A field wheat pest and disease identification method and device and storage medium

Abstract

This invention provides a method, device, and storage medium for identifying wheat diseases and pests in the field, belonging to the field of image processing. The method uses a neural network (DenseNet) to extract richer wheat disease and pest features; it optimizes the disease feature information output by each group of convolutions in DenseNet through feature pyramid fusion; it incorporates a convolutional attention module to optimize the output features of multiple convolutional layers to suppress the influence of complex background features on the performance of wheat disease and pest identification in the field; and it uses the Sobel filtering algorithm edge protocol header to detect the edge features of lesion areas, using these edge features to guide the model to predict the mask generation direction, thereby improving the identification speed of this invention. This method performs multiple feature optimizations on wheat disease and pest features through feature fusion and feature suppression during the feature extraction process, suppressing useless features introduced by complex backgrounds and highlighting the effective features of wheat disease and pest targets, thereby improving the accuracy of disease and pest identification.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a field wheat disease and pest identification method and device and a storage medium. BACKGROUND

[0002] Wheat is the most widely planted crop in the world, and its yield and quality have an important impact on human life. About 17.6% of the world's wheat is produced in China, and ensuring wheat yield safety is an important foundation for national security, social stability and economic development. However, in China, the annual yield loss caused by wheat diseases and pests is about 16.29% of the annual wheat yield. The four diseases that have a great harm and a wide impact on wheat include wheat stripe rust, wheat powdery mildew, wheat scab and wheat yellow dwarf disease; the three pests include the long-tubed aphid, the two-pronged aphid and the grass aphid. An effective method to reduce the yield loss caused by diseases and pests is to monitor wheat diseases and pests in a timely and accurate manner to take effective control measures. Therefore, efficient and accurate monitoring of wheat diseases and pests is a basic prerequisite for correct and effective control measures and for increasing wheat yield.

[0003] The current main method for disease monitoring and identification is manual field investigation and detection during the entire wheat cultivation process. Manual investigation of wheat diseases and pests not only requires a large amount of manpower, but also is easily affected by the subjective experience of researchers, leading to incorrect judgments, especially in the early stages of disease infection caused by similar phenotypic symptoms of different organisms and non-biological stress. However, the use of intelligent monitoring solutions can not only avoid the defects of the above-mentioned manual identification methods, but also play an indispensable and important role in accelerating the upgrading of the grain planting industry, improving the resource utilization rate of the grain planting industry, popularizing digital management of the grain planting industry, and increasing the income of the growers. The commonly used intelligent identification method is realized based on computer vision technology, including machine learning and image processing algorithms. In fact, under complex field conditions, machine learning-based methods still need to be designed by artificial features through feature engineering. Crop disease and pest detection based on image processing algorithms also faces many challenges, such as the slight symptom differences existing in the early stages of disease infection, limited shooting angles, and noise interference such as soil, weeds, water stains and shadows, which can greatly reduce the accuracy of the detection algorithm.

[0004] The existing technical solutions for intelligent identification of wheat diseases and pests are mainly divided into two categories: machine learning methods and deep learning methods, and the problems of the existing technical solutions are as follows:

[0005] There are two main problems in the method based on machine learning, one is that the algorithm used by the machine learning based technology is relatively simple, and the recognition accuracy and speed of the classification model based on machine learning need to be improved when facing complex field environment; the second is that the establishment period of the machine learning classification model needs to rely on manual design and extraction of excellent disease characteristics, and due to the large difference in the size and type of wheat disease area and the dependence on subjective experience judgment, the feature extraction work still exists the problems of low efficiency and inaccuracy;

