Research On Robust Recognition Of Crop Leaf Diseases Based On Deep Convolution Neural Network

Crop disease recognition plays an important role in the accurate diagnosis and sci-entific control of crop diseases,and it is also an important guarantee for the yield and quality of agricultural products.China is a big agricultural country.The agricultural products are important for people and country.Agriculture is not only the foundation of human food and living,but also the cornerstone of national economic construction and development.The safe production of crops is a necessary condition to improve agricultural economic efficiency and to promote the sustainable development of agri-culture.It is also an important guarantee to build a harmonious society and constructing a new socialist countryside.With the improvement of living standards,people are seek-ing high quality of agricultural products.Crop diseases with high impact and large harm importantly limit the development of agricultural industry,and thus directly affect the yield and quality of crops.Therefore,the goal of this thesis is to study the robustness recognition of crop diseases in real environments,to effectively overcome the limitation of inaccurate di-agnosis of crop diseases due to the lack of professional knowledge,to promote the pop-ularization of the knowledge and experience of agricultural experts based on artificial intelligence technology,and to significantly improve the diagnosis and control effect of crop diseases.The main innovations of this thesis are:1.To deal with the real environments having poor recognition accuracy,we pro-pose a high-order residual convolutional neural network(HOResNet)for accurate rec-ognizing crop diseases.The HOResNet is capable of exploiting low-level features with object details and high-level features with abstract representation simultaneously in or-der to improve the anti-interference ability,thus improving the recognition robustness of the approach.Extensive experiment results on the AES-CD9214 dataset collected in natural environment demonstrate that the HOResNet approach achieves the highest accuracy on the datasets tested.2.To address the problem of weak robustness of crop disease recognition in prac-tical environment,we propose a high-order residual and parameter-sharing feedback convolutional neural network(HORPSF)for high accurate and strong robust crop dis-ease recognition in real scenes.The effect of HORPSF for the robustness of crop disease recognition is discussed.The high-order residual subnetwork is able to provide rich and detail representations for crop desease regions,thus improving the recognition accuracy of crop disease.The parameter-sharing feedback subnetwork can effectively depress the background noises and enhance the robustness of model.Extensive experiment results on public MK-D2 dataset demonstrate that the proposed HORPSF approach sig-nificantly outperforms other competing methods in terms of recognition accuracy and robustness,especially demonstrating superior performance on the AES-CD9214 dataset when dealing with the real environments examples of crop disease recognition.3.The characteristics of the complex background in crop disease image,the small disease area,and the small contrast between disease region and background that easily causes confusion between them,seriously affect the recognition robustness and accu-racy.To address this issue,we propose a Self-Attention Convolutional Neural Network(SACNN)that is used to extract validate features of crop disease spots and to identify crop diseases.The SACNN method includes basic network and self-attention network.The basic network is mainly used to extract the global features of the image,and the self-attention network is mainly used to obtain the local features of the lesion area.Extensive experimental results show that the recognition accuracy of SACNN method on AES-CD9214 and MK-D2 is 95.31%and 97.0%,respectively.The recognition accuracy of SACNN method on MK-D2 is better than state-of-the-art methods by 1.9%,which implies that the convolutional neural network with self-attention mechanism can focus on the important areas of the image,and thus can improve the recognition accuracy.Adding different levels of noise to AES-CD9214 test set shows the anti-interference ability and the strong robustness of SACNN method for noise.In addition,we discuss the influence of the location selection,channel size setting,network number and other aspects of the self-attention network on the recognition performance,in order to fur-ther clearly show the self-attention network working mechanism,helping to provide inspiration for other researchers.In summary,based on the proposed approaches mentioned above,this thesis achieves high accuracy and robustness of image recognition for crop diseases in real environ-ments,and provides technical support for accurate and robust diagnosis and control of crops,which is very important for theoretical investigations and practical applications.