A Method for Detecting Fusarium Head Blight of Field Wheat Based on UAV Images

Abstract

The present invention relates to a method for detecting Fusarium head blight of field wheat based on UAV images, including: using a UAV equipped with a visible light sensor to obtain the image data U of field wheat suffering from Fusarium head blight, and simultaneously using a digital camera to continuously obtain the ground wheat image data G; cropping the wheat images in the field wheat image data U; using a dual regression network to perform super-resolution reconstruction on the cropped wheat images; performing data augmentation on the wheat images after super-resolution reconstruction to obtain a training set U Train ; constructing a detection network for Fusarium head blight of wheat based on feature enhancement and adaptive feature fusion; using the UAV to obtain the wheat images to be predicted and then performing the detection of Fusarium head blight to obtain the detection results. The present invention improves the resolution of the image data U of field wheat obtained by the UAV; adds a feature enhancement module on the basis of YOLOv5 and uses an adaptive feature fusion module to fuse multi-scale features, improving the detection ability of the network for small target lesions.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic detection of wheat scab, and in particular to a method for detecting wheat scab in a large field based on unmanned aerial vehicle (UAV) images. Background Art

[0002] Wheat is one of the most important food crops in the world. Its planting area and yield rank first among various food crops. The yield and quality of wheat have an important impact on human life. However, wheat is often infected by scab. Wheat scab can cause yield losses to wheat, and in severe cases, the yield loss can reach 30%. Moreover, wheat infected with scab will produce fungal toxins harmful to the human body. Wheat scab is caused by various Fusarium fungi. The pathogens can survive in seeds and soil and are spread by air currents, wind and rain. Therefore, it is crucial to quickly detect wheat scab with an effective and feasible method and spray pesticides to inhibit the spread of the pathogens.

[0003] Traditional methods for detecting wheat scab make corresponding disease judgments by manual observation, but this work is both time-consuming and susceptible to subjective influences. Detecting wheat scab on a large scale in a brand-new information-based manner and carrying out timely prevention and control is an inevitable choice for future agricultural development.

[0004] Computer vision and image processing technologies have been widely applied in the diagnostic monitoring of the agricultural field. Detecting crop diseases by processing images obtained by sensors is a simple and efficient method. However, most current methods process data taken close to the ground, and the ground data coverage is small. Moreover, plant diseases and pests often start to break out gradually from a small area. Therefore, detecting crop diseases and pests in the large-field environment in a timely manner can enable timely prevention and control. The UAV remote sensing system has a series of advantages such as small volume, light weight, low cost, and simple operation, and is less restricted by weather and cloud cover. Its flight height and flight time are flexible. It can obtain image data in a large range through the carried sensors, effectively making up for the defect of the small coverage of ground images and meeting the needs of precision agriculture management. Detecting wheat scab using the data taken by UAVs can quickly detect wheat scab in a large area, which is beneficial to timely prevention and control. Currently, there are relatively few methods for detecting wheat scab based on the images taken by UAVs. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for detecting wheat scab in a large field based on UAV images, which realizes the precise detection of wheat scab in a large field by super-resolution reconstruction of wheat images obtained by UAVs, constructing a detection network based on feature enhancement and adaptive feature fusion, and knowledge transfer of ground wheat images.

[0006] To achieve the above object, the present invention adopts the following technical solutions: A method for detecting Fusarium head blight of field wheat based on UAV images, the method comprising the following steps in sequence:

[0007] (1) Use a UAV to take wheat images to obtain field wheat image data U, and divide the field wheat image data U into a training set U Train and a test set U Test in a ratio of 2:1. Use a digital camera to take pictures at a height of 30 cm to 50 cm above the wheat to obtain ground wheat image data G;

[0008] (2) Crop the wheat images in the obtained training set U Train ;

[0009] (3) Use a dual regression network to perform super-resolution reconstruction on the cropped wheat images;

[0010] (4) Perform data augmentation on the wheat images after super-resolution reconstruction: Raise the pixel values of the RGB three channels of the wheat images after super-resolution reconstruction to the 3rd power and then normalize them to the range of 0 to 255 to obtain enhanced wheat image data U';