[0006] There are two main problems in the method based on deep learning, one is the problem of insufficient data set richness, although the deep learning method does not need manual feature design and extraction work, but training a deep learning neural network needs a large amount of data set support, and the data set used for training in the aspect of wheat disease and pest is mainly the disease and pest image in simple background, the accuracy of the network model trained by using the disease and pest image data set in simple background will be greatly reduced when tested in the actual field condition; the second is the problem of recognition accuracy and speed, in order to realize intelligent and efficient automatic wheat disease and pest monitoring, the accuracy and recognition rate of intelligent recognition model must be ensured. In fact, crop disease detection under complex field conditions faces many challenges, for example, the difference of slight symptoms in early infection stage, limited shooting angle and noise interference such as soil, weeds, water stains and shadow will have a great influence on the recognition performance of the model. In addition, since the disease characteristics of various wheat diseases and pests are shown on the wheat leaves, and the characteristics of the long and thin wheat leaves also bring difficulties to the intelligent recognition method of wheat diseases and pests. The existing deep learning recognition technology of wheat diseases and pests still has a large space for improvement in model accuracy and speed. SUMMARY

[0007] In order to solve the problems of low feature extraction efficiency, low accuracy and great influence of external factors on the recognition performance of the model in the existing wheat disease and pest recognition method, the present application provides a field wheat disease and pest recognition method, device and storage medium.

[0008] In order to achieve the above purpose, the present application provides the following technical scheme:

[0009] A field wheat disease and pest recognition method, comprising the following steps:

[0010] Collecting the disease and pest image of wheat, and constructing a data set;

[0011] Inputting the disease and pest image in the data set into a disease and pest recognition model, and outputting a feature map;

[0012] Sending the feature map into a region proposal network (RPN) to generate a candidate frame of the wheat disease area;

[0013] The feature map in the candidate frame is classified and recognized by the prediction network to obtain a wheat disease and pest prediction result map;

[0014] The disease and pest image in the data set is input into the disease and pest recognition model to output a feature map, and specifically includes:

[0015] The neural network DenseNet-169 is used as the backbone feature extraction network of the convolutional neural network Mask-RCNN, and the disease and pest image in the data set is preliminarily feature-extracted by the neural network DenseNet-169 to obtain a disease feature.

[0016] The feature pyramid FPN and the convolutional attention module CBAM are added to the backbone feature extraction network DenseNet-169 to multi-optimize and fuse the disease feature to generate a feature map.

[0017] Preferably, the disease and pest image is an image collected at different time periods for different varieties of wheat.

[0018] Preferably, the disease and pest image in the data set is input into the neural network DenseNet-169 to obtain a disease feature, and specifically includes:

[0019] The disease and pest image in the data set is input into the neural network DenseNet-169;

[0020] The neural network DenseNet-169 extracts the feature map C2, the feature map C3, the feature map C4 and the feature map C5 from the bottom-up pyramid structure DenseBlock(1), DenseBlock(2), DenseBlock(3) and DenseBlock(4) respectively.

[0021] Preferably, the disease and pest image in the data set is input into the neural network DenseNet-169 to obtain a disease feature, and specifically includes:

[0022] The top-level feature map P5 is generated by 1x1 convolution feature dimension reduction and CBAM adaptive feature refinement of the horizontal input feature map C4;

[0023] When the top-level feature map P5 is fused with the horizontal input feature map C4, the top-level feature map P5 is up-sampled by a 2x nearest neighbor up-sampling method;

[0024] Before generating a new feature map P4, the CBAM is applied to the feature map obtained in the previous step to re-allocate the feature map and the feature weight;

[0025] Based on the same strategy, feature map P3 is generated using feature map C3 and feature map P4, and feature map P2 is generated using feature map C2 and feature map P3;

[0026] Feature maps P2-P5 all need to be subjected to 3x3 convolution to generate a group of feature maps named Predict; Predict6 is obtained by performing maximum pooling downsampling on Predict5;

[0027] Feature maps Predict2-Predict6 are fused into one feature map as the input of the subsequent region generation network, and feature maps Predict2-Predict5 are fused into another feature map as the input of the subsequent region of interest network.

[0028] Preferably, the prediction network predictionNetwork comprises a classification branch, a prediction box branch and a mask branch;

[0029] The classification branch contains the category information of the diseases contained in the image, the prediction box branch contains the coordinate information of the disease position, and the mask branch contains the detailed edge information of the disease position. Combining the information in the three branches obtains the final wheat disease and pest prediction result map.

