A pine wood nematode disease tree detection method based on neural architecture search

Abstract

A pine wood nematode disease tree detection method based on neural architecture search includes the following steps: step 1: collecting images, marking spring and summer pictures containing disease tree samples, and making data sets; step 2: using the spring and summer data set to train an initial training model, loading the initial training model into the neural architecture search network, and using the autumn and winter data set to train the detection model; step 3: put the test picture set into the trained neural architecture search network for detection, make the misrecognized disease trees as negative sample set, mix the negative sample set with the autumn and winter training set, and train again to obtain the detection model; step 4: output the recognition result of the detection model as a vector to obtain the latitude and longitude coordinates of the disease tree. The purpose of the present application is to better extract the characteristics of the pine wood nematode disease tree for detection, and provide a robust method for detecting autumn and winter pine wood nematode disease trees based on unmanned aerial vehicle images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection in remote sensing images, and particularly relates to a detection method for pine wilt disease trees in autumn and winter. BACKGROUND

[0002] Pine wilt disease, also known as pine wilt, is a devastating pine disease caused by pine wood nematode, and belongs to a major invasive disease in China with an occurrence area of 10.26 million mu, causing direct economic losses and ecological service value losses of over 100 billion yuan. If the pine wilt disease epidemic is not effectively controlled, not only the landscape dominated by pine trees will no longer exist, but also the ecological environment will be severely damaged.

[0003] Traditional pine wilt disease identification usually adopts manual positioning method, however, the manual positioning method not only consumes time and resources, but also is difficult to implement in some places.

[0004] The method for target detection in remote sensing images has problems of multi-time phase and low picture acquisition efficiency in actual application. There is a significant difference in background environment between spring and summer remote sensing and autumn and winter remote sensing images. The forest environment is more complex in autumn and winter, there are a large number of normal trees similar to the target diseased trees in color and texture, and the target diseased trees are small in the image, so that the detection accuracy is reduced and the method cannot be practically applied.

[0005] In the research on the existing pine wilt disease tree detection method, Xu Xinluo et al. use Faster R-CNN target detection algorithm for automatic identification of diseased color pine trees, modify the anchor size in the region generation network according to the crown size of the damaged wood, and then detect the diseased trees. Li Hao et al. use Faster R-CNN model to detect pine trees with different degrees of disease. These detection methods do not consider that the background environment of spring and summer unmanned aerial vehicle images is significantly different from that of autumn and winter unmanned aerial vehicle images, and the forest environment is more complex in autumn and winter, so it is more difficult to extract the features of the target diseased trees. SUMMARY

[0006] The present application aims to solve the technical problem of the existing technology that it is difficult to detect pine wilt disease trees in autumn and winter as pointed out in the background, and the existing technology only detects pine wilt disease trees in spring and summer, ignoring the great influence of red broad-leaved trees and yellow broad-leaved trees on the detection of red pine wilt disease trees due to seasonal changes. Therefore, we provide a robust method for detecting pine wilt disease trees in autumn and winter based on image processing technology.

[0007] A pine wilt disease tree detection method in autumn and winter based on neural architecture search, comprising the following steps:

[0008] Step 1: image acquisition, marking pictures containing diseased tree samples in autumn and winter and pictures containing diseased tree samples in spring and summer respectively to make spring and summer dataset 1 and autumn and winter dataset 2, and making unmarked pictures into test picture set 3;

[0009] Step 2: training the initial network 4 using the spring and summer dataset 1 to obtain an initial training model, loading the initial training model into the neural architecture search network FED 5 as a pre-training model, constructing a sampling screening network and training the neural architecture search network FED 5 using the autumn and winter dataset 2 to obtain a detection model;

[0010] Step 3: placing the test picture set 3 into the trained neural architecture search network FED 5 for detection, counting the samples that are incorrectly identified and making them into a negative sample set 6, mixing the negative sample set with the autumn and winter training set 2 and training again, repeatedly iterating and training the detection model to improve the robustness of the detection model, and obtaining a detection model with optimal parameters;

[0011] Step 4: outputting the identification results of the detection model with optimal parameters as a vector and correcting the position to obtain a diseased tree center point latitude and longitude coordinate file;

[0012] Through the above steps, the pine wood nematode diseased trees are detected.

