CN111369540A - Plant leaf disease identification method based on mask convolutional neural network - Google Patents

Abstract

The invention discloses a plant leaf disease identification method based on a mask convolution neural network, which mainly solves the problem of low accuracy rate of identifying plant leaf diseases in the prior art. The scheme is as follows: enhancing and expanding the original data set to obtain a training set and a test set; performing semantic segmentation on the training set and the test set to obtain a corresponding mask set; adding a disease characteristic screening module between the full convolution layer and the mask branch of the model, inputting the training set and the mask set into a network for training, and obtaining the results of target classification and target detection; taking a characteristic graph belonging to disease leaves in a target classification result as the input of a mask branch, and obtaining a trained model after multiple iterations; and inputting the test set into the model, performing target classification and target detection on the leaves, and segmenting the leaves belonging to the disease category. The method improves the accuracy of leaf disease identification on the basis of the traditional method, and can be used for identifying and segmenting plant disease leaves in agricultural planting.

Description

Plant leaf disease identification method based on mask convolutional neural network

Technical Field

The invention belongs to the technical field of image processing, and particularly relates to a plant leaf disease identification method which can be used for segmenting and identifying plant disease leaves in agricultural planting.

Background

In modern wisdom agriculture, crop diseases are a great threat to grain safety, and plant diseases cause serious damage to crops by significantly reducing yield. Early blight is a typical disease that can severely reduce yield. Similarly, in humid climates, late blight is another very devastating disease that affects the leaves, stems and fruits of plants. Protection of plants from disease is critical to ensure the quality and quantity of crops. Crop protection should begin with the early detection of disease so that appropriate treatments can be selected at the right time to prevent disease transmission. The method for identifying the disease types of the greenhouse plant diseases mainly comprises various diseases such as bacterial spots, early blight, late blight, leaf mold, black spot and the like. However, in practice, the number of diseases is large, and the diseases are similar in performance on leaves, so that it is difficult to accurately determine the types of the diseases.

At present, the disease research on plant leaves mainly comprises detecting and classifying the plant leaves by using an image processing or deep learning method. In the Plant disease control, a foreign Plant Village group provides a method for detecting and classifying diseases through Plant leaves by utilizing deep learning in 2016, mainly classifies specific partial Plant diseases through a smart phone under a simple background, and mainly identifies colors, gray scales and segmented pictures through the deep learning method under data sets with different proportions and different networks, but the method can only classify the Plant disease leaves and cannot segment the Plant disease leaves and the positions of the diseases.

The same year Mads Dyrmann1 et al proposed the use of convolutional neural networks to classify plant species in color images. The network is built from scratch, trained and tested. These images are from six different local datasets, with data acquisition at different growth stages, in terms of illumination, resolution and soil type, but with low accuracy and without detection and classification of plant diseases.

Liuna et al in 2018 use an image processing technology and an artificial neural network technology to realize disease detection and disease degree classification of cucumber leaves, and mainly perform experimental research on cucumber downy mildew, powdery mildew and virus diseases with high morbidity and serious harm, but because the disease types of cucumbers are identified less, the number of training samples is less, and the identified disease types are less, the test accuracy is lower and overfitting is easily generated.

The framework of Mask convolutional neural network Mask R-CNN is provided by Heterokamm in 2017, and is characterized in that semantic segmentation is carried out on two branches of FasterR-CNN, namely, on classification and coordinate regression, one branch is added, image features are extracted mainly through a residual error network ResNet101/50 or a pyramid network FPN serving as a main network, the foreground and the background of a target region are obtained through a region recommendation network RPN, classification and example segmentation results are obtained through full convolutional layers on the obtained target region and the image features, and then the semantic segmentation results are obtained through convolutional network identification. The network is mainly used for target detection of a COCO data set, and is not used for the field of plant disease identification at present.

In summary, the current research on plant diseases mainly includes classification and identification of single plant disease types similar to cucumber, wherein the included samples include fewer disease types and samples, and the identification accuracy is low. The existing method for classifying the plant disease leaves through deep learning cannot separate the plant disease leaves and the positions of diseases, and the recognition rate is low.

Disclosure of Invention

The invention aims to provide a plant leaf disease identification method based on mask convolution neural network MaskR-CNN aiming at the defects in the prior art, so as to segment plant disease leaves and the positions of diseases, and improve the identification accuracy and efficiency.

