds of diseases and pests (divided into training set and test set at the ratio of 4∶1), under the transfer learning method, the CaffeNet+1 model with the learning rate of 0.005 was optimal, and the recognition accuracy of the test set was 98.9%. 【Conclusion】Compared with the CaffeNet model under the new learning model, this method can accelerate the convergence of the network model and has higher recognition accuracy. This technique has important application value in accurately identifying the image of symptoms of pests and diseases of cotton leaves in the field.