[0013] In step 2, when the detection model is obtained, the following sub-steps are adopted:

[0014] 2-1: the initial network 4 adopts a faster R-CNN network, and the faster R-CNN network is trained using the spring and summer dataset 1 to obtain an initial training model;

[0015] 2-2: loading the parameters of the initial training model into the neural architecture search network FED 5;

[0016] 2-3: at this time, the initial training model contains basic parameter information of the diseased trees in the spring and autumn dataset 1, the spring and summer dataset 1 and the autumn and winter dataset 2 are input into the neural architecture search network FED 5 to train the network, and the parameters are updated on the basis of the basic parameter information to obtain a better detection model.

[0017] In step 2-3, when the spring and summer dataset 1 and the autumn and winter dataset 2 are used to train the neural architecture search network FED 5, two losses are obtained, which are shown in the following formulas:

[0018] L1=L 1reg +L 1cls (1)

[0019] L2=L 2reg +L2cls (2)

[0020] wherein L1 represents the total loss obtained by training the network using the spring-summer data set 1, L 1reg and L 1cls respectively represent the regression loss and the classification loss obtained by training the network using the spring-summer training set 1, L2 represents the total loss obtained by training the network using the autumn-winter data set 2, L 2reg and L 2cls respectively represent the regression loss and the classification loss obtained by training the network using the autumn-winter training set 2.

[0021] The initial training model contains the parameters of the diseased trees in the spring-summer data set 1, and the neural architecture search network 5 is trained using the autumn-winter data set 2. When updating the parameters in the initial training model, part of the information may be lost, so the spring-summer data set 1 is added during training and a smaller weight is given to the loss to reduce information loss and improve the generalization ability of the network. The specific formula is as follows:

[0022]

[0023] wherein L is the total loss of network training, is a hyperparameter, which aims to control the influence of different losses, wherein L1 represents the total loss obtained by training the network using the spring-summer data set 1, and L2 represents the total loss obtained by training the network using the autumn-winter data set 2.

[0024] In step 3, when the detection model with the optimal parameters is obtained, the following steps are adopted:

[0025] 3-1: Put the test picture set into the trained neural architecture search network FED 5 for detection;

[0026] 3-2: Automatically divide the detection results according to the confidence threshold, and regard the positive samples as those with confidence greater than a specified value A, and regard the negative samples as those with confidence less than a specified value B;

[0027] 3-3: Add the negative samples to the autumn-winter data set 2, and then put them into the network for retraining;

[0028] 3-4: Repeat steps 3-2 to 3-3 until all the confidence in the detection results is greater than the specified value A.

[0029] The neural architecture search network FED includes a feature extraction module Resnet50, a feature fusion module ENF 7, a sampling and screening module RPN, and a detection head Double head, and the specific structure is as follows:

[0030] 4-1: The pictures of the autumn-winter data set are input into the feature extraction module Resnet50 to obtain the feature map of the diseased trees;

[0031] 4-2: Input the extracted feature map into the feature fusion module ENF 7 for feature fusion to obtain the fused diseased tree feature map;

[0032] 4-3: Input the fused diseased tree feature map into the sampling and filtering module RPN for sampling to obtain the pre-selection box;

[0033] 4-4: Map the preselected bounding boxes onto the fused disease tree feature map and input them into the Double head for classification and regression to obtain the final detection results.

[0034] In step 4-3, the sampling and filtering module RPN is constructed, specifically through the following steps:

[0035] 5-1: Sample diseased trees and use the sampling filtering module RPN to generate pre-selected boxes on the feature map through a sliding window method. The pre-selected boxes can be regarded as boxes predefined according to a fixed ratio.

[0036] 5-2: The preselected boxes are suppressed by nonmaximum to obtain the final candidate boxes;

[0037] 5-3: Calculate the Intersection over Union (IOU) between candidate bounding boxes and labeled ground truth values. The IOU calculation formula is:

[0038]

[0039] Samples with an IOU value greater than the set threshold of 0.6 are considered positive samples, samples with an IOU value less than the set threshold of 0.4 are considered negative samples, and samples with an IOU value between the threshold of 0.4 and 0.6 are considered difficult samples. Difficult samples will have an adverse effect on network training, so difficult samples with an IOU value between the threshold of 0.4 and 0.6 are discarded.