[0030] Preferably, before the preliminary feature extraction of the disease and pest images in the data set by DenseNet-169, the data set is expanded by using data enhancement technology, specifically: using mirror conversion, color enhancement, random rotation and brightness enhancement four methods to expand each original disease and pest image in the data set.

[0031] Preferably, after data augmentation of the data set, open source software LabelMe is used to label the image data after data augmentation in JSON format, and the labeled image data is divided into a training set, a test set and a validation set. The disease and pest recognition model is trained through the training set, tested through the test set, and verified through the validation set.

[0032] Preferably, when the disease and pest recognition model is trained through the training set, the edge protocol header is used to perform Sobel filtering on the prediction mask and the labeled mask to obtain the edge features of the prediction mask and the labeled mask, and an additional loss function L edge is added to the basic loss function, L edge The calculation formula of L

[0033]

[0034] Wherein, p represents the mean square error, M prepresents the p-th power of the generalized power mean, y represents the edge information of the annotation mask, represents the predicted mask edge information;

[0035] The loss function L of the pest and disease identification model our The calculation formula is:

[0036] L our = L cls + L box + L mask + L edge

[0037] Wherein, L cls , L box , L mask respectively are the classification loss, the prediction frame loss and the mask loss of the Mask-RCNN network;

[0038] The loss value is fed back to the pest and disease identification model through the loss function L our To optimize the weight parameters;

[0039] The pest and disease identification model is trained for multiple rounds of iteration until the model performance tends to be stable and the loss function L our Converge to save the weight parameters.

[0040] Another purpose of the present application is to provide a field wheat pest and disease identification device, comprising:

[0041] The data acquisition module is used for acquiring the pest and disease image of the wheat and constructing a data set;

[0042] The pest and disease identification module is used for inputting the pest and disease image in the data set into the pest and disease identification model and outputting a feature map;

[0043] The candidate frame generation module is used for sending the feature map into the region candidate network RPN to generate a candidate frame of the wheat disease area;

[0044] The result prediction module is used for classifying and identifying the feature map in the candidate frame through the prediction network Predict Network to obtain a wheat pest and disease prediction result map;

[0045] The pest and disease identification module comprises:

[0046] The feature extraction model is used for taking the neural network DenseNet-169 as the backbone feature extraction network of the convolutional neural network Mask-RCNN, preliminarily extracting the features of the pest and disease image in the data set through the DenseNet-169 to obtain disease features;

[0047] The feature generation model is used for adding a feature pyramid FPN and a convolution attention module CBAM into a backbone feature extraction network DenseNet-169, and multiple optimization and fusion are performed on disease features to generate a feature map.

[0048] The application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the method.

[0049] The field wheat disease and pest identification method provided by the application has the following beneficial effects:

[0050] (1) The application adopts the idea of a Mask-RCNN instance segmentation model, has a mask generation branch, and the generated disease mask can more accurately determine the disease position in the image.

[0051] (2) The method uses a densely connected neural network DenseNet-169 to extract more rich wheat disease and pest features, and the disease feature information output by each group of convolution of the DenseNet is fused and optimized by means of a feature pyramid; a convolution attention module is added to optimize the output features of multiple convolution layers to suppress the influence of complex background features on the field wheat disease and pest identification performance; a Sobel filter algorithm edge protocol head is used to detect the edge features of the disease spot area, and the edge features are used to guide the mask generation direction of the model to improve the identification speed of the application. The DenseNet-FPN-CBAM structure designed by the application has stronger feature extraction capability, fewer parameters, and the wheat disease and pest features are optimized multiple times through feature fusion and feature suppression in the feature extraction process, the useless features introduced by the complex background are suppressed, and the effective features of the wheat disease and pest target are highlighted, thereby improving the identification accuracy of the disease and pest. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the application and the design scheme thereof, the drawings required by the embodiments will be briefly introduced as follows. The drawings in the following description are only part of the embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0053] Figure 1 The flow chart of the field wheat disease and pest identification method provided by the application is shown in the figure.