[0011] (5) Construct a Fusarium head blight detection network for wheat: Construct a Fusarium head blight detection network for wheat through YOLOv5, a feature enhancement module, and an adaptive feature fusion module;

[0012] (6) Use the transfer learning method to pre-train the Fusarium head blight detection network for wheat using the ground wheat image data G, and then use the training set U Train to fine-tune the parameters of the Fusarium head blight detection network for wheat: Use a labeling tool to label the Fusarium head blight lesions in the ground wheat image data G and the training set U Train . Use the ground wheat image data G to train the Fusarium head blight detection network for wheat to obtain pre-trained weights P. Use the training set U Train to train the Fusarium head blight detection network for wheat loaded with the weights P, and optimize the weights P using the stochastic gradient descent algorithm to obtain weights W;

[0013] (7) Load the weights W, crop the wheat images in the test set U Test , perform super-resolution reconstruction and data augmentation processing on them, send the processed images to the trained Fusarium head blight detection network for wheat for testing, and splice the test results to obtain the detection results of Fusarium head blight of field wheat.

[0014] The specific steps of step (3) include the following steps:

[0015] (3.1) An image I with a resolution of 4H*4W HRPerform downsampling and convolution operations to obtain feature I with a resolution of 2H * 2W F ;

[0016] (3.2) Downsample I F to obtain an image I with a resolution of H * W LR1 ;

[0017] (3.3) Use the RCAB residual channel attention module to extract features from I F to obtain an image I with a resolution of 2H * 2W LR2 ;

[0018] (3.4) Perform feature extraction and upsampling operations on image I LR2 to obtain a super-resolution image I with a resolution of 4H * 4W SR ;

[0019] (3.5) Perform convolution and downsampling on the super-resolution image I SR to obtain an image I' with a resolution of 2H * 2W LR2 and an image I' with a resolution of H * W LR1 ;

[0020] (3.6) Calculate the loss: Use the L1 loss function to calculate the errors between I SR and I HR , I LR2 and I' LR2 as well as I LR1 and I' LR1 and then sum up these three errors to obtain the loss L S , and the formula of the L1 loss function is as follows:

[0021]

[0022] where N represents the number of pixel points on an image, x i ' and x i respectively represent the pixel values at the relative positions on two images for which the error needs to be calculated;

[0023] (3.7) Optimize the dual regression network by performing gradient backpropagation according to the value of the loss L S .

[0024] The specific steps of step (4) are as follows:

[0025] (4.1) Split the wheat image obtained in step (3) into three channels: R, G, and B;

[0026] (4.2) Input the three channels of R, G, and B into a non-linear increasing function respectively to obtain R', G', and B', and the non-linear increasing function is as follows:

[0027] Pixel(x i ,y i ) r，g，b = Pixel 3 (x i ,y i ) r，g，b

[0028] Among them, Pixel(x i ,y i ) r，g，b represents the pixel values of the R, G, and B channels of the image;

[0029] (4.3) Perform a normalization operation on R′, G′, and B′, and normalize the pixels to the range of 0 to 255 to obtain R″, G″, and B″. The normalization formula is as follows:

[0030]

[0031] (4.4) After splicing the three channels of R″, G″, and B″, the enhanced wheat image data U′ is obtained.

[0032] The specific content of the said step (5) refers to:

[0033] Build a wheat scab detection network based on YOLOv5. The wheat scab detection network consists of a backbone part, a neck, and a detection head. The backbone part is used for feature extraction, the neck is used for multi-scale information feature fusion, and the detection head is used for analyzing the finally obtained features to obtain the detection results; the feature enhancement module consists of two branches. The first branch uses a spatial attention mechanism to obtain the local information of the features, and this process is represented by the following formula:

[0034] Y 1 = (1 + sigmoid(Conv 3×3 ([AvgPool(X), MaxPool(X)] concat )))*X

[0035] Among them, X is the input feature, Y 1 is the obtained output feature, sigmoid is the sigmoid activation function, Conv 3×3 is a convolution operation with a convolution kernel size of 3×3, AvgPool and MaxPool respectively represent the average pooling operation and the maximum pooling operation, and [] concat represents the splicing operation;

