Research On Recognition Algorithm Of Wheat Head Scab And Its Severity Based On Deep Convolutional Neural Network

Crop diseases severely restrict the development of crops in our country and have a major impact on the food security of our country and even the world.As one of the staple foods of human beings,wheat is often affected by various diseases,among which head scab is one of the most important diseases in global wheat production.How to realize the accurate identification of wheat head scab and its severity has become an urgent problem in the current wheat production process.The conventional methods for diagnosing wheat head scab have the disadvantages of time-consuming,laborious,and high cost.Wheat head scab mainly causes ear rot.When the pathogen infects wheat ears,the chlorophyll,carotene and water content in the tissue cells of the wheat ears decrease rapidly,resulting in changes in the internal structure,morphology and color of the wheat ears.These changes can be sensitively captured by the image sensor.With the rapid development of artificial intelligence technology in image processing,recognition and other fields,new technical support is provided for the recognition of wheat head scab.The image processing technology and deep learning technology are combined to carry out in-depth research on wheat scab.The algorithms based on deep convolutional neural network are used to realize the recognition and severity estimation of wheat scab.The main research contents are as follows:(1)The wheat population image data set and wheat scab severity image data set in natural scenes are constructed.The Canon EOS 80 D handheld digital camera is used in this thesis to collect RGB images of wheat population in natural scenes and diseased wheat ears of different severity.A total of 1035 wheat image data are collected,including 510 wheat population images in natural scenes,525 images of wheat ears with different severity of head scab under controlled conditions.In order to improve data quality and increase sample diversity,this thesis preprocesses and augments the collected data samples.A total of 8513 wheat images are obtained after processing,including 940 images of wheat population in natural scenes and 7573 images of wheat ears with different severity of scab.(2)An image recognition algorithm for field wheat head scab based on multi-channel convolutional neural network is presented.Because of the interference of irrelevant targets in the wheat population images collected in the complex environment of the farmland,the identification of wheat head scab is seriously affected.In order to solve this problem,this thesis adopts a scab recognition algorithm based on a multi-channel convolutional neural network.The U-Net network is used first to perform segmentation processing operations to remove the interference of the complex environment of the farmland to obtain the segmented individual wheat spike image.Then,according to the color distribution characteristics of wheat head scab,a multi-channel convolutional neural network is constructed to extract the features of R,G,and B channels,and the fusion strategy is used to improve the feature strength.Finally,the joint loss function is used for training and learning to optimize the distance between data samples to realize the identification of wheat head scab.Through a large number of experimental studies.It is shown that the algorithm in this thesis can significantly improve the accuracy of the recognition of wheat head scab in the field environment,and its recognition effect is significantly better than other algorithms.(3)An algorithm for identifying the severity of wheat head scab based on multi-scale positioning learning is presented.Aiming at the problem of subtle feature differences between images of different severity of wheat head scab,a convolutional neural network based on multi-scale spatial positioning learning is designed to identify the severity of wheat head scab.The network is composed of a feature extractor module,a navigation network module and a scrutinizer network module,different modules promote and interact with each other.The feature extractor module is used to obtain the global features of the wheat ear image.The navigation network module is used to locate the local key areas of the disease with rich learning information.The scrutinizer network module is used to fuse the key area features and global features,and finally realize the severity of wheat scab Accurate identification.Experimental results show that the algorithm can achieve a recognition accuracy of 91.8%.The algorithm can effectively improve the accuracy of identifying the severity of wheat head scab compared with some commonly used networks,which can provide practical prevention for intelligent identification of the severity of crop diseases.