In order to achieve the purpose, the technical scheme of the invention comprises the following steps:

(1) sequentially enhancing, expanding and semantically segmenting an original data set to obtain a training and testing image set and a mask set;

(2) optimizing a Mask R-CNN network: a disease characteristic screening module is added between the full convolution layer and the Mask branch after the Mask R-CNN network ROI Align;

(3) training the optimized Mask R-CNN network:

(3a) setting main network parameters:

selecting a main network from the ResNet50 and ResNet101 residual error networks;

setting the size of the epoch and the step number of each epoch training;

setting a receiving threshold T of a diseased leaf in a diseased feature screening module₀The other parameters are default values of Mask R-CNN;

(3b) according to the known classification error L_clsDetecting error L_boxAnd a segmentation error L_maskAnd determining the optimized MaskR-CNN network loss function as follows: l is_loss＝L_cls+L_box+L_mask；

(3c) Inputting the training image set and the training Mask set into an optimized Mask R-CNN network for training to obtain a trained model;

(4) and inputting the test image into the trained model for testing.

Compared with the prior art, the invention has the following advantages:

firstly, compared with the GoogleNet and VGG methods, the method disclosed by the invention has the advantage that the identification accuracy of the damaged leaves is improved when the plant leaf diseases are identified.

Secondly, the Mask image of the target area is generated through the Mask branch of the Mask R-CNN network model, so that the plant disease leaves and the disease positions thereof can be accurately extracted.

Thirdly, because the disease characteristic screening module is added in the network structure of the Mask R-CNN, the invention only aims at unhealthy leaves to carry out the training and testing of Mask branches, reduces the burden of the Mask branches and improves the identification efficiency of the network on the premise of ensuring that the accuracy of the Mask R-CNN network is not changed.

Fourthly, in the invention, because the samples in the data set are increased by adopting image transformation and the data set is enhanced by adopting an adaptive contrast enhancement algorithm, the blurred image in the data set is improved.

Drawings

FIG. 1 is a general flow diagram of an implementation of the present invention;

FIG. 2 is a schematic diagram of the overall structure of an optimized Mask R-CNN network in the present invention;

FIG. 3 is a block diagram of a training sub-process for an optimized Mask R-CNN network in the present invention;

FIG. 4 is a training image and a training binary mask image acquired in the present invention;

FIG. 5 is a test image of healthy and diseased leaves taken in the present invention;

FIG. 6 is a result image of healthy leaves identified by simulation of the present invention;

FIG. 7 is a result image of diseased leaves identified by simulation in accordance with the present invention.

Detailed Description

The following detailed description of the embodiments and effects of the present invention will be made with reference to the accompanying drawings:

the application environment of the embodiment is an agricultural planting scene, and aims to detect and identify vegetation with diseases in agricultural planting and provide disease information for planting personnel more accurately.

Referring to fig. 1, the implementation steps of this example are as follows:

step 1, image enhanced data set D₁。

(1.1) plant disease leaf Dataset D was downloaded from public project PlantVillage-Dataset₀Using adaptive contrast enhancement algorithm to pair D₀Performing image enhancement to improve the blurred image in the database to obtain an image-enhanced database D₁：

(1.1a) acquiring the low frequency part m of the image x (i, j)_x(i, j) and a high frequency part h_x(i, j), obtaining the low frequency part of the image by means filtering:

h_x(i,j)＝x(i,j)-m_x(i,j)

wherein, (2n +1)²(ii) represents the window size with (i, j) as the image center point coordinate;

(1.1b) multiplying the high frequency part in the image by the gain value G (I, j) of the part to obtain an amplified high frequency part I (I, j):

I(i,j)＝G(i,j)h_x(i,j)

wherein G (I, j) may take a constant C greater than 1 to obtain an amplified high-frequency portion I_c(i, j) is:

I_c(i,j)＝Ch_x(i,j)

in this example the gain value G (i, j) is taken to be equal to the local mean square error σ_x(i, j) is inversely proportional to the change value

Wherein D is a constant, and the local mean square error of the image is:

obtaining an amplified high-frequency part I_σ(i, j) is:

(1.1c) recombining the high frequency part and the low frequency part to obtain an enhanced image f (i, j):

f(i,j)＝m_x(i,j)+I_σ(i,j)；

step 2, obtaining a training image set D₃And testing the image set D₄And a training mask set D₅And test mask set D₆。

(2.1) enhancing the image data set D by using a semantic segmentation labeling tool labelme₁Respectively plotting the image targets to generate the mask of the targets, and obtaining a mask set D containing mask information and label information₂The resulting training binary mask image is shown in fig. 4 b.