[0036] The second branch uses the form of cascaded dilated convolutions with different dilation rates to extract features and obtain the global information of the image. This process is represented by the following formula:

[0037] Y 2 = Conv d＝1 (X) + Conv d＝2 (Conv d＝1 (X)) + Conv d＝3 (Conv d＝2 (Conv d＝1 (X)))

[0038] where X is the input feature, and Y 2 is the obtained output feature, and Conv d＝1 , Conv d＝2 , Conv d＝3 represent dilated convolution operations with dilation rates of 1, 2, and 3 respectively;

[0039] After obtaining the global information and local information of the image, add the two, and learn the correlation between sub-regions and highlight the region of the object by establishing the interaction between the global information and the local information;

[0040] Perform multi-scale fusion on the features so that the features contain geometric information representation and semantic information representation, and use an adaptive feature fusion module to replace the path aggregation network used in YOLOv5 to fuse multi-scale information.

[0041] The step (6) specifically includes the following steps:

[0042] (6.1) Data annotation: Use the annotation tool Labelimg to annotate the Fusarium head blight lesions in the ground wheat image data G and the training set U Train to obtain the true position information of the lesions;

[0043] (6.2) Pre-training: Use the ground wheat image data G to pre-train the wheat Fusarium head blight detection network to obtain the pre-trained model weights P;

[0044] (6.3) Set the hyperparameters of training: The initial learning rate is 0.01, each training batch is set to 8, and the total number of training iterations is set to 500 epochs;

[0045] (6.4) The wheat Fusarium head blight detection network loads the trained weights P;

[0046] (6.5) Use the training set U without Train to train the wheat Fusarium head blight detection network loaded with the weights P: Input the training set U Train into the wheat Fusarium head blight detection network for feature extraction, then through feature enhancement and multi-scale feature adaptive fusion, and finally obtain three different prediction data through the detection head of the wheat Fusarium head blight detection network, including: predicted box coordinates, category, confidence;

[0047] (6.6) Calculate the loss: The cross - entropy loss function is used for the class loss and the confidence loss, and the GIOU loss is used for the location loss. Calculate the difference \(L\) between the predicted location and the true location, loc the difference \(L\) between the predicted class and the true class, cls and the difference \(L\) between the predicted confidence and the true confidence. cof Sum \(L\) loc , \(L\) cls and \(L\) cof to obtain the loss \(L\). D ;

[0048] (6.7) Optimize the wheat scab detection network by gradient backpropagation: Obtain the gradient according to the value of the loss \(L\), D and use the stochastic gradient descent algorithm to perform gradient backpropagation to update the weights, and finally obtain the weights \(W\).

[0049] As can be seen from the above technical solutions, the beneficial effects of the present invention are as follows: First, the present invention performs super - resolution reconstruction on the field wheat image data \(U\) obtained by the unmanned aerial vehicle with the ground wheat image data \(G\) as a reference to improve the resolution of the field wheat image data \(U\); Second, on the basis of YOLOv5, the present invention adds a feature enhancement module and uses an adaptive feature fusion module to replace the PANet path aggregation network to fuse multi - scale features, improving the network's detection ability for small target lesions; Third, the transfer learning method is used to pre - train the wheat scab detection network with the ground wheat image data \(G\), and then the training set \(U\) Train is used to fine - tune the parameters of the wheat scab detection network, enabling the network to have better generalization performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is the flowchart of the method of the present invention;

[0051] Figure 2 is the schematic diagram of training the wheat scab detection network using the pre - training and fine - tuning methods of the present invention;

[0052] Figure 3 is the schematic diagram of the wheat scab detection network in the present invention;

[0053] Figure 4 is the schematic diagram of the adaptive fusion method of multi - scale features used in the present invention;

[0054] Figure 5 is the process of training the wheat scab detection network loaded with weights \(P\) using the training set \(U\) Train in the present invention, and it is the change curve graph of the loss;

[0055] Figure 6 is the process of training the wheat scab detection network loaded with weights \(P\) using the training set \(U\) TrainCurve graph of the change in average precision (AP) during the process of training the Fusarium head blight detection network of wheat loaded with weights P. Detailed implementation manners

