Wheat scab recognition and lesion spot segmentation method based on convolutional neural network

Abstract

The present invention discloses a method for wheat scab identification and lesion segmentation based on a convolutional neural network, comprising the following steps: 1. obtaining images of healthy wheat ears and wheat ears with varying degrees of lesion and preprocessing them; 2. constructing a wheat lesion degree identification network model; 3. offline training the constructed wheat lesion degree identification network model; 4. constructing a wheat lesion segmentation network model; 5. offline training the constructed wheat lesion segmentation network model; and 6. using the constructed model to perform predictions to achieve the purpose of wheat lesion degree identification and lesion segmentation. The present invention uses an improved feature fusion convolutional neural network to process wheat lesion images, which can solve the problems of subtle key features in the image being lost during network transmission and lesion edge blurring, thereby improving the accuracy of wheat lesion degree identification and segmentation.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning and image processing, and specifically is a method for wheat fusarium head blight recognition and disease spot segmentation based on convolutional neural networks. Background Art

[0002] Wheat is the world's leading food crop, second only to rice. However, wheat production currently faces threats from a variety of extreme climate conditions and the resulting diseases, pests, low temperatures, high temperatures, drought, and insufficient natural resources, particularly water shortages. Common wheat diseases in my country include three rusts (stripe rust, leaf rust, and stem rust), powdery mildew, Fusarium head blight (FHB), sheath blight, take-all blight, smut, leaf blight, nematodes, and viral diseases. In recent years, the incidence of wheat head blight has become increasingly severe due to climate change and changes in wheat farming systems and agricultural production technologies. To effectively prevent and control wheat head blight, real-time and effective monitoring of wheat head blight severity is essential.

[0003] At present, due to the limitations of information technology development and the complexity of wheat field information, the degree of wheat fusarium wilt is mainly predicted and evaluated by manual field survey and statistical methods. This has high labor costs, takes a long time, and cannot predict and evaluate the degree of wheat fusarium wilt in a timely and effective manner. Summary of the Invention

[0004] The present invention aims to address the deficiencies of the above-mentioned prior art and proposes a method for wheat fusarium head blight identification and disease spot segmentation based on convolutional neural networks, in order to improve the accuracy of wheat disease spot degree identification and segmentation, thereby realizing the intelligent and automated development of agricultural production, reducing manpower and material resources, and improving production efficiency.

[0005] In order to achieve the above-mentioned object, the present invention adopts the following technical solutions:

[0006] The invention provides a method for identifying wheat scab and segmenting diseased spots based on a convolutional neural network, comprising the following steps:

[0007] Step 1: Obtain healthy wheat ear images and wheat ear images with different diseased spots and perform preprocessing to obtain the preprocessed wheat image dataset, which is denoted as X = {x1, x2, ..., x i ,...,x N}, where x i represents the i-th wheat image, i = 1, 2, ..., N, where N is the total number of images in the wheat image dataset X;

[0008] The diseased spot degree label set corresponding to the wheat image dataset X is recorded as Y = {y1, y2, ..., y i ,...,y N}, where y i Represents the i-th wheat image x i The lesion degree label, and y i ∈[0,M], M is the level of the lesion; when y i =0, indicating x i For healthy wheat ears image;

[0009] The lesion segmentation label set corresponding to the wheat image dataset X is denoted as Z = {z1,z2,...,z i ,...,z N}, where z i Represents the i-th wheat image x i The lesion segmentation label;

[0010] Step 2: Construct a wheat spot degree recognition network and use it to identify the i-th wheat image x i Processing is performed to extract the wheat lesion degree characteristic map FD i , and get the i-th wheat image x i The predicted label y' i ;

[0011] Step 3: Construct a multi-category cross entropy loss function L CE and Dice loss function L D The total loss function L composed of

[0012] The wheat image dataset X is input into the disease spot degree recognition network, and the gradient descent method is used to train the disease spot degree recognition network. At the same time, the loss function L is calculated. When L converges, the training is stopped to obtain the optimal wheat disease spot degree recognition network, which is used to identify the disease degree of wheat disease images.