[0054] Figure 2 The effect diagram of the field wheat disease and pest identification method provided by the application is shown in the figure. DETAILED DESCRIPTION

[0055] In order to better understand the technical solutions of the present application and to enable one skilled in the art to carry out the present application, the present application will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0056] Firstly, the proper nouns involved in the present application are explained.

[0057] 1. Mask Region-based Convolutional Neural Network (Mask-RCNN)

[0058] Mask-RCNN is an object detection and instance segmentation algorithm. The algorithm model includes a backbone network module for feature extraction, a region proposal network module for region of interest extraction, and a prediction network module for object and mask prediction.

[0059] 2. Densely Connected Convolutional Networks (DenseNet)

[0060] DenseNet is a feature extraction network with dense connection channels between the front-end convolutional layer and all back-end convolutional layers.

[0061] 3. Feature Pyramid Network (FPN)

[0062] FPN is a basic component in a recognition system for enriching the semantic information and location information contained in the feature map to detect objects of different scales.

[0063] 4. Convolutional Block Attention Module (CBAM)

[0064] CBAM is a structure that sequentially cascades a channel attention submodule and a spatial attention submodule to optimize the weight distribution of the input image feature map.

[0065] 5. Edge Agreement Head (EAH)

[0066] EAH is a method of guiding the prediction direction of the instance mask edge by edge filtering and participating in the calculation of the loss function.

[0067] Embodiment 1

[0068] The application is based on the Mask-RCNN network model, and is designed and innovated in view of the phenotype characteristics of wheat diseases and insect pests, and provides a field wheat disease and insect pest identification method, specifically a precise and efficient field wheat disease and insect pest intelligent identification method based on multiple feature optimization. Specifically, a dense connected neural network (DenseNet) is used to extract more rich wheat disease and insect pest features; the feature pyramid (FPN) is used to fuse and optimize the disease feature information output by each group of convolutions of the DenseNet; a convolution attention module (CBAM) is added to optimize the output features of multiple convolution layers to suppress the influence of complex background features on the field wheat disease and insect pest identification performance; the Sobel filter algorithm edge protocol head is used to detect the edge features of the disease spot area, and the edge features are used to guide the model to predict the mask generation direction, thereby improving the identification speed of the application.

[0069] Specifically, as shown in Figure 1 The field wheat disease and insect pest identification method provided by the application comprises the following steps:

[0070] Step 1, constructing a complete wheat disease and insect pest data set.

[0071] Step 101, collecting wheat disease and insect pest images in a complex background in a wheat field. In order to increase the richness and complexity of the data set and fully restore the complex situations that the model needs to cope with in actual application, more than 200 different wheat varieties are collected, the collection equipment includes cameras and smartphones, and the collection time includes early, middle and late three time periods, a total of 2236 images are collected.

[0072] Step 102, expanding the data set by using data enhancement technology. The data enhancement technology can increase the diversity of samples and improve the robustness of the model. The application uses four enhancement methods, namely mirror conversion, random rotation, color enhancement and brightness enhancement, to increase the diversity of image samples. The application respectively performs four kinds of transformation on each original disease and insect pest image in the data set, and then obtains four new images, that is, the data set is expanded by 1:5. In this way, the data set is expanded by four times, from 2236 images to 11180 images, so that more images can be used for model training, so that more suitable model parameters for production requirements can be fitted in the training process.

[0073] Step 103, using the open source software LabelMe to mark the image data after data enhancement in JSON format. The marked image data is divided into a training set, a test set and a validation set, the disease and insect pest identification model is trained through the training set, the disease and insect pest identification model is tested through the test set, and the disease and insect pest identification model is verified through the validation set.

[0074] Step 2, the disease and pest recognition model DenseNet-FPN-CBAM is used to extract the wheat disease characteristics and optimize the extracted characteristics.