[0040] In step 4-2, when using the feature fusion module ENF, such as Figure 3 As shown, five layers of feature maps are input, a large number of features are fused on the bottom feature maps, and a small number of features are fused on the high feature maps, and five layers of feature maps are output.

[0041] Its specific structure is as follows:

[0042] Input layer: Input feature maps of 7 layers: C1, C2, C3, C4, C5, C6, and C7. However, layers C1 and C2 would greatly increase the computational parameters of the network, resulting in increased training time, so they were discarded. Layers C6 and C7 are obtained by downsampling C5.

[0043] First layer: Perform 1×1 convolution on feature maps C7, C6, C5, C4, and C3 respectively to obtain C71, C61, C51, C41, and C31;

[0044] The second layer: fusing the feature maps C5, C51, C61 and C71 to obtain a feature map C52, fusing the feature maps C4, C41 and C52 to obtain a feature map C42, and fusing the feature maps C3, C31 and C42 to obtain a feature map C32;

[0045] The output layer: outputting the feature maps C71, C61, C52, C42 and C32 after convolution and fusion operations to obtain final output feature maps P7, P6, P5, P4 and P3, respectively.

[0046] Finally, the obtained five-layer feature maps are automatically searched to search for an optimal structure in a given space, and the network for extracting features is updated.

[0047] Compared with the prior art, the present application has the following technical effects:

[0048] 1) The present application proposes a method for detecting pine wood nematode disease trees in autumn and winter with high robustness through unmanned aerial vehicle images. By automatically searching the artificially designed feature extraction network, the feature extraction capability is strengthened, the problem of complex forest environment due to the change of time phase and the existence of a large number of normal trees with high similarity in color and texture features to the target disease trees is solved, thereby improving the detection accuracy of disease trees in autumn and winter and obtaining a network model with better robustness.

[0049] 2) The present application solves the problem of reduced detection accuracy due to seasonal changes in actual application by using a transfer learning training strategy. By transferring the image information in spring and summer, the detection accuracy of pine wood nematode disease trees in autumn and winter is improved. A good detection accuracy can be obtained with only a small amount of unmanned aerial vehicle labeled samples, greatly reducing the labor cost of labeling new data sets and improving the generalization ability of the model.

[0050] 3) The present application converts the position of the diseased tree into a vector output, which can accurately locate the latitude and longitude of the diseased tree in the remote sensing image after registration, avoiding manual inspection and reducing labor costs.

[0051] 4) Compared with the prior art, the present application combines the artificially designed feature extraction network with the neural network search architecture, proposes a feature extraction network that is more suitable for a specific data set and meets the target expectations, and can strengthen the extraction of target features to improve the detection accuracy. The present application takes into account the change of background environment due to the change of season, and proposes to use a transfer learning training strategy to transfer the image information in spring and summer to the image in autumn and winter, thereby improving the generalization ability of the network model. The above improvement methods can greatly improve the accuracy of the detection framework. BRIEF DESCRIPTION OF DRAWINGS

[0052] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0053] Figure 1 This is a flowchart of the training strategy of the present invention;

[0054] Figure 2 for Figure 1 FED network flowchart;

[0055] Figure 3 for Figure 2 Network flowchart of the ENF module; Detailed Implementation

[0056] A method for detecting pine wilt disease in autumn and winter based on neural architecture search includes the following steps:

[0057] Step 1: Collect images and label the images containing diseased tree samples in autumn and winter and spring and summer to create Spring / Summer Dataset 1 and Autumn / Winter Dataset 2 respectively. Create a test image set 3 from the unlabeled images.

[0058] Step 2: Train the initial network 4 using the spring / summer dataset 1 to obtain the initial training model. Load the initial training model as a pre-trained model into the neural architecture search network FED 5, construct a sampling and filtering network, and train the neural architecture search network FED 5 using the autumn / winter dataset 2 to obtain the detection model.