[0056] As Figure 1 shown, a method for detecting Fusarium head blight of field wheat based on UAV images, the method includes the following steps in sequence:

[0057] (1) Use a UAV to take wheat images to obtain field wheat image data U, divide the field wheat image data U into a training set U Train and a test set U Test in a ratio of 2:1, and use a digital camera to take pictures at a height of 30 cm to 50 cm above the wheat to obtain ground wheat image data G;

[0058] (2) Crop the wheat images in the obtained training set U Train ;

[0059] (3) Use a dual regression network to perform super-resolution reconstruction on the cropped wheat images;

[0060] (4) Perform data augmentation on the wheat images after super-resolution reconstruction: cube the pixel values of the three RGB channels of the wheat images after super-resolution reconstruction and then normalize them to the range of 0-255 to obtain enhanced wheat image data U';

[0061] (5) Construct a Fusarium head blight detection network: construct a Fusarium head blight detection network through YOLOv5, a feature enhancement module, and an adaptive feature fusion module;

[0062] (6) Use the transfer learning method to pre-train the Fusarium head blight detection network with the ground wheat image data G, and then use the training set U Train to fine-tune the parameters of the Fusarium head blight detection network: use an annotation tool to annotate the Fusarium head blight lesions in the ground wheat image data G and the training set U Train , use the ground wheat image data G to train the Fusarium head blight detection network to obtain pre-trained weights P, use the training set U Train to train the Fusarium head blight detection network loaded with weights P, and optimize the weights P using the stochastic gradient descent algorithm to obtain weights W;

[0063] (7) Load the weights W, crop the wheat images in the test set U Test , perform super-resolution reconstruction and data augmentation processing on them, send the processed images to the trained Fusarium head blight detection network for testing, and splice the test results to obtain the detection results of field wheat Fusarium head blight.

[0064] The specific steps of step (3) include the following steps:

[0065] (3.1) Downsample and perform convolution on the image I with a resolution of 4H * 4W HR to obtain the feature I with a resolution of 2H * 2W F ;

[0066] (3.2) Downsample I F to obtain the image I with a resolution of H * W LR1 ;

[0067] (3.3) Use the RCAB residual channel attention module to extract features from I F to obtain the image I with a resolution of 2H * 2W LR2 ;

[0068] (3.4) Perform feature extraction and upsampling on the image I LR2 to obtain the super-resolution image I with a resolution of 4H * 4W SR ;

[0069] (3.5) Perform convolution and downsampling on the super-resolution image I SR to obtain the image I' with a resolution of 2H * 2W LR2 and the image I' with a resolution of H * W LR1 ;

[0070] (3.6) Calculate the loss: Use the L1 loss function to calculate the errors between I SR and I HR , I LR2 and I', LR2 and between I LR1 and I' LR1 and then sum up these three errors to obtain the loss L S . The formula of the L1 loss function is as follows:

[0071]

[0072] where N represents the number of pixel points on an image, x i ' and x i respectively represent the pixel values at the relative positions on two images for which the error needs to be calculated;

[0073] (3.7) Optimize the dual regression network by performing gradient backpropagation according to the value of the loss L S .

[0074] The specific steps of the said step (4) include the following steps:

[0075] (4.1) Segment the wheat image obtained in step (3) into three channels: R, G, and B;

[0076] (4.2) Input the R, G, and B channels into a non - linear increasing function respectively to obtain R′, G′, and B′. The non - linear increasing function is as follows:

[0077] Pixel(x i , y i ) r，g，b =Pixel 3 (x i , y i ) r，g，b

[0078] Among them, Pixel(x i , y i ) r，g，b represents the pixel values of the R, G, and B channels of the image;

[0079] (4.3) Perform a normalization operation on R′, G′, and B′ to normalize the pixels to the range of 0 - 255 to obtain R″, G″, and B″. The normalization formula is as follows:

[0080]

[0081] (4.4) Concatenate the R″, G″, and B″ channels to obtain the enhanced wheat image data U′.

