Study On Real-time Diagnosis Of Verticillium Wilt Of Cotton Based On Image Recognition

Verticillium wilt is the most serious and widely distributed cotton disease,known as "cotton cancer".When Verticillium wilt is serious in cotton,it is easy to cause yield reduction or even crop failure.Currently,cotton disease monitoring mainly relies on plant protection personnel to monitor cotton diseases by observing cotton leaves,based on experience or combined with modern laboratory testing methods.These traditional disease monitoring methods are time-consuming and laborious,and the means are lagging behind.Therefore,it is of great significance to improve the accuracy of information technology in the diagnosis of Cotton Verticillium wilt and provide a real-time and reliable diagnosis method of Verticillium wilt.In this paper,a real-time diagnosis method of Cotton Verticillium wilt disease based on image recognition is proposed by using cotton field surveillance video as data source.Through the research of video processing technology,the key frame extraction model of cotton field surveillance video was constructed,and the key frame image containing suspected Cotton Verticillium wilt disease information was acquired.The color space technology,morphological processing method and leaf segmentation method were studied,and the leaf segmentation model of suspected Cotton Verticillium wilt disease was constructed to realize the cotton leaf segmentation in the key frame image of cotton disease under complex background.The method of image processing and machine learning is studied,and the diagnosis model of Cotton Verticillium wilt disease is constructed,which realizes the real-time identification of Cotton Verticillium wilt disease.The main results of this study are as follows:(1)A key detection and extraction method of cotton field surveillance video is proposed.In this paper,key frame extraction techniques are discussed,which are k-Means Clustering Algorithms(k-Means),video framing,enhanced filtering,feature extraction and k-Nearest Neighbor(kNN).Firstly,all frames are extracted according to the video frame rate,and then the k-Means algorithm is used to perform the clustering operation.From the clustered result images,the nearest frame to the center of the class is found to be extracted as the key frame.According to the length of the video,the required number of pictures are output.Taking the experimental video as an example,250 frames are extracted from the 10 s video.After clustering these images,four pictures are obtained.The image is processed by low-pass filter,and the previous image is smoothed to reduce the error and noise.Then the structured data of the target image is generated according to the color histogram,colorcorrelation graph,color moment and wavelet parameters,and the data set is the feature value of the image.Finally,the kNN algorithm is used to train the image using the training set,and then the eigenvalues of the recognized image are substituted into the algorithm for recognition and judgment.The experimental results show that the accuracy of key frame extraction is more than 90%.(2)A leaf segmentation model based on key frame image of cotton disease in complex background is proposed.In this paper,plant image is segmented from the key frame image,and each easily recognizable leaf in the plant is extracted separately,i.e.image segmentation under complex background,which makes full preparations for subsequent image recognition,aiming at solving the basic problems of cotton disease recognition.Firstly,the RGB color feature of cotton leaves is used to separate the plant from the soil;secondly,the stem of the plant is removed by morphological processing technology to achieve color segmentation and retain the cotton leaves;finally,three image segmentation algorithms are used to separate the leaves,which are based on breadth search segmentation algorithm,OpenCV contour search function segmentation algorithm and watershed segmentation algorithm.Three segmentation algorithms are used to segment and extract cotton leaf images.Among them,the combination of breadth-based search segmentation algorithm and edge detection technology has significant effect on image segmentation with clear leaf structure,but compared with OpenCV contour search function,the latter has wider applicability and clearer contour hierarchy structure.Finally,the watershed segmentation algorithm and the improved watershed transform algorithm based on distance transformation are compared with the first two methods.The results show that the OpenCV based contour search function segmentation algorithm has the best effect.(3)The identification and diagnosis model of Cotton Verticillium wilt disease was constructed.In this paper,image processing technology and machine learning algorithm are used to validate the segmented cotton leaves.Normal cotton leaves,Verticillium wilt cotton leaves,Fusarium wilt cotton leaves and rotated disease cotton leaves are used as training sets.Support Vector Machine(SVM),Back Propagation Neural Network(BP)and Convolutional Neural Network(CNN)are studied respectively,and the three algorithms are compared.According to its unique local perception,weight sharing and downsampling methods,convolutional neural network can establish a training model combining image feature extraction,hierarchical structure and image classification,which has good invariance to image size change,feature location and rotation angle.The convolution neural network designed in this paper consists of nine layers,i.e.input-output layer,three convolution layer,three pooling layer and one fully connected layer.The influence of activation function,classifier,learning rate and iteration times on the final results of the algorithm is studied.The experimental results show that the accuracy of the convolution neural network to the test of diseased leaves is 96.67%,and the prediction accuracy is 88.75%.The feasibility and efficiency of the model in cotton Verticillium wilt identification and diagnosis were verified.(4)A real-time identification and diagnosis system for Verticillium wilt of cotton was designed and developed.Based on the research of theory and method,according to the method of software engineering,this paper designs and develops a whole system through the analysis of system functional requirements,the design of system business logic and overall architecture,the design of system database,the realization of system and the system test,etc.The cotton yellow is preliminarily realized.