(2.2) Using image transformation on the enhanced dataset D₁Sum mask set D₂Sequentially translating, rotating and overturning to increase the sample data volume and obtain a data set D after sample expansion₃Sum mask set D₄；

(2.2) expanding the data set D₃Sum mask set D₄Dividing the training image into training image sets D according to the proportion of 8:2₅And testing the image set D₆And a training mask set D₇And test mask set D₈The obtained training set image is shown in fig. 4a, and the test set image is shown in fig. 5a and 5 b;

and step 3, continuously optimizing the Mask R-CNN network structure.

The existing Mask R-CNN network comprises a main network, a regional recommendation generation network RPN, a full convolution layer and a Mask branch, namely the full convolution layer and a full connection layer, in the embodiment, a disease characteristic screening module is added between the full convolution layer and the Mask branch to obtain an optimized Mask R-CNN network structure, which is shown in figure 2.

Referring to fig. 2, the Mask R-CNN network structure after optimization in this example is: the main network → the area recommendation generation network RPN → the full convolution layer → the disease feature screening module → the full convolution layer → the full connection layer.

The disease characteristic screening moduleThe block is to judge the confidence coefficient T of the fault blade output by the full convolution layer₁Receiving threshold T of damaged leaf given by network initialization parameter₀The characteristic maps of the diseased leaves in batch processing are screened out, and screening results are input to mask branches, namely, full convolution layers and full connection for processing.

Step 4, training the optimized Mask R-CNN network to obtain a trained network model:

referring to fig. 3, the specific implementation of this step is as follows:

(4.1) setting main network parameters:

selecting a main network from the ResNet50 and ResNet101 residual error networks, wherein ResNet101 is selected as the main network in the example;

setting the number of labels to be 11 according to the category of the images in the database, wherein the number of labels comprises 1 background label and 10 image labels;

setting the iteration times epoch of all samples to be 100, the iteration times of each epoch to be 100, the learning rate to be 0.001 and the weight attenuation coefficient to be 0.0001; setting the number of GPUs to be 1, the number of images processed by each GPU to be 2, and setting the receiving threshold value T of the damaged leaf₀The other parameters are default values of Mask R-CNN;

(4.2) based on the known classification error L_clsDetecting error L_boxAnd a segmentation error L_maskAnd determining the optimized MaskR-CNN network loss function as follows: l is_loss＝L_cls+L_box+L_mask；

(4.3) training the optimized Mask R-CNN network:

(4.3a) initializing the network parameters in (3a) and training the image set D₅And training mask set D₇Inputting the data into an optimized Mask R-CNN network;

(4.3b) extracting the characteristic map F of the training image through the training residual error network₀；

(4.3c) mapping the feature pattern F₀Inputting the information into a region recommendation generation network RPN to obtain the foreground F of the target region₁And background F₂；

(4.3d) reactingUsing ROI Align method to map the foreground F of the target area₁Mapping to a feature map F₀Generating a feature map F with a fixed size₃：

First, the foreground F of the target area is calculated₁The characteristic layer of the system is as follows:

wherein ,w₀ and h₀Respectively representing the width and height of the target region, k₀A value of 4;

second, in the target area foreground F₁After finding out a corresponding characteristic layer k, obtaining a step length s corresponding to the characteristic layer;

then, the foreground F of the target area is calculated₁Breadth mapping to feature map

And height

And obtaining a target area Z on the characteristic diagram according to the two parameters:

Z＝w₁*h₁

next, the target region Z on the feature map is divided into n²Obtaining divided target areas Z_i，

Z_i＝w₂*h₂,i＝1,2,…n²

wherein ,w₂ and h₂Represents Z_iWidth and height, respectively

Then, each target region Z_iDividing the image into four parts, obtaining pixel values of four points by taking the position of a central point of each part, and taking the maximum value of the four pixel values as each target area Z_iTo obtain n in the target zone Z²Pixel values which form a characteristic map of n × n size;

(4.3e) feature map F₃By full rollingStacking to obtain target classification result and target detection result, and calculating classification error L_clsAnd the detection error L_box：

Probability p corresponding to target classification result u_uObtaining the classification error: l is_cls＝-logp_u；

Let t_i ^u＝{t_x,t_y,t_w,t_hV 4 parameterized coordinates of the target detection result, v_i＝{v_x,v_y,v_w,v_hCalculating the detection error L by the following formula as a target translation scaling parameter_box：

Wherein smooth_L1Is a smoothed norm loss function expressed as:

(4.3f) judging whether the classification result belongs to the diseased leaf:

if the leaf is not the diseased leaf, continuing to judge the target classification of the next leaf;

if the leaf is a diseased leaf, the confidence coefficient T of the classification result is determined₁And a diseased leaf reception threshold T₀Comparing;

when T is₁＞T₀Then, in the feature diagram F₃The confidence coefficient of the selected one is T₁Characteristic diagram F of₄；

When T is₁＜＝T₀If so, continuing to judge the target classification of the next blade;

(4.3g) feature map F selected in (4.3d)₄And training mask set D₇Inputting the binary mask into a mask branch for training to obtain a binary mask of a target area, namely a segmentation result of a disease leaf and a disease position of the disease leaf:

first, the feature map F is transformed by deconvolution₃Amplifying to obtain binary mask regions M of all categories_k；

Then, traverse all the categories of binary mask regions M_kFor each class binary mask region M_iApplying Sigmoid activation functions

After classification operation is carried out, a classification probability vector H is obtained, and the maximum probability y in the H is a binary mask of a target area corresponding to target classification;

(4.3h) calculating the segmentation error L of the binary mask of the target region obtained in (4.3e)_mask；

Where y is the predicted probability of the binary mask of the target region,

a real label of the binary mask of the target area;

(4.3i) calculating loss value L of network_loss＝L_cls+L_box+L_maskUpdating the network weight by back propagation of the loss value after each iteration;

(4.3g) judging whether the iteration times of all samples are more than the set 100 times:

if the iteration times of all samples are larger than 100, stopping network training to obtain a trained network model;

if the iteration number of all samples is less than or equal to 100, repeating the steps from (4.3c) to (4.3g) until the iteration number of all samples is more than 100.

And 5, obtaining the identification result of the plant leaf diseases.

(5.1) collecting the test image D obtained in (1.2)₆Healthy leaves in (1) such as FIG. 5a and diseased leaves in (5 b) are input into the trained network model, feature vectors are extracted, and feature vectors A of each real target classification in n real leaf classes are calculated_iCosine similarity cos with feature vector B of prediction target classification_i(θ), resulting in a probability vector P for the classification:

P＝{cos₁(θ)，cos₂(θ)，…，cos_i(θ)，…cos_n(θ)}，i＝1,2…,n

wherein the cosine similarity cos_iThe formula for (θ) is as follows:

wherein ,||A_iThe | | represents the two-norm of the feature vector of each real target classification, and the | B | represents the two-norm of the feature vector of the prediction target classification;

(5.3) selecting the maximum probability value q in the probability vector P, and taking the category corresponding to q as a target classification result;

(5.2) judging whether the category corresponding to the q belongs to the diseased leaf:

if the class corresponding to q is a healthy leaf, directly outputting a target detection result and a target classification result;

if the category corresponding to q belongs to the damaged leaf, the q and the damaged leaf receiving threshold T are used₀And (3) comparison:

if q > T₀Outputting target detection, target classification and segmentation results of disease positions of the targets;

if q < ═ T₀And directly outputting the target detection result and the target classification result.

The effect of the invention can be further illustrated by the following simulation experiment:

experimental conditions

The software platform for the experiment training is as follows: *** coliab; the hardware platform is as follows: tesla P4 GPU; the development environments are keras and Tensorflow;

the software platform for the experimental test is as follows: windows 10; the hardware platform is as follows: a CPU;

the test image is selected from (256, 3) size blade image shown in FIG. 5, and the test set is selected from the test set D obtained in step 2₃。

Second, the experimental contents

Experiment 1. A Mask R-CNN network model and the method of the invention are used for respectively carrying out comparison experiments on single test images.

Fig. 5 is a test image, in which fig. 5a is a healthy leaf and fig. 5b is a diseased leaf. The detection result of the Mask R-CNN network model is shown in FIG. 6, in which FIG. 6a is the detection result of healthy leaves, and FIG. 6b is the detection result of diseased leaves. The detection results of the method of the present invention are shown in fig. 7, wherein fig. 7a is the detection results of healthy leaves, and fig. 7b is the detection results of diseased leaves. From the detection results of the two methods, the method reduces redundant operation of dividing the diseased region of the blade compared with a MaskR-CNN network when detecting the healthy blade.

Experiment 2. comparative experiments were carried out with the method of the present invention and GoogleNet and VGG networks, respectively.