[0082] The specific content of step (5) is as follows:

[0083] Build a wheat scab detection network based on YOLOv5. The wheat scab detection network consists of a backbone part, a neck, and a detection head. The backbone part is used for feature extraction, the neck is used for multi - scale information feature fusion, and the detection head is used for analyzing the finally obtained features to get the detection results; the feature enhancement module consists of two branches. The first branch uses the spatial attention mechanism to obtain the local information of the features. This process is represented by the following formula:

[0084] Y 1 =(1 + sigmoid(Conv 3×3 ([AvgPool(X), MaxPool(X)] concat )))*X

[0085] Among them, X is the input feature, Y 1 is the obtained output feature, sigmoid is the sigmoid activation function, Conv 3×3 is a convolution operation with a convolution kernel size of 3×3, AvgPool and MaxPool represent the average pooling operation and the maximum pooling operation respectively, [] concat represents the concatenation operation;

[0086] The second branch extracts features in the form of cascaded dilated convolutions with different dilation rates to obtain the global information of the image. This process is represented by the following formula:

[0087] Y 2 = Conv d＝1 (X) + Conv d＝2 (Conv d＝1 (X)) + Conv d＝3 (Conv d＝2 (Conv d＝1 (X)))

[0088] where X is the input feature, Y 2 is the obtained output feature, and Conv d＝1 , Conv d＝2 , and Conv d＝3 represent dilated convolution operations with dilation rates of 1, 2, and 3 respectively;

[0089] After obtaining the global information and local information of the image, the two are added together, and the correlation between sub-regions is learned by establishing the interaction between the global information and the local information, and the region of the object is highlighted;

[0090] The features are fused at multiple scales so that the features contain geometric information representation and semantic information representation, and an adaptive feature fusion module is used to replace the path aggregation network used in YOLOv5 to fuse the multi-scale information.

[0091] Step (6) specifically includes the following steps:

[0092] (6.1) Data annotation: Use the annotation tool Labelimg to annotate the scab lesions in the ground wheat image data G and the training set U Train to obtain the true position information of the lesions;

[0093] (6.2) Pre-training: Use the ground wheat image data G to pre-train the wheat scab detection network to obtain the pre-trained model weights P;

[0094] (6.3) Set the hyperparameters for training: The initial learning rate is 0.01, each training batch is set to 8, and the total number of training iterations is set to 500 epochs;

[0095] (6.4) The wheat scab detection network loads the trained weights P;

[0096] (6.5) Use the training set U without Train to train the wheat scab detection network loaded with the weights P: Use the training set U TrainInput into the wheat scab detection network for feature extraction, then through feature enhancement and multi-scale feature adaptive fusion, and finally obtain three different prediction data through the detection head of the wheat scab detection network, including: predicted box coordinates, categories, and confidence levels;

[0097] (6.6) Calculate the loss: The category loss and confidence loss use the cross-entropy loss function, and the position loss uses GIOU loss. Calculate the differences L between the position and the true position loc , the difference L between the category and the true category cls and the difference L between the confidence level and the true confidence level cof , and sum L loc , L cls and L cof to obtain the loss L D ;

[0098] (6.7) Optimize the wheat scab detection network by gradient backpropagation: Obtain the gradient according to the value of the loss L D , and use the stochastic gradient descent algorithm to perform gradient backpropagation and update the weights, and finally obtain the weights W.

[0099] As Figure 2 shown, the present invention uses the ground wheat image data G to pre-train the wheat scab detection network to obtain the pre-trained weights P, and then uses the drone wheat scab image training set U Train to train the wheat scab detection network loaded with the weights P.

[0100] As Figure 3 shown, the wheat scab detection network consists of a backbone part, a neck, and a detection head. The backbone part is responsible for feature extraction, the neck is responsible for fusing multi-scale features, and the detection head is responsible for parsing the fused features to obtain the detection results.

[0101] As Figure 4 shown, the adaptive feature fusion module used in the present invention adaptively fuses multi-scale features by giving a normalized weight to the multi-scale features.