[0013] Step 4: Construct a wheat lesion segmentation network and perform a FD on the wheat lesion degree feature map. i Process and get the i-th wheat image x i Lesion segmentation feature map T i ;

[0014] Step 5: Use formula (6) to construct the cyclic focus loss function L CFL :

[0015]

[0016] In formula (6), ρ is a dynamically adjusted weight coefficient;

[0017] The wheat spot segmentation network is trained using the gradient descent method, and the loss function L is calculated. CFL , until L CFL The training is stopped when convergence occurs, thereby obtaining the optimal wheat lesion segmentation network, which is used to segment wheat lesion images.

[0018] The wheat scab identification and lesion segmentation method based on convolutional neural network of the present invention is also characterized in that the wheat lesion degree identification network includes: a feature extraction sub-network module, a feature downsampling sub-network module, and a feature fusion and classification sub-network module;

[0019] Step 2.1: The feature extraction sub-network module uses the convolution layer to perform a sliding window on the i-th wheat image x i Perform feature extraction to obtain K mapping features in, represents the kth mapping feature;

[0020] The kth mapping feature Input n spatial pooling blocks of different sizes for processing in sequence, obtain the local detail features output by each spatial pooling block, and after splicing the n local detail features, obtain the kth standardized mapping feature F i,k ;

[0021] Step 2.2: The feature downsampling sub-network module consists of s-level strip pooling layers, and sequentially i,k After downsampling, the feature vector output by each strip pooling layer is obtained, and after splicing the s feature vectors, the kth downsampled feature vector FVD is obtained. i,k ;

[0022] Step 2.3: The feature fusion and classification sub-network module includes: a fully connected layer and a Softmax classification module:

[0023] The K downsampled feature vectors {FVD i,k |k=1,2,…,K} is input into the fully connected layer for processing, and the fused wheat lesion degree feature map FD is obtained. i ;

[0024] FD i Input into the Softmax classification module for dimensionality reduction processing to obtain the predicted probability matrix F of the lesion degree i and from F i The degree of lesions corresponding to the maximum probability is selected as the i-th wheat image x i The predicted label y' i .

[0025] The total loss function L in step 3 is obtained by the following steps:

[0026] Step 3.1: Use formula (1) to construct the multi-category cross entropy loss function L CE :

[0027]

[0028] Step 3.2: Use formula (2) to construct the Dice loss function L D :

[0029]

[0030] Step 3.3: Use formula (3) to construct the total loss function L:

[0031] L = α × L CE +β×L D (3)

[0032] In formula (3), α and β are the multi-category cross entropy loss function L CE and Dice loss function L D The weight coefficient of .

[0033] The wheat spot segmentation network includes: an encoder module, a decoder module and a dense skip connection module;

[0034] Step 4.1, the encoder module includes P convolutional contraction blocks CDB1, CDB2, ..., CDB p ,...,CDB P And Q downsampling blocks DB1, DB2, ..., DB q ,...,DB Q , where CDB p represents the pth convolutional contraction block, DB q represents the qth downsampling block;

[0035] When p=1,q=1, for FD i After adjusting the image size and channel, the adjusted feature map is obtained And input the pth convolutional contraction block CDB p In this way, global average pooling is used to The absolute value of the dimensionality reduction process is performed, and the obtained one-dimensional feature vector Then input to the qth downsampling module DB q The size is reduced so that the output size is half of the input size, thus obtaining the qth low-level feature map

[0036] When p=2,3,...,P,q=2,3,...,Q,the p-1th low-level feature map Input the pth convolutional contraction block CDB p The one-dimensional feature vector obtained by processing Then input to the qth downsampling module DB q The size is reduced in the process to obtain the qth low-level feature map Thus we get the Qth downsampling module DB Q The Qth low-level feature map output

