Research On Tomato Leaf Disease Recognition Based On Machine Learning

According to the Food and Agriculture Organization of the United Nations(FAO)statistics for the period 2008-2017,the annual reduction in food production due to pests and diseases accounts for about 20-40 percent of total food production and economic losses of$120 billion worldwide.Crop disease control remains a major challenge for the agricultural sector.Early identification of crop diseases is critical to reduce pesticide use and mitigate economic losses.Fast,accurate and real-time information on the specific growth status of a crop can be obtained to enable timely variable application based on specific crop morbidity for accurate disease control.Research into the accurate and real-time diagnosis of crop diseases is of great strategic importance for the development of the agricultural economy of China and the world.Current image-based crop disease diagnostic methods can be broadly divided into two categories based on the differences in the way of extracting disease features from images:traditional machine learning-based methods and deep learning-based methods.Traditional machine learning methods rely on artificially designed disease features,usually set based on knowledge of the disease pathology of a specific target object.The deep learning approach,on the other hand,uses convolutional neural networks to automatically learn disease features from the training data.Although there have been studies on the application of deep learning methods to the identification of crop disease images,limited by the single source of training data,most of the training data images used are laboratory background images,which are different from the actual growing environment of the crop,resulting in low generalization performance of the trained models,and few studies have applied their results to actual production.Therefore,in this paper,we first collected tomato disease images from multiple sources,and expanded the image data by image processing to enrich the diversity of images and reduce the gap between the data sample sizes.Secondly,an image segmentation algorithm with an adaptive clustering number that incorporates the color index of crop disease images was investigated,and a identification study was conducted on tomato disease images from different sources,and an Android application was developed to provide a viable tomato disease diagnosis tool for crop production.Finally,by analyzing the characteristics of field background images,the accurate diagnosis of field background disease images containing multiple leaves was achieved using object detection algorithms.The main research in this paper is as follows.(1)A data set was established using images of tomato disease from three different sources.Tomato seedlings of Tomato yellow leaf curl virus disease and Tomato mosaic virus disease were transplanted in the laboratory greenhouse,and tomato disease images were taken during the full growth cycle using both white paper background and field background to form the basic data set.Collected tomato disease images from multiple disease mapping databases,such as the Arkansas Plant Diseases database,the American Phytopathological Society(APS)database,the Bugwood image database,and a number of academic research repositories at universities,to form the Internet data set.Also introduced 10 kinds of tomato disease images in the Plant Village data set.Through histogram equalization,grayscale transformation,whiteness,centralization and normalization,as well as image processing with multiple angles of rotation,panning,flipping,mirroring,scaling and cropping of the above three different sources of image data,the three different sources of image data can be expanded,eliminating the problem of unbalanced sample size in the disease image data set and enriching the diversity of images.(2)An adaptive clustering crop disease image segmentation algorithm with fused color index was investigated.Based on the need for background separation of leaves in crop diseases diagnostic studies of in crop images,two methods with lower arithmetic quantities,the color index segmentation method and the clustering segmentation method,were investigated in combination with the fact that most leaves are green.Using the color index as input data and then clustering by the k-means clustering method of adaptive clustering number,an adaptive clustering number crop disease image segmentation algorithm with fused color index is proposed to achieve fast and accurate crop disease image segmentation.The F1 and Entropy(E)values of the proposed algorithm and k-means,Ex G-Ex R,Mean Shift and DBSCAN methods were analyzed,and the corresponding F1 and E mean values were[0.982,0.118],[0.894,0.317],[0.947,0.246],[0.805,0.485] and [0.867,0.396],respectively.It is shown that the average F1 value and E value of the segmentation results of the proposed algorithm are the highest and the smallest,so the proposed algorithm can achieve accurate and fast background segmentation of tomato disease image with the best results.(3)The method of crop disease identification was investigated using data sets of crop disease images from different sources,which solved the problem of poor generalization performance of convolutional neural models trained from single-source data and improved the accuracy and robustness of crop disease identification in real-world.From the perspectives of network optimization,overfitting problem optimization and operational efficiency optimization,the important role of raw data diversity is studied,and the generalization performance of the model is effectively improved using global pooling,drop out,and batch normalization,with the final model test accuracy of 99.75%.Saliency maps and visualizations of maximum activation were also performed on the trained models to explore whether the features learned by the models matched the knowledge of crop disease pathology.To bring the research closer to real-world applications,an Android app was developed to help users quickly identify crop diseases.The results showed that the application enables rapid identification of nine leaf diseases and healthy leaves of tomato with high precision.(4)The effectiveness of the object detection method for diagnosing crop disease images with complex backgrounds in the field was investigated,and a diagnostic method for crop disease images with complex backgrounds and many leaves was proposed to enhance the robustness of the diagnosis.Training object detection models using data containing information on the location of target objects in the images,so as to achieve a single diagnosis in images of complex background diseases in fields containing multiple leaves,and outputting the results for each leaf in the image.The experimental results show that the mean average precision(m AP)of the proposed algorithm is 0.662(Io U >0.5).