Research On Recognition And Segmentation Of Plant Leaf Diseases Based On Deconvolution And Weakly Supervised

Plant disease is one of the major factors that reduce crop productivity,effective plant disease detection is crucial in the development of efficient,ecological,and environmental agriculture.Making an early diagnosis of the disease according to the observed characteristics of leaf surface,which could helps to spray pesticide more precisely and efficiently,serves as an important plant disease detection application.With the development of computer vision technology,deep learning has been widely applied to the recognition and segmentation of plant leaf disease.Different from traditional methods that need to extract the features of interest manually,deep learning methods are able to extract the features automatically,leading to the reduction of research and development cycle of disease detection model as well as the improvement of the detection accuracy.However,the generalization ability of deep learning methods usually relies on a large number of accurately labeled data,which are often not accessible in real-world applications.For example,the label of disease needs to be determined by professional experts,which is time-consuming and laborious because of the variable shape and large number of disease spots,making it hard to collect large amount of accurate disease labels.When meeting leaf images collected from real farmland,the feature extraction process of deep learning models will be disturbed by the complex background environment,variable illumination conditions or other issues.In addition,the symptoms of some diseases are quite similar.As a result,the existing semantic segmentation models are not able to distinguish such fine-grained disease categories.To address these drawbacks,in this paper,we focus on the plant leaf diseases recognition and segmentation with a weakly supervised framework,and aim at designing a novel recognition and segmentation method for plant leaf diseases,where the segmentation results are used to estimate the severity of the diseases.To sum up,the main contributions of this work can be listed as follows:(1)Traditional recognition models of plant leaf disease usually lack transparency,since these end-to-end deep classifiers are susceptible to shadow,occlusion,and light intensity.The segmentation models of plant leaf diseases are not capable to distinguish the category of disease area and heavily rely on a lot of pixel-level labels.To address the drawbacks mentioned above,we proposed a novel recognition and segmentation model for plant leaf disease based on guided deconvolution,called Deconvolution-Guided VGGNet(DGVGGNet).On the one hand,the skip connections and reversed fully connected layers are used to transfer information of the feature and the category to the deconvolution module.On the other hand,only a few pixel-level labels are deployed by the deconvolution module to train the segmentation model,and the back propagation process is used particularly to transfer the location of disease to the recognition model.After the mutual information transmission,the model can realize the recognition and segmentation of plant leaf diseases simultaneously under the limited pixel-level labels.Experimentally,the results demonstrate that DGVGGNet had strong robustness under the condition of shielding and light interference.(2)Existing deep learning recognition models usually require a lot of labeled data,which is often not practical in real-world applications.To solve this challenge,we proposed a novel recognition and segmentation model of plant leaf disease under a semi-supervised learning framework,where DGVGGNet and Mix Match are employed.Firstly,we constructed the pseudo labels of unlabeled data by consistency regularization and entropy minimization.Then,the labeled and unlabeled samples were mixed by the Mix Up operation,and the mixed images were used for semi-supervised disease classification.Finally,we upsample the category labels,and the resulting label vector as well as a few pixel-level labels were used for semi-supervised disease segmentation.Moreover,during model training,we deployed the exponential weighted moving average method to guarantee the model robustness.(3)After the disease segmentation process,it is necessary to segment the leaves to estimate the severity of the disease.In order to solve the problem that semantic segmentation model under deep learning frameworks ususlly heavily rely on a lot of pixel-level labels,a segmentation model based on leaf semantics is further proposed by using leaf species information,called Teacher and Student Network(TSNet).For laboratory background images,only image-level labels were used to train TSNet model,leading to the initial leaf segmentation results.For the complex background images,a few pixel-level labels were used to fine-tune the teacher and deconvolution module in the TSNet model,which can obtain Teacher Net.The fine-tuning method could improve the segmentation accuracy of the leaves in the complex background.Finally,these two leaf segmentation models and the DGVGGNet model were used to estimate the severity of the disease and the percentage of the disease area.