Tobacco Disease Recognition And Detection Based On Convolutional Neural Network

Tobacco is an important economic crop in my country,and the tobacco industry has become an important industry in my country’s social economy.Tobacco diseases severely restrict the output and quality of tobacco,and affect the economic development of the tobacco industry and the income of tobacco farmers.Therefore,effective identification and reasonable control of tobacco diseases are essential to ensure the physiological health of tobacco leaves and increase the yield and quality of tobacco leaves.Tobacco leaf diseases have a wide variety and complicated pathology,but the current diagnosis method of tobacco diseases is still manual identification,which is poor in accuracy and low in efficiency.It is easy to cause misjudgment and cannot be effectively prevented.In response to the above problems,this paper proposes a method for identifying and detecting tobacco diseases based on Convolutional Neural Network(CNN),and preliminary studies on the precise identification and detection of tobacco diseases.The main work is as follows:(1)Construct a tobacco disease data set,and study a tobacco image preprocessing method based on image fusion.From May to September,the author collected tobacco leaves during the growth period of the tobacco fields in Ningyang,Laiwu,Linyi and other regions,and constructed the construction that contains common tobacco mosaic diseases.The image datasets of five tobacco diseases,tobacco,cucumber mosaic,brown spot,wildfire,and climate speckle,provide basic data for research on tobacco diseases.Considering that the data collection time span is large and the image is easily affected by factors such as light and shadow,this paper studies the tobacco image preprocessing method,uses the image enhancement method to expand the data set,and then uses the MSRCR method to eliminate the influence of light,and then use The Gamma correction method adjusts the image contrast,and finally the MSRCR image and the Gamma correction image are fused by the weighted average method to obtain an image fusion data set,thereby obtaining high-quality tobacco image data.(2)In this paper,the five most harmful tobacco diseases(common tobacco mosaic disease,tobacco cucumber mosaic disease,brown spot disease,wildfire disease,and climate spot disease)are the research objects,and a tobacco recognition model is constructed to realize a single species on a single disease picture.Classification of diseases.The proposed tobacco disease recognition model is based on the Inception V3 framework and is constructed using the transfer learning method,and uses four different data sets: the original data set,the data enhanced data set,the MSRCR data set,and the image fusion data set to be performed on the Inception V3 and Alex Net models.Training and testing.The test results show that the Inception V3 model achieves a disease recognition accuracy rate of 90.80% on the image fusion data set,and the accuracy rate on the original data set is 70.00%,compared with an accuracy increase of 20.80%.The recognition accuracy rate of the Alexnet model on the image fusion data set is 87.3%,which is also higher than the accuracy rate of 70.4% on the original data set,and the overall performance of the Inception V3 model is better than the Alexnet model.The experimental results show that image preprocessing can enhance the image quality,which is helpful for the training of the network model and the improvement of the model effect.The tobacco disease identification model proposed in this paper based on Inception V3 realizes the rapid and accurate identification of tobacco diseases,and provides a theoretical basis for the prevention and control of tobacco diseases.(3)This paper constructs a tobacco disease recognition and detection model based on YOLOv3,which realizes the recognition and detection of multiple diseases on a single leaf.The proposed method uses the darknet53 feature extraction network to extract the input disease image features and fuse the disease features at different scales,and generate a bounding box through a multi-scale detection network.Improve the YOLOv3 network,propose the kmeans++algorithm to cluster the width and height of the anchor box,obtain the optimal width and height,predict the category and position of the fusion feature,and finally remove the redundant frame through the non-maximum suppression algorithm(NMS),Get the final bounding box.This paper compares the proposed YOLOv3 disease detection model with the SSD(Single Shot multibox Detector)model.The test results show that the average IOU of the YOLOv3 and SSD models are 0.81 and 0.73,respectively.When IOU=0.5,the mAP of the YOLOv3 model used in this paper on the tobacco disease data set is 0.77,which is better than the SSD’s 0.69,which validates the model can effectively locate the tobacco disease area and provide information support for precise prevention and control.