[0075] Although Mask-RCNN has high target positioning accuracy according to its ROI pooling mechanism, the classic Mask-RCNN model still has problems of large calculation amount, slow speed and rough target mask contour, which is difficult to meet the requirements of high-efficiency and real-time identification of wheat diseases and pests in the field. Therefore, based on the classic Mask-RCNN framework, a model suitable for extracting crop disease and pest characteristics is designed from the perspective of feature optimization.

[0076] Step 201, the backbone feature extraction network DenseNet with better feature extraction capability and less parameters and calculation amount is used to preliminarily extract the features of the wheat disease and pest picture. Compared with the backbone feature extraction network ResNet in the classic Mask-RCNN model, DenseNet improves the feature extraction capability of the network and greatly reduces the number of network parameters through feature reuse and residual channel. After rigorous control experiments, the 169-layer DenseNet (DenseNet-169) with the best performance for wheat disease and pest feature extraction is selected as the backbone feature extraction network of the application.

[0077] Step 202, the FPN-CBAM module is designed to add to the backbone feature extraction network DenseNet-169 to optimize the extracted disease features.

[0078] As Figure 1 As shown in the feature extraction and feature map generation module in the flowchart, the entire DenseNet-FPN-CBAM module is connected in a pyramid shape. Specifically, the model is composed of two parts, the feature extraction part is the DenseNet-169 network from bottom to top pyramid ( Figure 1 left structure in the middle), and the feature fusion part is the top-down pyramid with horizontal connection ( Figure 1 right structure in the middle).

[0079] In the feature extraction part, the neural network DenseNet-169 extracts feature maps C2, C3, C4 and C5 from the bottom-up pyramid structures DenseBlock(1), DenseBlock(2), DenseBlock(3) and DenseBlock(4), respectively. C2-C5 are the preliminary disease feature maps extracted by DenseNet-169, which contain the disease features needed in the subsequent structure of the application.

[0080] In the feature fusion part, the features extracted by the feature extraction part are subjected to multiple optimization, which is specifically implemented as follows: the top feature map P5 is generated by 1*1 convolution feature dimension reduction and CBAM adaptive feature refinement of the horizontal input C4. Then, when P5 is fused with the horizontal input C4, P5 needs to be up-sampled by a 2 times nearest neighbor up-sampling method. In addition, before generating a new feature map P4, CBAM is applied to the feature map obtained in the previous step to re-allocate the feature map and feature weight. Based on the same strategy, P3 is generated using C3 and P4, and P2 is generated using C2 and P3. Finally, in order to eliminate the aliasing effect of up-sampling, the feature maps P2-P5 all need to be subjected to 3*3 convolution to generate a group of feature maps named Predict. Predict6 is obtained by maximum pooling down-sampling on Predict5.

[0081] Step 203, the feature maps Predict2-6 are fused into one feature map as the input of the subsequent region generation network, and the feature maps Predict2-5 are fused into another feature map as the input of the subsequent region of interest network.

[0082] Through the DenseNet-FPN-CBAM model, both the strong abstract semantic features of the upper layer and the bottom layer features of the horizontal input can be considered, and useless features generated by the complex background are also inhibited. The feature extraction method of the application performs multiple optimization processing on the disease features, so that the extracted wheat disease and pest target information is more prominent.

[0083] Step 3, the feature maps generated in step 2 are sent into a region candidate network (Region Proposal Network, RPN) and a prediction network (Predict Network) to classify and identify the wheat disease and pest image.

[0084] The RPN generates a candidate frame of the wheat disease region. The prediction network includes three branches, a classification branch, a prediction frame branch and a mask branch. The classification branch contains the category information of the disease contained in the image, the prediction frame branch contains the coordinate information of the disease position, and the mask branch contains the detailed edge information of the disease position. The information in the three branches is combined to obtain the final wheat disease and pest prediction result image.

