Research On Rice Image Data Enhancement Method Based On WGAN-G

Rice is one of the three major grains in the world,and over half of the population in China relies on rice as their main food.The safety of rice production is crucial to the country’s food security.With the development of intelligent agriculture,the automation of rice cultivation is more demanding,and the accurate and rapid diagnosis of rice diseases is one of the research hotspots.At present,the main method of rice disease recognition is deep neural network,but the sample number is limited because of the difficulty of collecting rice disease leaves,it is impossible to provide large-scale and high-quality training samples for deep neural network,so it is necessary to enhance the small sample data set.Therefore,it is necessary to enhance the small sample dataset.On site investigation in Heilongjiang reclamation area,it was found that rice blast,brown spot,and leaf blight are the most common.Therefore,this thesis mainly focuses on the data augmentation methods for these three diseases in the dataset.A progressive WGAN-GP(PWGAN-GP)is used to generate rice disease images for small sample dataset augmentation.The algorithm principle of WGAN-GP network is investigated,and a multiscale progressive approach is introduced on this structure to propose a multiscale progressive WGAN-GP network to progressively enhance the resolution of the generated images and to enhance the stability of the model training and generated images.The validation experimental results show the optimal image generation effect of PWGAN-GP.The recognition accuracy of VGG-16,Goog Le Net and Res Net-50 using PWGAN-GP data augmentation method was improved by 10.9%,13.4% and 13.3%,which proved that the method can effectively improve the accuracy of convolutional neural network in recognizing disease images.Based on the PWGAN-GP rice disease image generation model,the Vi T-PWGAN-GP network is proposed by introducing the Vi T module with Transformer structure.The generation model uses the Vi T module to learn global features in the feature space of the target samples and gradually increases the resolution of the generated images.The discriminant model is constrained based on the WGAN-GP structure.The validation experimental results show that the Vi TPWGAN-GP image generation is better than the comparison model.Image data augmentation using Vi T-PWGAN-GP improves the recognition accuracy of VGG-16,Goog Le Net and Res Net-50 by 15.3%,14.7% and 14.3%,respectively.The results of the validation experiments on the rice image dataset showed that Res Net-50 achieved 98.7% accuracy for the recognition of three rice diseases after data augmentation with Vi T-PWGAN-GP.Design and implementation of rice disease image data augmentation system.The system includes three functional modules: image acquisition,image generation and image recognition.The system integrated image generation model can enhance the small sample data set and realize the fast diagnosis of three kinds of common rice disease images.