Rice Variety Recognition Based On Deep Learning

Rice is one of the main food crops in China,and its variety identification is of great significance for the development of precision agriculture and crop yield estimation.The traditional rice variety identification mainly relies on manual work,which is subjective and inefficient,so it is impossible to analyze the large-scale production comprehensively.With the rapid development of machine learning technology,automatic identification of rice varieties has become possible.This paper takes RGB images of rice during the full growth period as the research object,and uses image processing techniques and deep learning methods to identify varieties.The main research work is as follows:(1)Enhancement and preprocessing of rice image data.In the natural light environment,1694 rice images of 8 varieties were collected by RGB cameras.The image data covers the full growth period of rice from the tillering stage to the mature stage after transplanting rice,and each picture contains rice and complex background.In order to improve the robustness of the deep learning model,the paper expanded the total amount of rice image data to 9722 by zooming out 0.8 times,0.7 times and 0.5times,as well as geometric transformations such as horizontal and vertical flips.Further,the size of the rice image data is uniformly adjusted to 32 * 32 pixels and normalized for rice variety identification.(2)Rice variety identification based on deep learning.A simple convolutional neural network model with 9-layer structure and a Res Net-18 deep residual network model are constructed,and the activation function,learning rate and other parameters of the model are optimized for rice variety identification.The rice image data sets were divided by 5-fold cross-validation,and the average recognition accuracy of five independent predictions was used as the model evaluation index.The experimental results show that the recognition accuracy of the two network models on the expanded rice image data is much higher than that of the original data,and the recognition rate on the rice panicle image data is better than that on the whole growth period data.Among them,on the expanded panicle image dataset EDHM,the recognition accuracy of the deep residual network model is 91.36%,and the training speed is faster;The simple convolutional neural network model with activation function Selu,the recognition accuracy is up to 94.03%,but the training speed is slow.(3)Rice variety identification based on the characteristics of convolutional neural networks and K nearest neighbor classifier.Based on simple convolutional neural network and deep residual network,the rice morphological features are extracted,and a rice variety recognition model based on K-nearest neighbor was constructed.The experimental results show that on the EDHM dataset,the recognition rate of KNN rice varieties using deep residual network features is 92.63%,which is significantly higher than the direct recognition rate,and is also better than the recognition rate of the deep residual network model itself.