The method and the existing GoogleNet and VGG network models are used for carrying out the pair of test image sets D obtained in the step 1₆1000 tests are carried out, the obtained identification accuracy rate results are shown in table 1,

TABLE 1 identification accuracy results

Method of producing a composite material	Average rate of accuracy	Time/s
			VGG	0.8846	3.17
GoogleNet	0.9040	3.32
			Optimized Mask R-CNN	0.9257	3.59

From table 1, it can be seen that the method of the present invention improves the recognition accuracy compared to the conventional GoogleNet and VGG network models.

The foregoing description is only a specific example of the present invention and is not intended to limit the invention, so that it will be apparent to those skilled in the art that modifications and variations in form and detail may be made without departing from the principles and structures of the invention, but such modifications and variations are within the scope of the invention as defined by the appended claims.

Claims (10)

1. A plant leaf disease identification method based on a mask convolution neural network is characterized by comprising the following steps:

(1) sequentially enhancing, expanding and semantically segmenting an original data set to obtain a training and testing image set and a mask set;

(2) optimizing a Mask R-CNN network: a disease characteristic screening module is added between the full convolution layer and the Mask branch after the Mask R-CNN network ROI Align;

(3) training the optimized Mask R-CNN network:

(3a) setting main network parameters:

selecting a main network from the ResNet50 and ResNet101 residual error networks;

setting the size of the epoch and the step number of each epoch training;

setting a receiving threshold T of a diseased leaf in a diseased feature screening module₀The other parameters are default values of Mask R-CNN;

(3b) according to the known classification error L_clsDetecting error L_boxAnd a segmentation error L_maskAnd determining the optimized Mask R-CNN network loss function as follows: l is_loss＝L_cls+L_box+L_mask；

(3c) Inputting the training image set and the training Mask set into an optimized Mask R-CNN network for training to obtain a trained model;

(4) and inputting the test image into the trained model for testing.

2. The method of claim 1, wherein (1) the original data set is sequentially enhanced, augmented and semantically segmented to obtain a trained and tested image set and a mask set, as follows:

(1a) the Plant disease leaf data set D is downloaded from the public project Plant Village-Dataset₀Using adaptive contrast enhancement algorithm to pair D₀Performing image enhancement to improve the blurred image in the database to obtain an image-enhanced data set D₁；

(1b) Tagging tool for semantic segmentation on enhanced data set D₁Respectively plotting the image targets to generate the mask of the targets, and obtaining a mask set D containing mask information and label information₂；

(1c) Pair of enhanced data sets D with image transformations₁Sum mask set D₂Sequentially translating, rotating and overturning to increase the sample data volume and obtain a data set D after sample expansion₃Sum mask set D₄；

(1d) Scaling the augmented data set D by 8:2₃Sum mask set D₄Divided into training image sets D₅And testing the image set D₄And a training mask set D₇And test mask set D₈。

3. The method of claim 2, wherein the image enhancement in (1a) is performed by using an adaptive contrast enhancement algorithm, and the following is performed:

(1a1) dividing the image into high-frequency parts h_x(i, j) and a low frequency part m_x(i, j), wherein (i, j) refers to pixel points of an image;

(1a2) multiplying the high frequency part in the image by the gain value G (I, j) of the part to obtain an amplified high frequency part I (I, j):

I(i,j)＝G(i,j)h_x(i,j)

(1a3) recombining the high frequency part and the low frequency part to obtain an enhanced image f (i, j):

f(i,j)＝m_x(i,j)+I(i,j)。

4. the method according to claim 1, wherein the optimized Mask R-CNN network obtained in (2) has the following structure:

residual error network ResNet101/50 or pyramid network FPN → area recommendation generation network RPN → full convolution layer → disease feature screening module → mask branch, namely full convolution layer and full connection layer;

the disease characteristic screening module is used for judging the confidence coefficient T of the disease blade output by the full convolution layer₁With a given diseased leaf reception threshold T₀Screening out the characteristic maps of the diseased leaves in batch processing, and inputting the screening result into a mask branch for processing.