[0037] Using formula (4) for the qth low-level feature map Perform soft thresholding to obtain the qth enhanced low-level feature map

[0038]

[0039] In formula (4): σ represents the set threshold;

[0040] Step 4.2: The encoder module includes P convolutional contraction blocks CDB' P+1 ,CDB' P+2 ,...,CDB' P+p ,...,CDB' 2P and Q upsampling blocks UB1, UB2, ..., UB q ,...,UB Q , where CDB' P+p Indicates the P+pth convolutional contraction block; UB q represents the qth upsampling block;

[0041] When p=1, q=1, Input the P+pth convolution shrinkage block CDB' P+p The one-dimensional feature vector obtained is processed Then input to the qth upsampling module UB q In this way, bilinear interpolation is used to increase the output size to twice the input size to obtain the P+pth high-level feature map

[0042] When p=2,3,...,P,q=2,3,...,Q,the P+p-1th high-level feature map Input the P+pth convolution shrinkage block CDB' P+p The one-dimensional feature vector obtained by processing Then input to the qth upsampling module UB q Processed in, get the P+pth high-level feature map

[0043] Using formula (4) to calculate the high-level feature map Perform soft thresholding to obtain enhanced high-level feature maps

[0044] Step 4.3: The dense skip connection module uses formula (5) to obtain the i-th wheat image x i Lesion segmentation feature map T i :

[0045]

[0046] In formula (5), Conv represents the convolution operation; D represents the downsampling operation; Up represents the upsampling operation; [] represents the feature map concatenation operation.

[0047] The electronic device of the present invention includes a memory and a processor, and is characterized in that the memory is used to store a program that supports the processor to execute the method for identifying wheat fusarium wilt and segmenting diseased spots, and the processor is configured to execute the program stored in the memory.

[0048] The present invention provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, which is characterized in that when the computer program is run by a processor, the steps of the method for identifying wheat fusarium head blight and segmenting diseased spots are executed.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. The present invention uses feature fusion to fuse the local features of wheat ear diseases learned by the network with the global features of wheat ear diseases, so that the network can effectively process scenes where spatial (color, texture, etc.) features are severely lost, while also processing scenes where abstract features are not obvious, such as blurred images. This solves the problem of subtle key features in the image being lost during network transmission, and ultimately improves the accuracy of disease degree identification.

[0051] 2. The present invention uses wheat multi-channel fusion images as input, which can make the prediction results of wheat lesion degree more accurate; at the same time, the degree is identified according to the classification model based on the lesion degree, thereby solving the problem of unclear edges of some lesion areas in the wheat lesion degree prediction results in the wheat lesion degree loss identification technology, thereby accurately evaluating the lesion degree level.

[0052] 3. This method uses adaptive soft thresholding to refine lesion edges, and multi-scale feature fusion to aggregate high-resolution information from low-level feature maps to restore small lesion features. EfficientNet-B4 is used as the network encoder, and a redesigned skip connection structure improves network performance and reduces the number of parameters. Finally, a decoder aggregates low-level to high-level features. During model training, a weighted combination loss is used to calculate loss, ultimately effectively improving lesion segmentation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart for identifying the degree of disease of the present invention;

[0054] Figure 1a This is a structural diagram of the feature extraction subnetwork module of the present invention;

[0055] Figure 1b This is a structural diagram of the feature downsampling subnetwork module of the present invention;

[0056] Figure 1c This is a structural diagram of the feature fusion and classification sub-network module of the present invention;

[0057] Figure 2 This is a diagram of the lesion segmentation structure of the present invention;

[0058] Figure 3a These are images of wheat with different disease degrees according to the present invention;

[0059] Figure 3b This is the wheat image disease degree recognition image of the present invention;

[0060] Figure 3c This is the diseased spot segmentation image of the wheat image of the present invention. DETAILED DESCRIPTION