[0102] As Figure 5 shown, during the process of training the wheat scab detection network loaded with the weights P using the drone wheat scab image training set U Train in the present invention, the loss (Loss) gradually decreases and the network gradually fits;

[0103] As Figure 6 shown, the present invention uses the drone wheat scab image training set U TrainDuring the process of training the Fusarium head blight detection network loaded with weights P, the average precision (AP) detected gradually increases. However, when the network is overfitted after fitting to a certain extent, the average precision detected will decrease.

[0104] In summary, the present invention performs super-resolution reconstruction on the field wheat image data U obtained by the drone with reference to the ground wheat image data G, improving the resolution of the field wheat image data U. The present invention adds a feature enhancement module on the basis of YOLOv5 and uses an adaptive feature fusion module to replace the PANet path aggregation network to fuse multi-scale features, improving the network's detection ability for small target lesions. The present invention uses the transfer learning method to pre-train the Fusarium head blight detection network with the ground wheat image data G, and then uses the training set U Train to fine-tune the parameters of the Fusarium head blight detection network, enabling the network to have better generalization performance.

Claims

1. A method for detecting Fusarium head blight of field wheat based on UAV images, characterized in that: This method includes the following steps in sequence: (1) Use a drone to capture wheat images to obtain field wheat image data U, and divide the field wheat image data U into a training set U and a test set U at a ratio of 2:

1. Train Test Use a digital camera to take pictures at a height of 30 cm to 50 cm above the wheat to obtain ground wheat image data G; (2) Crop the wheat images in the obtained training set U Train ; (3) Use a dual regression network to perform super-resolution reconstruction on the cropped wheat images; (4) Perform data augmentation on the wheat images after super-resolution reconstruction: cube the pixel values of the RGB three channels of the wheat images after super-resolution reconstruction and then normalize them to the range of 0-255 to obtain the enhanced wheat image data U'; (5) Construct a Fusarium head blight detection network for wheat: Construct a Fusarium head blight detection network for wheat through YOLOv5, a feature enhancement module, and an adaptive feature fusion module; (6) Use the transfer learning method to pre-train the wheat scab detection network with the ground wheat image data G, and then use the training set U Train Fine-tune the parameters of the wheat scab detection network: Use the annotation tool to annotate the scab lesions in the ground wheat image data G and the training set U Train Train the wheat scab detection network with the ground wheat image data G to obtain the pre-trained weights P, and use the training set U Train Train the wheat scab detection network loaded with the weights P, and optimize the weights P using the stochastic gradient descent algorithm to obtain the weights W; (7) Load the weights W, crop the wheat images in the test set U Test and perform super-resolution reconstruction and data augmentation on them. Then send the processed images to the trained wheat scab detection network for testing. After obtaining the test results, splice them to get the detection results of wheat scab in the field; The specific description of step (5) is as follows: Based on YOLOv5, construct a Fusarium head blight detection network for wheat. The Fusarium head blight detection network for wheat consists of three parts: a backbone part, a neck, and a detection head. Among them, the backbone part is used for feature extraction, the neck is used for multi-scale information feature fusion, and the detection head is used for analyzing the finally obtained features to obtain the detection results; the feature enhancement module consists of two branches. The first branch uses a spatial attention mechanism to obtain local information of the features, and this process is represented by the following formula: Y 1 = (1 + sigmoid(Conv 3×3 ([AvgPool(X), MaxPool(X)] concat )))*X Where X is the input feature, Y 1 is the obtained output feature, sigmoid is the sigmoid activation function, Conv 3×3 is a convolution operation with a convolution kernel size of 3×3, AvgPool and MaxPool represent average pooling operation and max pooling operation respectively, conca t represents the concatenation operation; The second branch uses a cascaded form of dilated convolutions with different dilation rates to extract features and obtain the global information of the image, and this process is represented by the following formula: Y 2 = Conv d＝1 (X) + Conv d＝2 (Conv d＝1 (X)) + Conv d＝3 (Conv d＝2 (Conv d＝1 (X))) Where X is the input feature and Y 2 is the obtained output feature, and Conv d＝1 , Conv d＝2 , Conv d＝3 represent dilated convolution operations with dilation rates of 1, 2, and 3 respectively; After obtaining the global information and local information of the image, add the two to learn the correlation between sub-regions and highlight the regions of the object by establishing the interaction between the global information and the local information. Perform multi-scale fusion on the features so that the features contain geometric information representation and semantic information representation, and use an adaptive feature fusion module to replace the path aggregation network used by YOLOv5 to fuse multi-scale information.