[0059] Step 3: Input the test image set 3 into the trained neural architecture search network FED 5 for detection, count the samples that are incorrectly identified and make them into negative sample set 6. Mix the negative sample set with the autumn and winter training set 2 and train again. Iterate and train the detection model repeatedly to improve the robustness of the detection model and obtain the detection model with the optimal parameters.

[0060] Step 4: Output the identification results of the detection model with optimal parameters as a vector, correct the position, and obtain the latitude and longitude coordinate file of the center point of the diseased tree;

[0061] The above steps were used to detect pine wilt disease in trees.

[0062] Figure 1 This is a flowchart of the training strategy of the present invention. Figure 1As can be seen, the initial network faster R-CNN is trained using the spring-summer data set 1 to obtain an initial training model, and the parameters of the initial training model are loaded into the neural architecture search network 5, then the neural architecture search network 5 is trained using the autumn-winter data set 2 to obtain a detection model, the test picture set 3 is put into the trained neural architecture search network 5 for detection, the samples that are incorrectly recognized are counted and made into a negative sample set 6, the negative sample set is mixed with the autumn-winter training set 2 and then trained again, and the detection model is repeatedly iterated and trained; the robustness of the detection model is improved through repeated iteration, and a detection model with optimal parameters is obtained;

[0063] In step 2, when the detection model is obtained, the following substeps are adopted:

[0064] 2-1: The initial network 4 adopts the faster R-CNN network, and the faster R-CNN network is trained using the spring-summer data set 1 to obtain an initial training model;

[0065] 2-2: The parameters of the initial training model are loaded into the neural architecture search network FED 5;

[0066] 2-3: At this time, the initial training model contains the basic parameter information of the diseased trees in the spring-summer data set 1, and the spring-summer data set 1 and the autumn-winter data set 2 are input into the neural architecture search network FED to train the network, and the parameters are updated on the basis of the basic parameter information to obtain a better detection model.

[0067] In step 2-3, when the neural architecture search network FED 5 is trained using the spring-summer data set 1 and the autumn-winter data set 2, two losses are obtained, as shown in the following formulas:

[0068] L1=L 1reg +L 1cls (1)

[0069] L2=L 2reg +L 2cls (2)

[0070] Wherein L1 represents the total loss obtained by training the network using the spring-summer data set 1, L 1reg and L 1cls represent the regression loss and the classification loss obtained when the network is trained using the spring-summer training set 1, L2 represents the total loss obtained by training the network using the autumn-winter data set 2, L 2reg and L 2cls represent the regression loss and the classification loss obtained when the network is trained using the autumn-winter training set 2;

[0071] The initial training model includes parameters related to diseased trees from images in the Spring / Summer dataset 1. The Neural Architecture Search Network 5 is then trained using the Autumn / Winter dataset 2. Updating the parameters in the initial training model may result in some information loss. Therefore, the Spring / Summer dataset 1 is added during training and given a smaller weight for its loss function to reduce information loss and improve the network's generalization ability. The specific formula is as follows:

[0072]

[0073] Where L is the total training loss of the network. L1 is a hyperparameter designed to control the impact of different losses, where L1 represents the total loss obtained by training the network using the Spring / Summer dataset 1, and L2 represents the total loss obtained by training the network using the Autumn / Winter dataset 2.

[0074] In step 3, the following steps are used to obtain the detection model with optimal parameters:

[0075] 3-1: The test image set is fed into the trained neural architecture search network FED 5 for detection;

[0076] 3-2: The detection results are automatically divided according to the confidence threshold. Result with a confidence level greater than 0.7 is regarded as a positive sample and result with a confidence level less than 0.3 is regarded as a negative sample.

[0077] 3-3: Add the negative samples to the Autumn / Winter dataset 2, and then put them into the network for retraining;

[0078] 3-4: Repeat steps 3-2 to 3-3 until all confidence levels in the test results are greater than 0.7.

