Study On Leaf Lesion Segmentation And Recognition Strategy Under Complex Background

The appearance of crop pests and diseases will be directly reflected in the appearance,and the shape,color and texture of the lesion area and non-marking area will be greatly changed.Whether it is by means of detection tools or manual identification,there will be many cases of missed judgments and misjudgments,and it takes a long time and costs,and cannot meet the rapid judgment standard of precision agriculture.At present,most of the automatic detection techniques for crop pests and diseases at home and abroad are based on digital image processing and spectral analysis imaging technology.With the development of artificial intelligence technology,computer vision and image processing technology are widely used in the field of plant disease identification,and image processing and deep learning methods can be used to identify plant disease types more quickly and accurately.In this paper,deep neural network is used to study the deep features of plant lesions,and the deep convolutional neural network model is trained to realize the pretreatment and recognition of 20 lesions in 9 crop leaves under complex background.The main research contents are as follows:(1)The traditional GAC model is improved and the AD-GAC model is established by constructing a new edge detection function by adding anisotropic diffusion filter operator.According to the difference of pixel values between the leaf part and the lesion part,the AD-GAC model and the maximum entropy threshold method were used to segment the crop leaf lesions under complex background.(2)Using the operations of rotation,flipping,zooming,and panning to expand the sample,effectively solve the problem of insufficient sample size;then,using the raw data as input,the lesion image is obtained through the AD-GAC model,and resize the size of the lesion image according to the backbone network;finally,use transfer learning to fine-tune it.By redefining the loss function,the joint supervision mechanism of cross entropy and central loss function is used to expand the difference between the features of the class,and the difference of the features within the class is reduced.Then a AD-GAC CNN with two-stage network structure is proposed.Experiments show that the model trained by AD-GAC CNN based on transfer learning and joint supervision can accomplish the task of crop leaf lesion identification well under the condition of insufficient sample size and complex background.(3)A software integration system integrating the functions of denoising,segmentation and recognition of lesions has been developed.There are several models for each part of the function for the user to choose,the user can choose the model type according to actual needs.