Rice Disease Image Recognition Research Based on Convolutional Neural Network

PurposeIn order to solve the problem that traditional rice disease recognition technology has strong dependence on specific features of images and low recognition efficiency, we proposed to apply deep learning theory to rice disease identification in order to obtain better recognition results.

MethodThe rice disease recognition model was established by using the deep convolution neural network. The data of three main diseases of rice were normalized, and deep learning framework Keras was used to perform deep CNN training. By setting different convolution kernel sizes and pooling functions, the three common diseases of rice were identified and identified.

ResultExperimental results show that the convolution kernel size was 9×9 and the pooling function adopted the maximum pooling, and the model recognition rate was the highest; the accuracy rate of the model after 5 iterations could reach more than 90%; the image tended to be stable and the model basically reached convergent after 6 iterations. From the perspective of model performance analysis, the loss function showed a gradient descent trend, the change was relatively stable, and the prediction loss deviation gradually decreased.

ConclusionThis model has the characteristics of strong generalization ability, high accuracy, good robustness and small loss rate, which provides a reference for the research of plant disease identification.