[0079] like Figure 2 As shown, Figure 2 for Figure 1 The FED network flowchart in China, from Figure 2 As can be seen, after obtaining the pre-trained model, the network model is trained using the Autumn / Winter Dataset 4. After passing through the backbone network, it is input into our hand-designed feature extraction network for feature extraction. Then, after passing through the sampling and filtering network and the detection head, the final detection model is obtained.

[0080] The neural architecture search network FED 5 includes a feature extraction module ResNet50, a feature fusion module ENF 7, a sampling and filtering module RPN, and a detection head Double head. Its specific structure is as follows:

[0081] 4-1: Input the images from the autumn / winter dataset into the ResNet50 feature extraction module to obtain the feature map of the diseased tree;

[0082] 4-2: Input the extracted feature map into the feature fusion module ENF 7 for feature fusion to obtain the fused diseased tree feature map;

[0083] 4-3: Input the fused diseased tree feature map into the sampling and filtering module RPN for sampling to obtain the pre-selection box;

[0084] 4-4: Map the preselected bounding boxes onto the fused disease tree feature map and input them into the Double head for classification and regression to obtain the final detection results.

[0085] In step 4-3, the sampling and filtering module RPN is constructed, specifically through the following steps:

[0086] 5-1: Sample diseased trees and use the sampling filtering module RPN to generate pre-selected boxes on the feature map through a sliding window method. The pre-selected boxes can be regarded as boxes predefined according to a fixed ratio.

[0087] 5-2: The preselected boxes are suppressed by nonmaximum to obtain the final candidate boxes;

[0088] 5-3: Calculate the Intersection over Union (IOU) between candidate bounding boxes and labeled ground truth values. The IOU calculation formula is:

[0089]

[0090] Samples with an IOU value greater than the set threshold of 0.6 are considered positive samples, samples with an IOU value less than the set threshold of 0.4 are considered negative samples, and samples with an IOU value between the threshold of 0.4 and 0.6 are considered difficult samples. Difficult samples will have an adverse effect on network training, so difficult samples with an IOU value between the threshold of 0.4 and 0.6 are discarded.

[0091] Figure 3 for Figure 2 The network flowchart of the ENF module, from Figure 3 As can be seen, after the image passes through the backbone network, it obtains 5 feature maps. After multiple feature fusions of the 5 feature maps, they are input into the neural architecture search network for automatic search to obtain the feature extraction network that best matches our dataset.

[0092] In step 4-2, in the feature fusion module ENF 7, a network structure such as... Figure 3 The network shown performs feature fusion. It takes five layers of feature maps as input, performs extensive feature fusion on the bottom feature maps, performs minimal feature fusion on the high feature maps, and outputs five layers of feature maps.

[0093] Its specific structure is as follows:

[0094] Input layer: Input feature maps of 7 layers: C1, C2, C3, C4, C5, C6, and C7. However, layers C1 and C2 would greatly increase the computational parameters of the network, resulting in increased training time, so they were discarded. Layers C6 and C7 are obtained by downsampling C5.

[0095] First layer: Perform 1×1 convolution on feature maps C7, C6, C5, C4, and C3 respectively to obtain C71, C61, C51, C41, and C31;

[0096] Second layer: Fuse feature maps C5, C51, C61 and C71 to obtain feature map C52 → Fuse feature maps C4, C41 and C52 to obtain feature map C42 → Fuse feature maps C3, C31 and C42 to obtain feature map C32;

[0097] Output layer: The feature maps C71, C61, C52, C42, and C32 after convolution and fusion operations are output respectively to obtain the final output feature maps P7, P6, P5, P4, and P3;

[0098] Finally, the obtained 5-layer feature maps are automatically searched to find the optimal structure in the given space and the feature extraction network is updated.

[0099] Finally, the detection model with optimal parameters is obtained.

[0100] Example:

[0101] In this embodiment, the training epock parameter is 12, the batchsize parameter for the number of input images per cycle is 8, and hyperparameters control the impact of different losses.

[0102] By iterating and training the detection model repeatedly, the robustness of the detection model is improved, and the detection model with optimal parameters is obtained. The model is then used to detect actual diseased trees, and the detection accuracy and recall rate are obtained.