[0085] Step 4, calculate the loss function

[0086] Step 401, in order to guide the prediction direction of the wheat disease and pest mask and speed up the identification speed of the model, the application uses an edge protocol header to perform Sobel filtering on the predicted mask and the labeled mask, obtains the edge features of the predicted mask and the labeled mask, and calculates an additional loss function L edge which is added to the basic loss function. L edge The calculation formula of L wherein p represents mean square error, M p represents the p-th power of the generalized power mean, and in the present application, p is 2, and y represents the labeled mask edge information, represents the predicted mask edge information. The present application uses Sobel filtering to detect mask edge features, and calculates edge loss to guide the generation direction of the predicted mask, thereby avoiding redundant calculation and improving the identification speed of plant diseases and insect pests.

[0087] Step 402, calculate the loss function of the basic Mask-RCNN. The role of the loss function is to calculate the difference value between the predicted value and the true value, that is, the loss value. After obtaining the loss value, the model updates the parameters through back propagation to reduce the loss between the true value and the predicted value. The basic Mask-RCNN loss function includes three parts, the classification loss L cls , the predicted frame loss L box and the mask loss L mask . The calculation formula is L Mask-RCNN =L cls +L box +L mask .

[0088] Step 403, calculate the final loss L our of the method proposed in the present application, and feed the loss value back to the neural network to optimize the weight parameters, L our =L cls +L box +L mask +L edge .

[0089] Step 5, perform multi-round iterative training on the model proposed in the present application until the model performance tends to be stable and the loss function converges, and save the weight parameters.

[0090] At this point, the model construction of the field wheat plant disease and insect pest intelligent identification method based on multiple feature optimization proposed in the present application is completed.

[0091] Based on the same inventive concept, the present application also provides a field wheat plant disease and insect pest identification device, which comprises a data acquisition module, a plant disease and insect pest identification module, a candidate frame generation module and a result prediction module.

[0092] Specifically, the data acquisition module is used for the plant disease and insect pest image of wheat to construct a data set; the plant disease and insect pest identification module is used for inputting the plant disease and insect pest image in the data set into a plant disease and insect pest identification model to output a feature map; the candidate frame generation module is used for sending the feature map into a region candidate network RPN to generate a candidate frame of the wheat disease area; and the result prediction module is used for classifying and identifying the feature map in the candidate frame through a prediction network Predict Network to obtain a wheat plant disease and insect pest prediction result image.

[0093] Further, the pest and disease identification module comprises a feature extraction model and a feature generation model. The feature extraction model is configured to use a neural network DenseNet-169 as a backbone feature extraction network of a convolutional neural network Mask-RCNN, to perform preliminary feature extraction on the pest and disease images in the data set by the DenseNet-169, and to obtain disease features. The feature generation model is configured to add a feature pyramid network FPN and a convolutional attention module CBAM to the backbone feature extraction network DenseNet-169, to perform multiple optimization and fusion on the disease features, and to generate feature maps.

[0094] The above-mentioned various modules in the field wheat pest and disease identification device can be realized by software, hardware and combinations thereof in whole or in part. The above-mentioned various modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to call and execute the operations corresponding to the above-mentioned various modules by the processor.

[0095] Meanwhile, the present application also provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the above-mentioned method.

[0096] Collecting pest and disease images of wheat to construct a data set.

[0097] Inputting the pest and disease images in the data set into a pest and disease identification model to output feature maps.

[0098] Sending the feature maps into a region proposal network RPN to generate candidate boxes of wheat disease regions.

[0099] Classifying and identifying the feature maps in the candidate boxes by a prediction network Predict Network to obtain a wheat pest and disease prediction result map.

[0100] Inputting the pest and disease images in the data set into a pest and disease identification model to output feature maps, specifically comprising:

[0101] Using a neural network DenseNet-169 as a backbone feature extraction network of a convolutional neural network Mask-RCNN, and performing preliminary feature extraction on the pest and disease images in the data set by the neural network DenseNet-169 to obtain disease features.

[0102] Adding a feature pyramid network FPN and a convolutional attention module CBAM to the backbone feature extraction network DenseNet-169 to perform multiple optimization and fusion on the disease features to generate feature maps.