5. The method according to claim 1, wherein the optimized Mask R-CNN network is trained in (3c) as follows:

(3c1) initializing the network parameters in (3a) and assembling the training image set D₅And training mask set D₇Inputting the data into the optimized MaskR-CNN network;

(3c2) extraction of D by training residual error network₅To obtain a feature map F₀；

(3c3) The characteristic map F₀Inputting the information into a region recommendation generation network RPN to obtain the foreground F of the target region₁And background F₂；

(3c4) Foreground F of target area by using ROIAlign method₁Mapping to a feature map F₀Generating a feature map F with a fixed size₃；

(3c5) Will feature chart F₃Obtaining a target classification result and a target detection result through a full convolution layerAnd calculating a classification error L_clsAnd the detection error L_box；

(3c6) If the classification result is a diseased leaf, the confidence coefficient T of the classification result is determined₁And a diseased leaf reception threshold T₀By comparison, when T is₁＞T₀Then, a disease characteristic screening module is utilized to screen a characteristic diagram F₃The confidence coefficient of the selected one is T₁Characteristic diagram F of₄；

(3c7) The feature map F selected in (3c6)₄And training mask set D₇Inputting the binary mask into a mask branch for training to obtain a binary mask of a target area, namely a segmentation result of a disease leaf and a disease position of the disease leaf, and calculating a segmentation error L_mask；

(3c8) Calculating a loss value L of a network_loss＝L_cls+L_box+L_maskAnd updating the network weight by utilizing the back propagation of the loss value after each iteration, and stopping the network training when the training epoch is greater than the initialized epoch to obtain the trained network model.

6. The method of claim 5, wherein the ROI Align method is used to map the foreground F of the target region in (3c4)₁Mapping to a feature map F₀The corresponding positions of (a) are implemented as follows:

(3c4a) calculating a feature layer to which the target region belongs:

wherein ,w₀ and h₀Respectively representing the width and height of the target region, k₀A value of 4;

(3c4b) after finding the corresponding characteristic layer k in the target area, obtaining the step length s corresponding to the characteristic layer;

(3c4c) calculating the width w of the target region mapping to the feature map₁And a height h₁And obtaining a target area Z on the characteristic diagram:

Z＝w₁*h₁

wherein ,

(3c4d) dividing the target zone Z on the feature map into n²Obtaining divided target areas Z_i，

Z_i＝w₂*h₂,i＝1,2,…n²

wherein ,w₂ and h₂Represents Z_iWidth and height, respectively

(3c4e) Each target zone Z_iDividing the image into four parts, obtaining pixel values of four points by taking the position of a central point of each part, and taking the maximum value of the four pixel values as each target area Z_iTo finally obtain n in the target zone Z²And (4) forming a characteristic map with the size of n × n by using the pixel values.

7. The method of claim 5, wherein the classification error L is calculated in (3c5)_clsAnd the detection error L_locThe implementation is as follows:

(3c5a) probability p corresponding to the target classification result u_uObtain a classification error L_cls：

L_cls＝-logp_u

(3c5b) setting t_i ^u＝{t_x,t_y,t_w,t_hV 4 parameterized coordinates of the target detection result, v_i＝{v_x,v_y,v_w,v_hCalculating the detection error L by the following formula as a target translation scaling parameter_box：

wherein

For the smoothed norm loss function, tableShown as follows:

8. the method of claim 5, wherein (3c7) the binary mask is obtained through a mask branch, and the following is implemented:

(3c7a) feature map F is transformed by deconvolution₃Amplifying to obtain binary mask regions M of all categories_k；

(3c7b) traversing all classes of binary mask regions M_kFor each class binary mask region M_iApplying Sigmoid activation functions

After classification operation is carried out, a classification probability vector H is obtained, and the maximum probability y in the H is a binary mask of a target area corresponding to target classification;

(3c7c) calculating the segmentation error L of the mask branch_mask：

wherein ,

the true label of the binary mask of the target area.

9. The method of claim 1, wherein the test image is input into the trained model for testing in (4) as follows:

(4a) collecting the test image D obtained in (1a)₆Inputting the feature vectors into a trained network model, extracting the feature vectors, obtaining classified probability vectors P by calculating the similarity of the feature vectors, selecting the maximum probability value q in P, and taking the category corresponding to q as a target classification result;

(4b) judging whether the category corresponding to q belongs to a diseased leaf or not:

if the class corresponding to q is a healthy leaf, directly outputting a target detection result and a target classification result;

if the category corresponding to q belongs to the damaged leaf, the q and the damaged leaf receiving threshold T are used₀And (3) comparison:

if q > T₀Outputting target detection, target classification and segmentation results of disease positions;

if q < ═ T₀And directly outputting the target detection result and the target classification result.

10. The method of claim 9, wherein the probability vector of the classification obtained by calculating the similarity of the feature vectors in (4a) is obtained by calculating the cosine similarity cos (θ) between the feature vector a of each real target classification and the feature vector B of the predicted target classification, and the formula is as follows:

wherein, | a | | represents the two-norm of the feature vector of the real target classification, and | B | | represents the two-norm of the feature vector of the prediction target classification.

Images