[0103] In this embodiment, a network based on transfer learning and neural architecture search was used to detect diseased trees, achieving a precision of 86.7% and a recall of 93.6%. Without transfer learning and using only neural architecture search, the precision reached 82.4% and the recall was 89.5%. Without using either transfer learning or neural architecture search, the precision was only 78.2%, and the recall was 85.3%. This demonstrates that the present invention can effectively detect pine wilt diseased trees in autumn and winter.

[0104] In this embodiment, the experimental results are compared with networks that do not use transfer learning and neural architecture search, as well as other networks. The results are shown in Table 1:

[0105] Table 1: Experimental results of pine wilt disease on trees in autumn and winter

[0106]

[0107] In this embodiment, the identification results of the detection model with optimal parameters are output as vectors, and the position is corrected to obtain the latitude and longitude coordinates file of the center point of the diseased tree. The results are shown in Table 2.

[0108] Table 2: Coordinate information of diseased trees

[0109]

[0110] This invention provides a method based on transfer learning and neural architecture search. It takes into account the limitations of manually designed feature extraction networks, which cannot fit the dataset with the best structure to perform the most effective feature extraction. This results in low accuracy in detecting pine wilt disease trees in the complex environment of autumn and winter. Simply searching the original feature extraction network from scratch will greatly increase the training time and the training cost is too high.

[0111] This method combines a manually designed feature extraction network with a neural architecture search method. The designed feature extraction network is searched to obtain the network that best fits our dataset and meets our expectations. This effectively extracts features from the target diseased trees, significantly reducing search time and extracting features from the target diseased trees in the fastest and most efficient way. Furthermore, this method uses a transfer learning training strategy, employing feature parameters from spring and summer pine wilt diseased trees to assist in detecting autumn and winter pine wilt diseased trees, training a more robust detection model to improve the detection accuracy of autumn and winter pine wilt diseased trees.

Claims

1. A method for detecting pine wilt disease in autumn and winter based on neural architecture search, characterized in that, It includes the following steps: Step 1: Collect images, label the images containing diseased tree samples in autumn and winter and the images containing diseased tree samples in spring and summer respectively, and create spring / summer datasets and autumn / winter datasets. Create a test image set from the unlabeled images. Step 2: Train the initial network using the spring / summer dataset to obtain the initial training model. Load the initial training model as a pre-trained model into the neural architecture search network FED, construct the sampling and filtering network, and train the neural architecture search network FED using the autumn / winter dataset to obtain the detection model. Step 3: Input the test image set into the trained neural architecture search network FED for detection, count the samples that are incorrectly identified and make them into a negative sample set. Mix the negative sample set with the autumn / winter dataset and train again. Iterate and train the detection model repeatedly to improve the robustness of the detection model and obtain the detection model with the optimal parameters. Step 4: Output the identification results of the detection model with optimal parameters as a vector, correct the position, and obtain the latitude and longitude coordinate file of the center point of the diseased tree; The above steps were used to detect pine wilt disease in trees. The initial network adopts the Faster R-CNN network, and the Faster R-CNN network is trained using the Spring / Summer dataset to obtain the initial training model; The neural architecture search network FED includes a feature extraction module Resnet50, a feature fusion module ENF, a sampling and filtering module RPN, and a detection head Double head; In step 3, when obtaining the detection model with optimal parameters, the following steps are adopted: 3-1: Feed the test image set into the trained neural architecture search network FED for detection; 3-2: The detection results are automatically divided according to the confidence threshold. Samples with a confidence level greater than the specified value A are considered positive samples, and samples with a confidence level less than the specified value B are considered negative samples. 3-3: Add negative samples to the autumn / winter dataset and then put them into the network for retraining; 3-4: Repeat steps 3-2 and 3-3 until all confidence levels in the test results are greater than the specified value A.

2. The method according to claim 1, characterized in that, In step 2, the following sub-steps are used when obtaining the detection model: 2-1: The initial network uses the Faster R-CNN network, and the Spring / Summer dataset is used to train the Faster R-CNN network to obtain the initial training model; 2-2: Load the parameters of the initial trained model into the neural architecture search network FED; 2-3: At this point, the initial training model contains the basic parameter information of diseased trees in the spring and summer datasets. The spring and summer datasets and the autumn and winter datasets are input into the neural architecture search network FED to train the network. Based on the basic parameter information, the parameters are updated to obtain a better detection model.