[0103] After testing, the mean average precision (mAP) of the method proposed in the application for identifying field wheat diseases and insect pests can reach 96.02%, and the identification speed of a single image can reach 0.08 s / pic (second per picture), and the effect diagram is as shown in Figure 2 The results show that the method proposed in the application has the advantages of high precision and high efficiency, and has excellent comprehensive performance, can meet the needs of real-time monitoring of field wheat diseases and insect pests, can provide technical support for intelligent transformation and upgrading of the wheat industry, and is worth popularizing. However, the method proposed in the application is only for the identification and detection of wheat diseases and insect pests, and does not have the function of grading and predicting the severity of diseases and insect pests. Subsequently, the data set can be continuously expanded, and the method proposed in the application can be improved and perfected for multiple crops and multiple severity detection functions.

[0104] The above-described embodiments are only preferred specific embodiments of the application, and the protection scope of the application is not limited thereto, and any simple change or equivalent replacement of the technical solutions within the technical range disclosed in the application can be obtained by those skilled in the art, and all belong to the protection scope of the application.

Claims

1. A field wheat disease and pest identification method, characterized in that, The method comprises the following steps: Collecting disease and pest images of wheat to build a data set; Inputting the disease and pest images in the data set into a disease and pest recognition model to output feature maps; Sending the feature maps into a region proposal network (RPN) to generate candidate boxes of wheat disease regions; Classifying and identifying the feature maps in the candidate boxes through a prediction network to obtain a wheat disease and pest prediction result map; The step of inputting the disease and pest images in the data set into the disease and pest recognition model to output feature maps comprises the following steps: The neural network DenseNet-169 is taken as a backbone feature extraction network of the convolutional neural network Mask-RCNN, preliminary feature extraction is performed on the plant disease and pest images in the data set by the neural network DenseNet-169, and disease features are obtained; specifically, the plant disease and pest images in the data set are input into the neural network DenseNet-169; the neural network DenseNet-169 extracts feature maps C2 , C3 , C4 , C5 from the bottom-up pyramid structures DenseBlock(1), DenseBlock(2), DenseBlock(3) and DenseBlock(4) respectively. Adding a feature pyramid network (FPN) and a convolutional attention module (CBAM) into a DenseNet-169 backbone feature extraction network to perform multiple optimization and fusion on disease features to generate feature maps; The step of performing multiple optimization and fusion on disease features to generate feature maps comprises the following steps: top-level feature map P5 is generated by 1x1 convolution feature dimension reduction and CBAM adaptive feature refinement on the horizontal input feature map C4 ​ When the top-level feature map P5 is fused with the horizontal input feature map C4 , the top-level feature map P5 is up-sampled by a 2x nearest neighbor up-sampling method. Before generating a new feature map P 4CBAM is applied to the feature map obtained in the previous step to redistribute the feature map and feature weights; Based on the same strategy, using feature maps C3 and feature maps P4 Generating feature maps P3 , using feature maps C2 and feature maps P3 Generating feature maps P2 ; feature maps P2 - P5 All need to be 3 x 3 convolved to generate a set of feature maps named Predict ; Perdict6 are obtained by max-pooling down-sampling on Predict5 ; The feature maps Predict2- The feature maps Predict6 Fused into one feature map as input for the subsequent region generation network. The feature map Predict2 - feature map Predict 5 fused into another feature map as input for the subsequent region of interest network.

2. The field wheat disease and pest recognition method according to claim 1, characterized in that, The disease and pest images are images collected at different time periods for different varieties of wheat.

3. The field wheat disease and pest recognition method according to claim 1, characterized in that, The prediction network comprises a classification branch, a prediction box branch and a mask branch; The classification branch contains category information of diseases contained in the images, the prediction box branch contains coordinate information of disease positions, and the mask branch contains detailed edge information of disease positions.