[0061] In this embodiment, a method for wheat scab recognition and lesion segmentation based on convolutional neural networks mainly uses convolutional neural networks to extract features of wheat head scab images and fuses the extracted features using feature pyramid technology; then an adaptive soft threshold method is used and a multi-scale feature fusion method is used to segment lesions. Specifically, Figure 1 As shown, the steps are as follows:

[0062] Step 1: Use geometric transformation method and pixel transformation method to preprocess the healthy wheat ear images and wheat ear images with different diseased spots, and obtain the preprocessed wheat image dataset denoted as X = {x1, x2, ..., x i ,...,x N}, where x irepresents the i-th wheat image, i = 1, 2, ..., N, where N is the total number of images in the wheat image dataset X;

[0063] In this example, 500 wheat images were used, each containing multiple wheat ears, including both healthy wheat ears and wheat ears with varying degrees of disease. The collected images were 3900 pixels × 2600 pixels in RGB format.

[0064] First, the image was preprocessed to remove the background of the wheat ears from the collected image, leaving only the wheat ears, and the image size was unified to 224 pixels × 224 pixels. Finally, a total of 1250 wheat ear images of different disease levels (mild, moderate and severe) and healthy wheat ears were obtained. The processed wheat ear images are as follows Figure 3a shown.

[0065] The geometric transformation method used operates on the pixel positions in the image, only changing the pixel distribution and not adding new information to the image. Color transformation methods used include adding noise and color dithering. In this study, Gaussian noise and salt and pepper noise were added to the image.

[0066] The experimental dataset applies real-time data augmentation operations under the TensorFlow framework, and randomly uses the above data augmentation methods in each training batch to expand the dataset.

[0067] The diseased spot degree label set corresponding to the wheat image dataset X is recorded as Y = {y1, y2, ..., y i ,...,y N}, where y i Represents the i-th wheat image x i The lesion degree label, and y i ∈[0,M], M is the level of the lesion; when y i =0, indicating x i For healthy wheat ears image;

[0068] The lesion segmentation label set corresponding to the wheat image dataset X is denoted as Z = {z1,z2,...,z i ,...,z N}, where z i Represents the i-th wheat image x i The lesion segmentation labels.

[0069] Lesions were segmented using semantic segmentation based on a fully convolutional neural network, and the pixels in the image were divided into two categories: lesion part and background. In this experiment, a certain number of lesion images were labeled, and pixel-level labeling of fusarium susceptibility was achieved in the labeling tool LabelMe.

[0070] Step 2: Construct a wheat spot degree recognition network, such as Figure 1 As shown, it includes: feature extraction sub-network module, feature downsampling sub-network module, feature fusion and classification sub-network module;

[0071] Step 2.1, feature extraction sub-network module, such as Figure 1a As shown in Figure 1, all convolution and pooling modules are set up the same as before the last pooling layer of VGG16. Each module has multiple convolution layers with rectified linear units (ReLU) activations, using 3×3 filters, stride and padding of 1, and 2×2 max pooling layers with stride of 2. The convolution layers are used to perform a sliding window on the i-th wheat image x. i Perform feature extraction to obtain K mapping features of arbitrary size in, represents the kth mapping feature;

[0072] The kth mapping feature Input n spatial pooling blocks of different sizes for processing in sequence, obtain the local detail features output by each spatial pooling block, and after splicing the n local detail features, obtain the kth standardized mapping feature F i,k .

[0073] Step 2.2, feature downsampling sub-network module, such as Figure 1b As shown in Figure 1, it consists of s-level strip pooling layers. This module replaces the last pooling layer in VGG16. It mainly pools the feature map extracted in the first part through multi-level strip windows. On the one hand, it obtains global information through global pooling with windows of the same size as the feature map, and on the other hand, it obtains long-range local information through strip windows of other sizes. i,k After downsampling, the feature vector output by each strip pooling layer is obtained, and after splicing the s feature vectors, the kth downsampled feature vector FVD is obtained. i,k ;

[0074] Step 2.3, feature fusion and classification sub-network module, such as Figure 1c As shown, it includes: fully connected layer and Softmax classification module:

[0075] The K downsampled feature vectors {FVD i,k |k=1,2,…,K} is input into the fully connected layer for processing, and the fused wheat lesion degree feature map FD is obtained. i ;

[0076] FD iInput into the Softmax classification module for dimensionality reduction processing to obtain the predicted probability matrix F of the lesion degree i and from F i The degree of lesions corresponding to the maximum probability is selected as the i-th wheat image x i The predicted label y' i .