3. The method according to claim 2, characterized in that, In steps 2-3, when training the neural architecture search network FED using the spring / summer and autumn / winter datasets, two losses are obtained, as shown in the following formulas: ； ； in This represents the total loss obtained by training the network using the spring / summer dataset. and These represent the regression loss and classification loss obtained when training the network on the spring and summer datasets, respectively. This represents the total loss obtained by training the network using the autumn / winter dataset. and These represent the regression loss and classification loss obtained when training the network on the autumn / winter dataset, respectively. The initial training model includes parameters related to diseased trees from images in the spring and summer datasets. Then, the neural architecture search network is trained using the autumn and winter datasets. Updating the parameters in the initial training model may result in the loss of some spring and summer diseased tree information. Therefore, the spring and summer datasets are added during training and assigned a weight to reduce information loss and improve the network's generalization ability. The specific formula is as follows: ； in L For the total training loss of the network, It is a hyperparameter whose purpose is to control the impact of different losses, among which This represents the total loss obtained by training the network using the spring / summer dataset. This represents the total loss obtained by training the network using the autumn / winter dataset.

4. The method according to claim 1, characterized in that, The neural architecture search network FED includes a feature extraction module ResNet50, a feature fusion module ENF, a sampling and filtering module RPN, and a detection head Double head, specifically: 4-1: Input the images from the autumn / winter dataset into the ResNet50 feature extraction module to obtain the feature map of the diseased tree; 4-2: Input the extracted feature map into the feature fusion module ENF for feature fusion to obtain the fused diseased tree feature map; 4-3: Input the fused diseased tree feature map into the sampling and filtering module RPN for sampling to obtain the pre-selection box; 4-4: Map the preselected bounding boxes onto the fused disease tree feature map and input them into the Double head for classification and regression to obtain the final detection results.

5. The method according to claim 4, characterized in that, In step 4-3, the sampling and filtering module RPN is constructed, specifically through the following steps: 5-1: Sample diseased trees and use the sampling filtering module RPN to generate pre-selected boxes on the feature map through a sliding window method. The pre-selected boxes can be regarded as boxes predefined according to a fixed ratio. 5-2: The preselected boxes are suppressed by nonmaximum to obtain the final candidate boxes; 5-3: Calculate the Intersection over Union (IOU) between candidate bounding boxes and labeled ground truth values. The IOU calculation formula is: ； Samples with an IOU value greater than the set threshold of 0.6 are considered positive samples, samples with an IOU value less than the set threshold of 0.4 are considered negative samples, and samples with an IOU value between the threshold of 0.4 and 0.6 are considered difficult samples. Difficult samples will have an adverse effect on network training, so difficult samples with an IOU value between the threshold of 0.4 and 0.6 are discarded.

6. The method according to claim 4, characterized in that, In step 4-2, when using the feature fusion module ENF, five layers of feature maps are input, a large amount of feature fusion is performed on the bottom feature maps, and a small amount of fusion is performed on the high-level feature maps, outputting five layers of feature maps; specifically as follows: Input layer: Input feature maps of 7 layers: C1, C2, C3, C4, C5, C6, and C7. However, layers C1 and C2 would greatly increase the computational parameters of the network, resulting in increased training time, so they were discarded. Layers C6 and C7 are obtained by downsampling C5. First layer: Perform 1×1 convolution on feature maps C7, C6, C5, C4, and C3 respectively to obtain C71, C61, C51, C41, and C31; Second layer: Fuse feature maps C5, C51, C61 and C71 to obtain feature map C52 → Fuse feature maps C4, C41 and C52 to obtain feature map C42 → Fuse feature maps C3, C31 and C42 to obtain feature map C32; Output layer: The feature maps C71, C61, C52, C42, and C32 after convolution and fusion operations are output respectively to obtain the final output feature maps P7, P6, P5, P4, and P3; Finally, the obtained 5-layer feature maps are automatically searched to find the optimal structure in the given space and the feature extraction network is updated.

Images