4. The field wheat disease and pest recognition method according to claim 1, characterized in that, Before performing preliminary feature extraction on the disease and pest images in the data set through the DenseNet-169, the method further comprises expanding the data set by using a data augmentation technique, specifically by using four methods of mirror conversion, color enhancement, random rotation and brightness enhancement to expand each original disease and pest image in the data set.

5. The field wheat pest and disease recognition method according to claim 4, characterized in that, After data augmentation on the data set, the method further comprises annotating the image data after data augmentation in JSON format by using an open-source software LabelMe, dividing the annotated image data into a training set, a test set and a validation set, training the disease and pest recognition model through the training set, testing the disease and pest recognition model through the test set, and verifying the disease and pest recognition model through the validation set.

6. The field wheat pest and disease recognition method of claim 5, wherein, The edge protocol header is used to perform Sobel filtering on the prediction mask and the labeled mask to obtain edge features of the prediction mask and the labeled mask, and an additional loss function is calculated according to the edge features is added to the basic loss function, The calculation formula is where p represents the mean square error, M p represents the p-th power of the generalized power mean, y represents the labeled mask edge information, and ŷ represents the predicted mask edge information; The loss function of the pest and disease identification model The calculation formula is: wherein, , , are the classification loss, the bounding box loss and the mask loss of the Mask-RCNN network, respectively. through a loss function feedback the loss value to the pest and disease identification model to optimize the weight parameter; The pest recognition model is trained for multiple rounds of iteration until the model performance tends to be stable and the loss function converges. The weight parameters are saved when the model performance tends to be stable and the loss function 7. A field wheat disease and pest identification device characterized by, The method comprises the following steps: A data collection module is configured to collect disease and pest images of wheat to build a data set; A disease and pest recognition module is configured to input the disease and pest images in the data set into a disease and pest recognition model to output feature maps; A candidate box generation module is configured to send the feature maps into a region proposal network (RPN) to generate candidate boxes of wheat disease regions; A result prediction module is configured to classify and identify the feature maps in the candidate boxes through a prediction network to obtain a wheat disease and pest prediction result map; The disease and pest recognition module comprises: The feature extraction model is used for taking the neural network DenseNet-169 as a backbone feature extraction network of the convolutional neural network Mask-RCNN, and the disease characteristics are obtained by performing preliminary feature extraction on the disease and pest image in the data set through the DenseNet-169; specifically, the disease and pest image in the data set is input into the neural network DenseNet-169; the neural network DenseNet-169 extracts feature maps C2 , feature maps C3 , feature maps C4 and feature maps C5 from the bottom-up pyramid structures DenseBlock(1), DenseBlock(2), DenseBlock(3) and DenseBlock(4) respectively. A feature generation model is configured to add a feature pyramid network (FPN) and a convolutional attention module (CBAM) into a DenseNet-169 backbone feature extraction network to perform multiple optimization and fusion on disease features to generate feature maps; The step of performing multiple optimization and fusion on disease features to generate feature maps comprises the following steps: top-level feature map P5 is generated by 1x1 convolution feature dimension reduction and CBAM adaptive feature refinement on the horizontal input feature map C4 is generated by 1x1 convolution feature dimension reduction and CBAM adaptive feature refinement on the horizontal input feature map When the top-level feature map P5 is fused with the horizontal input feature map C4 , the top-level feature map P5 is up-sampled by a 2x nearest neighbor up-sampling method. In generating a new feature map P 4 Before, CBAM is applied to the feature map obtained in the last step to reassign feature maps and feature weights; Based on the same strategy, using feature maps C3 and feature maps P4 generate feature maps P3 , using feature maps C2 and feature maps P3 generate feature maps P2 ; feature maps P2 - P5 All need to be 3 x 3 convolved to generate a set of feature maps named Predict ; Perdict6 are obtained by max-pooling down-sampling on Predict5 ; The feature maps Predict2- The feature maps Predict6 Fused into one feature map as input for the subsequent region generation network. The feature map Predict2 - feature map Predict 5 fused into another feature map as input for the subsequent region of interest network.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, which is executed by a processor, implements the steps of the method according to any one of claims 1 to 6.

Images