[0077] Step 3: Train the model:

[0078] Step 3.1: Use formula (1) to construct the multi-category cross entropy loss function L CE :

[0079]

[0080] Step 3.2: Use formula (2) to construct the Dice loss function L D :

[0081]

[0082] Step 3.3: Use formula (3) to construct the total loss function L:

[0083] L = α × L CE +β×L D (3)

[0084] In formula (3), α and β are L CE and L D The weight coefficient of the Dice loss function and the cross entropy loss function are linearly weighted. While paying attention to the overall image loss, more attention is paid to the loss of foreground targets, which effectively improves the prediction ability of the model.

[0085] Step 3.4: Input the wheat image dataset X into the disease spot degree recognition network, and use the gradient descent method to train the disease spot degree recognition network. At the same time, calculate the loss function L. When L converges, stop training to obtain the optimal wheat disease spot degree recognition network, which is used to identify the disease degree of wheat disease images. The disease spot degree recognition effect is as follows: Figure 3b shown.

[0086] Step 4: Construct a wheat spot segmentation network, such as Figure 2 As shown, it includes: an encoder module, a decoder module and a dense skip connection module;

[0087] Step 4.1, the encoder module includes P convolutional contraction blocks CDB1, CDB2, ..., CDB p ,...,CDB P And Q downsampling blocks DB1, DB2, ..., DB q ,...,DBQ , where CDB p represents the pth convolutional contraction block, DB q represents the qth downsampling block;

[0088] When p=1,q=1, for FD i After adjusting the image size and channel, the adjusted feature map is obtained And input the pth convolutional contraction block CDB p In this way, global average pooling is used to The absolute value of the dimensionality reduction process is performed, and the obtained one-dimensional feature vector Then input to the qth downsampling module DB q The size is reduced so that the output size is half of the input size, thus obtaining the qth low-level feature map

[0089] When p=2,3,...,P,q=2,3,...,Q,the p-1th low-level feature map Input the pth convolutional contraction block CDB p The one-dimensional feature vector obtained by processing Then input to the qth downsampling module DB q The size is reduced in the process to obtain the qth low-level feature map Thus we get the Qth downsampling module DB Q The Qth low-level feature map output

[0090] Using formula (4) for the qth low-level feature map Perform soft thresholding so that each value of the scale vector parameter is between 0 and 1. In order to avoid the threshold being a negative number, multiply the scale vector parameter by the channel average of the absolute value of the shrinkage layer input feature map to obtain the threshold of each channel. Get the qth enhanced low-level feature map Filter out noise and remove redundant information;

[0091]

[0092] In formula (4), σ represents the set threshold.

[0093] Step 4.2, the encoder module includes P convolutional contraction blocks CDB' P+1 ,CDB' P+2 ,...,CDB' P+p ,...,CDB' 2P and Q upsampling blocks UB1, UB2, ..., UB q ,...,UB Q , where CDB'P+p Indicates the P+pth convolutional contraction block; UB q represents the qth upsampling block;

[0094] When p=1, q=1, Input the P+pth convolution shrinkage block CDB' P+p The one-dimensional feature vector obtained by processing Then input to the qth upsampling module UB q In this way, bilinear interpolation is used to increase the output size to twice the input size to obtain the P+pth high-level feature map

[0095] When p=2,3,...,P,q=2,3,...,Q,the P+p-1th high-level feature map Input the P+pth convolution shrinkage block CDB' P+p The one-dimensional feature vector obtained by processing Then input to the qth upsampling module UB q Processed in, get the P+pth high-level feature map

