Rice Disease Detection Technology Based On Semantic Segmentation

Crop diseases and insect pests not only severely restrict the yield of crops,but also affect people’s quality of life.Therefore,the timely diagnosis of crop diseases and insect pests is a very important task.However,the diagnosis of crop diseases and insect pests is still based on manual observation.This traditional method does not meet the needs of the development of intelligent agriculture.Therefore,with the development of deep learning,more and more researchers have begun to pay attention to the intelligent identification methods of crop diseases and insect pests.This paper takes three common types of rice diseases as research objects,and proposes a rice disease detection method based on semantic segmentation.The main research contents include: data set establishment,data analysis and network selection,and training and improvement of rice disease detection models.Firstly,establishment of three common rice disease data sets.At present,in the field of agricultural pest detection,it is difficult to obtain sample images,and image annotation requires a lot of manpower and material resources.Therefore,there is a lack of data available for training.In order to increase the diversity of samples,this paper expands the collected samples by image transformation,and manually annotates to make a semantic segmentation data set.Secondly,data analysis and network selection.The scale of the disease sample images collected in this paper is small.In order to reduce the risk of overfitting the model,a lightweight semantic segmentation network should be used for training.By comparing the performance of commonly used strong supervised semantic segmentation models and combining the characteristics of the data used in this study,U-Net was selected as the rice disease detection network in this paper.Thirdly,research on rice disease detection method based on improved U-Net.First,the self-made data set is used to train a rice disease detection model based on the U-Net network,and the experimental results are analyzed.Aiming at the problem of poor multiscale target detection effect in the experiment,the improved hollow space pyramid pooling(ASPP)module was introduced to improve U-Net,and the improved U-Net rice disease detection method was proposed,and a comparative experiment was conducted.The experimental results on the data set show that through the fusion of multiscale target information,the improved U-Net network has better detection performance for lesions of different scales,and the effectiveness of the improved network is verified.The rice disease detection method based on semantic segmentation proposed in this paper,the output result is a mask image,which realizes end-to-end detection,and the effect is intuitive,which can provide reference and reference for the automatic identification research of other diseases and insect pests in the crop field.