Research On Segmentation And Recognition Of Tomato Leaf Lesions Based On Multiple Scale Convolution Structure

Tomato is one of the important agricultural economic products in China.In the process of greenhouse tomato cultivation,it is easy to produce a variety of diseases under the influence of environmental conditions,resulting in the decline of Tomato yield and quality.Timely and effective analysis of tomato leaf spot characteristics is helpful to quickly determine the type of disease,provide corresponding guidance and suggestions for disease control,so as to reduce economic losses.The traditional disease diagnosis model needs to segment the complex disease spots first,then design the feature extraction manually,and finally classify the image features to identify the disease types.It is difficult to meet the needs of efficient production of facility tomato due to the large workload and lack of timeliness.At present,deep learning uses convolution kernel to extract disease image features automatically,which provides a new solution for crop disease segmentation and recognition.However,It is difficult to extract the global and local features of tomato leaf lesions by using convolution structure of fixed receptive field,which affects the performance of disease segmentation and recognition.Moreover,the structure of tomato leaf lesions is complex and the number is large.It is difficult to obtain the marker data of tomato leaf lesions,which affects the segmentation training of tomato leaf lesions.In order to solve the above problems,two models based on multiple scale convolution structure are proposed to segment and identify tomato leaf diseases,and a tomato leaf disease diagnosis system for farmers is developed based on the two models.(1)Propose a model for segmentation and recognition of lesions with a combined structure of multiple receptive fieldsA multiple scale u-network model with multiple receptive field combination structure was proposed to solve the problems of different size,irregular shape and large number of pixel level markers in tomato leaf lesion segmentation.Firstly,the multiple scale residual module is used to set up a variety of receptive field combinations in advance.The model can freely select different receptive fields to extract disease features to adapt to the dynamic changes of lesion size and shape.Secondly,CB module is introduced Bridge)connects the disease feature extraction stage with the disease spot segmentation stage,uses a large number of image class labeling samples to train CB module,classifies the disease features,and obtains the key information of specific categories of diseases,so as to reduce the demand for a large number of disease spot labeling samples in the segmentation training stage;finally,uses a small number of pixel level labeling to train deconvolution module,so as to make the segmentation of the disease spot more accurate Accurate.In the original test set,the pixel accuracy,average pixel accuracy,average hybrid ratio and frequency weighted hybrid ratio of the model are94.37%,86.15%,75.25% and 90.27% respectively;in the interference test set with 30% brightness reduction and salt and pepper noise addition,the recognition accuracy of the model is 95.10% and 99.20% respectively.Experimental results show that the proposed method can improve the segmentation and recognition effect of tomato leaf lesions.(2)Propose a disease segmentation and recognition model with adaptive receptive field structureThe shape and size of tomato leaf lesions are complex.At present,the multiple scale feature extraction method used in the field of crop disease image recognition can only provide limited scale selection,but the size of plant lesions is diverse,and the convolution kernel with fixed size can not fully perceive all sizes of lesions.In order to further obtain a flexible receptive field range and enhance the adaptive ability of convolution kernel to extract image features,this paper improves the current popular hole convolution,proposes adaptive scale convolution,and designs an adaptive scale convolution network,which can learn the appropriate void ratio for different sizes of lesions in the disease image to obtain a more accurate receptive field range.At the same time,adaptive scale convolution network uses adaptive global average pooling layer and convolution layer instead of full connection layer for disease classification,which can reduce the amount of parameters.The experimental results show that the model can adaptively learn the receptive field range according to the input image,and the performance of lesion segmentation and disease recognition is further improved.In the original test set,the pixel accuracy,average pixel accuracy,average hybrid ratio,frequency weighted hybrid ratio and the recognition accuracy of the model are94.37%,86.15%,75.25%,90.27% and 99.54%respectively.(3)Develop a diagnosis system based on tomato leaf disease imagesIn this paper,based on a large number of tomato leaf disease image and a small number of disease mark data,with the help of convolution neural network model of multiple scale convolution structure,a tomato leaf disease diagnosis system for farmers is developed and deployed.The system is a web-based application system,which can quickly respond to the needs of disease identification and disease spot segmentation under the condition of smooth network.