[0096] Using formula (4) to calculate the high-level feature map Perform soft thresholding to obtain enhanced high-level feature maps

[0097] Step 4.3: The dense skip connection module makes each node interconnected, providing rich semantic features, and maps the output to the horizontal output layer. After upsampling, the high-level and low-level feature maps of different scales are aggregated together to achieve the final segmentation. Using formula (5), we can get the i-th wheat image x i Lesion segmentation feature map T i :

[0098]

[0099] In formula (5), Conv represents the convolution operation; D represents the downsampling operation; Up represents the upsampling operation; [] represents the feature map concatenation operation.

[0100] Step 4.4, training model:

[0101] Use formula (6) to construct the cycle focus loss function L CFL , which optimizes the update speed of neural network parameters by periodically adjusting the learning rate to avoid falling into the local optimal solution:

[0102]

[0103] In formula (6), ρ is a dynamically adjusted weight coefficient;

[0104] Step 4.5: Draw the wheat lesion degree characteristic map FD i Input the wheat lesion segmentation network and use the gradient descent method to train the wheat lesion segmentation network, and calculate the loss function L CFL , when L CFL The training is stopped when convergence occurs, and the optimal wheat lesion segmentation network is obtained, which is used to segment the wheat lesion image. The lesion segmentation effect is as follows: Figure 3c shown.

[0105] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0106] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are executed.

Claims

1. A method for wheat head blight identification and disease spot segmentation based on convolutional neural network, characterized in that: The steps include: Step 1: Obtain healthy wheat ear images and wheat ear images with different diseased spots and preprocess them to obtain the preprocessed wheat image dataset, which is recorded as ,in, Indicates the An image of wheat, , is the total number of images in the wheat image dataset X; Step 2: Construct a wheat spot degree recognition network, including: feature extraction sub-network module, feature downsampling sub-network module, feature fusion and classification sub-network module; and use it to identify the wheat spot degree. wheat image Processing is performed to extract the characteristic map of wheat lesion degree , and get the i-th wheat image The predicted label ; Step 2.1: The feature extraction sub-network module uses the convolution layer to perform a sliding window operation on the first wheat image Perform feature extraction to obtain K mapping features ;in, represents the kth mapping feature; The kth mapping feature Input n spatial pooling blocks of different sizes for processing in sequence, obtain the local detail features output by each spatial pooling block, and after splicing the n local detail features, obtain the kth standardized mapping feature ; Step 2.2: The feature downsampling sub-network module consists of s-level strip pooling layers, and sequentially After downsampling, the feature vector output by each strip pooling layer is obtained, and after splicing the s feature vectors, the kth downsampled feature vector is obtained. ; Step 2.3: The feature fusion and classification sub-network module includes: a fully connected layer and a Softmax classification module: The K downsampled feature vectors { |k=1,2,…,K} is input into the fully connected layer for processing, and the fused wheat spot degree feature map is obtained. ; Will Input into the Softmax classification module for dimensionality reduction processing to obtain the predicted probability matrix F of the lesion degree i and from F i The degree of lesions corresponding to the maximum probability is selected as the i-th wheat image The predicted label ; Step 3: Construct a wheat lesion segmentation network including an encoder module, a decoder module, and a dense skip connection module; and perform a segmentation on the wheat lesion degree feature map. Process and get the i-th wheat image Lesion segmentation feature map ; Step 3.1, the encoder module includes P convolutional contraction blocks CDB1, CDB2, ..., CDB p ,...,CDB P And Q downsampling blocks DB1, DB2, ..., DB q ,...,DB Q , where CDB p represents the pth convolutional contraction block, DB q represents the qth downsampling block; When p=1, q=1, After adjusting the image size and channel, the adjusted feature map is obtained , and input the pth convolutional contraction block CDB p In this way, global average pooling is used to The absolute value of the dimensionality reduction process is performed, and the obtained one-dimensional feature vector , and then input to the qth downsampling module DB q The size is reduced so that the output size is half of the input size, thus obtaining the qth low-level feature map ; When p=2,3,...,P,q=2,3,...,Q, the p-1th low-level feature map Input the pth convolutional contraction block CDB p The one-dimensional feature vector obtained is processed , and then input to the qth downsampling module DB q The size is reduced in the process to obtain the qth low-level feature map , thus obtaining the Qth downsampling module DB Q The Qth low-level feature map output ; Using formula (4) for the qth low-level feature map Perform soft thresholding to obtain the qth enhanced low-level feature map : = (4) In formula (4): Indicates the set threshold; Step 3.2: The encoder module includes P convolutional shrinkage blocks , ,..., ,..., and Q upsampling blocks , ,..., ,..., ,in, represents the P+pth convolutional contraction block; represents the qth upsampling block; When p=1, q=1, Input the P+pth convolution shrinkage block The one-dimensional feature vector obtained is processed , and then input to the qth upsampling module In this way, bilinear interpolation is used to increase the output size to twice the input size to obtain the P+pth high-level feature map ; When p=2,3,...,P,q=2,3,...,Q,the P+p-1th high-level feature map Input the P+pth convolution shrinkage block The one-dimensional feature vector obtained is processed , and then input to the qth upsampling module Processed in, get the P+pth high-level feature map ; Using formula (4) to calculate the high-level feature map Perform soft thresholding to obtain enhanced high-level feature maps ; Step 3.3: The dense skip connection module uses formula (5) to obtain the i-th wheat image Lesion segmentation feature map : (5) In formula (5), Conv represents the convolution operation; D represents the downsampling operation; Up represents the upsampling operation; [ ] represents the feature map concatenation operation.