2. The method for detecting Fusarium head blight of field wheat based on UAV images according to claim 1, characterized in that: The specific description of step (3) includes the following steps: (3.1) Downsample and convolve the image I with a resolution of 4H * 4W to obtain the feature I with a resolution of 2H * 2W HR F ;​ (3.2) Downsample I F to obtain an image I with a resolution of H*W LR1 ; (3.3) Use the RCAB residual channel attention module to process I F to extract features and obtain an image I with a resolution of 2H*2W LR2 ; (3.4) Perform feature extraction and upsampling operations on the image I LR2 to obtain a super-resolution image I with a resolution of 4H * 4W SR ; (3.5) Convolve and downsample the super-resolution image I SR to obtain an image I' with a resolution of 2H * 2W LR2 and an image I' with a resolution of H * W LR1 ; (3.6) Calculate the loss: Use the L1 loss function to calculate the error between I SR and I HR 、I LR2 and I′ LR2 as well as the error between I LR1 and I′ LR1 and sum up these three errors to obtain the loss L S . The formula for the L1 loss function is as follows: where N represents the number of pixel points on an image, and x i ' and x i respectively represent the pixel values of the relative positions on two images for which the error needs to be calculated; (3.7) Optimize the dual regression network by performing gradient backpropagation according to the value of the loss L S ​ 3. The method for detecting Fusarium head blight of field wheat based on UAV images according to claim 1, characterized in that: The specific description of step (4) includes the following steps: (4.1) Split the wheat images obtained after step (3) into three channels: R, G, and B; (4.2) Input the R, G, and B channels into a non-linear increasing function respectively to obtain R', G', and B'. The non-linear increasing function is as follows: Pixel(x i ,y i ) r，g，b =Pixel 3 (x i ,y i ) r，g，b Among them, Pixel(x i , y i ) r，g，b represents the pixel values of the R, G, and B channels of the image; (4.3) Perform a normalization operation on R', G', and B' to normalize the pixels to between 0 and 255 to obtain R'', G'', and B''. The normalization formula is as follows: (4.4) Concatenate the R'', G'', and B'' channels to obtain the enhanced wheat image data U'.

4. The method for detecting Fusarium head blight of field wheat based on UAV images according to claim 1, characterized in that: The specific description of step (6) includes the following steps: (6.1) Data annotation: Use the annotation tool Labelimg to annotate the Fusarium head blight lesions in the ground wheat image data G and the training set UTrai n to obtain the true position information of the lesions; (6.2) Pre-training: Use the ground wheat image data G to pre-train the Fusarium head blight detection network for wheat to obtain the pre-trained model weights P; (6.3) Set the hyperparameters for training: the initial learning rate is 0.01, each training batch is set to 8, and the total number of training iterations is set to 500 epochs; (6.4) The wheat scab detection network loads the trained weights P; (6.5) Use the untrained set U Train Train the Fusarium head blight detection network loaded with weights P: Input the training set U Train into the Fusarium head blight detection network for feature extraction, then through feature enhancement and multi-scale feature adaptive fusion, and finally obtain three different prediction data through the detection head of the Fusarium head blight detection network, including: predicted box coordinates, category, and confidence; (6.6) Calculate the loss: The categorical loss and the confidence loss adopt the cross-entropy loss function, and the localization loss adopts the GIOU loss. Calculate the difference L between the predicted location and the ground truth location loc , the difference L between the predicted category and the ground truth category cls , and the difference L between the predicted confidence and the ground truth confidence cof . Sum L loc , L cls , and L cof to obtain the loss L D ; (6.7)Gradient backpropagation to optimize the wheat scab detection network: Calculate the gradient based on the value of the loss L D and use the stochastic gradient descent algorithm to backpropagate the gradient for weight update, finally obtaining the weight W.

Images