Research On Accurate And Real Time Detection Of Corn Leaf Blight Based On Improved Convolution Neural Network

The prevention and control of crop diseases has become an important issue in agricultural production.Accurate disease detection is an important part of disease control.At present,the research on crop disease detection based on computer vision technology has made important progress.However,the disease detection methods based on traditional machine vision need to rely on artificial feature selection,and cannot fully reflect the characteristics of crop diseases,the detection accuracy needs to be improved.Convolutional neural network,as one of the representative algorithms of deep learning,has strong autonomous learning ability,and does not need to select features manually.The detection accuracy is higher than that of traditional machine vision method.But in the field complex environment,the existing disease detection model based on convolutional neural network still has the problems of low detection accuracy and poor real-time performance.In addition,although the acquisition of disease image has gradually changed from manual shooting to machine shooting,the quality of the images taken by automatic equipment is uneven.Some image shot by mechanical arms or UAV have been too bright,and the detection effect of the disease image remains to be tested.In view of this,this paper takes the image of corn leaf blight obtained by hand-held camera,mechanical arm and UAV as the research object,and constructs the improved image preprocessing model and disease detection model.New model realizes the high precision detection of maize blight.The specific contents and results are as follows:(1)Aiming at the problem of lack of image preprocessing in general domain target detection,this paper proposes a method of image decomposition and reconstruction based on convolutional neural network to realize the preprocessing of high light image.The software optimized image and the original image are used as input,and the separation of incident image and reflection image is realized by convolution.In the decomposition process,the image decomposition process is regulated by reflectivity and light smoothness.In the illumination optimization module,in order to obtain more image information,a multi-scale cascade structure is designed.The continuous convolution layer and activation function will sample the image to obtain a wide range of illumination distribution.The cascade structure not only makes use of the large-scale illumination information reasonably,but also brings the module with adaptive adjustment ability to optimize the illumination intensity.The optimized three kinds of data sets have the detection accuracy of 75.85%,71.95% and 68.77% respectively in the traditional SSD.(2)To solve the problems of low accuracy of disease detection in the one-stage network and slow speed of disease detection in the two-stage network,this paper proposes a fine-tuning module in one-stage network.Through the layer screening of generated anchors,the detection module can obtain more effective information and solve the problem of detection accuracy decrease due to too many negative samples.In addition,the reasonable and effective anchor screening mechanism also improves the detection speed of the detection module on the target diseases.In order to further improve the performance of the model,a feature fusion transmission module is designed between the fine-tuning module and the detection module.Through this module,the information sharing between the fine-tuning module and the detection module can be realized,and the traditional algorithm layer by layer information transmission mode is replaced,the overall performance of the model is effectively improved.In addition,the feature fusion module also strengthens the information fusion between different layers.Due to the effective information acquisition method of the front layer network,this paper designs less detection layers in the detection module,which further improves the detection efficiency.The optimal precision of the improved model in three different data sets is 91.37%,89.32% and 85.77%,of which the highest FPS is 32.9.It meets the requirements of real-time detection of disease.In conclusion,the disease detection model based on convolutional neural network can accurately detect the corn leaf blight in real time.The results show that the model can provide technical support for the accurate detection of corn leaf blight,and can provide reference for the software development of crop disease detection.It has certain theoretical research significance and practical application value.