2. The method for wheat head blight identification and disease spot segmentation based on convolutional neural network according to claim 1, wherein: Wheat Image Dataset The corresponding lesion degree label set is recorded as Y = {y1,y2,...,y i ,...,y N }, where y i Indicates the wheat image The lesion degree label, and ∈[0,M], M is the degree of the lesion; when =0, indicating For healthy wheat ears image; Construct a multi-category cross entropy loss function L CE and Dice loss function L D The total loss function L composed of Wheat Image Dataset The data is input into the lesion degree recognition network, and the gradient descent method is used to train the lesion degree recognition network. At the same time, the loss function L is calculated. When L converges, the training is stopped to obtain the optimal wheat lesion degree recognition network, which is used to identify the disease degree of wheat disease images.

3. The method for wheat head blight identification and disease spot segmentation based on convolutional neural network according to claim 2, wherein: The total loss function L is obtained by the following steps: Step 3.1: Use formula (1) to construct the multi-category cross entropy loss function L CE : L CE = (1) Step 3.2: Use formula (2) to construct the Dice loss function L D : L D = 1 - (2) Step 3.3: Use formula (3) to construct the total loss function L: L = α×L CE + β×L D (3) In formula (3), α and β are the multi-category cross entropy loss function L CE and Dice loss function L D The weight coefficient of .

4. The method for wheat head blight identification and disease spot segmentation based on convolutional neural network according to claim 1, wherein: Wheat Image Dataset The corresponding lesion segmentation label set is recorded as Z = {z1,z2,...,z i ,...,z N }, where z i Indicates the wheat image The lesion segmentation label; Use formula (6) to construct the cycle focus loss function : (6) In formula (6), It is a dynamically adjusted weight coefficient; The wheat spot segmentation network is trained using the gradient descent method, and the loss function is calculated. , until The training is stopped when convergence occurs, thereby obtaining the optimal wheat lesion segmentation network, which is used to segment wheat lesion images.

5. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the method for identifying wheat fusarium head blight and segmenting diseased spots as described in any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for identifying wheat fusarium head blight and segmenting diseased spots according to any one of claims 1 to 4 are executed.

Images