Plant Disease Recognition And Pest Detection Under Unstructured Environment

Diseases and pests are the major enemy in the crop growth,they occur throughout the growth period of crops because of the diversity and variety.In recent years,with the increase in production level and climate change,diseases and pests have continued to increase in incidence and recurrence,which directly affects national food security.With the rapid development of computer science,the ability of automatic detection and recognition of plant diseases and pests based on machine learning technology has surpassed some agricultural experts in some aspects.However,most of the machine learning methods are indirect feature weight learning method,the accuracy and robustness of the disease recognition as well as pest detection depend on the quality of the feature extraction model and the disease data as well as pest data need to be periodically verified and expanded during the training process.Thus,how to improve the fitting degree of the pest detection as well as disease recognition model under unstructured environment in order to carry out the large-scale on-site,real-time and accurate pest detection and disease recognition under unconstrained as well as unstructured in-field environment and realize engineering application is the key technical issue that need to be resolved.This dissertation focuses on the in-depth research and discussion of the disease recognition as well as pest detection towards unstructured environment based on the essential principles of image classification and object detection technology.The main research content in this dissertation is as follows:1.This dissertation propose a plant disease recognition method based on cascading convolutional neural network.Its main idea is completing the feature classification and feature mapping of multi-class plant disease as well as their disease severity through cascading ideology and constructing multi-level plant disease classification method under unstructured environment.First,considering the discrepancy of different deep learning models,we employ different convolutional neural networks as backbone and construct a parallel-voting algorithm for plant disease image classification,so that the accurate plant disease category is calculated.Next,the weighted Euclidean distance between different disease severities can be measured by the weighted siamese network and the corresponding plant disease severity is determined.Combining the results of the two networks,the category and severity of plant disease under unstructured environment are calculated at the same time.Besides,in order to further improve the robustness of the plant disease classification model,this dissertation propose a new test enhancement method to optimize and guide the plant disease classification process.A large number of experiments show that the method proposed in this dissertation can effectively improve the accuracy of plant disease classification and disease severity estimation.2.This dissertation propose a pest detection method which integrates in-field multimodal environmental information.The main idea is to integrate the occurrence conditions of in-field pest as multimodal information into pest detection algorithm and provide semantic prior-knowledge guidance for pest detection models.First,we need build a rough classification network for pest images.Due to the diversity in harmed crop,occurrence time,location,and environmental information of different pests,we construct a multimodal information extraction network,so that the pest data can be roughly divided according to the multimodal information including crop categories.Then,we need design a fine pest detection network.For the pest detection model based on the specific crop,we introduce the projection convolution module which can extract pest multimodal information to alleviate the small pest feature vanishing issue in convolutional neural network.Compared with other object detection methods,the proposed method in the dissertation is more effective and robust.3.This dissertation proposed a new object detection loss function to effectively solve the feature conflict issue in the object detection pipeline which provides a new idea for the further optimization of the two-stage object detection technology.Through the experiment of the region proposal assignment process in the object detection pipeline,the feature conflict issue between the training stage and the inference stage is analyzed and quantified.Aiming to the deviation of the classification confidence of the positive and negative samples predicted by the network as well as the intersection over union between predicted samples and ground truth,the dissertation uses a strong supervision method to suppress such feature conflict issue.A large number of experimental results demonstrate that the loss function proposed in this dissertation makes the object detection pipeline based on deep neural networks has better performance.