Crop pest detection method and system based on image super-resolution reconstruction network
By using an image super-resolution reconstruction network, which utilizes hierarchical residual feature aggregation blocks and a hybrid attention module, the problem of insufficient utilization of feature information at different levels in existing technologies is solved, thereby improving the accuracy of crop pest detection and reducing equipment deployment costs.
Patent Information
- Application Number
- CN202310077864.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-01-16
- Publication Date
- 2025-12-16
- Estimated Expiration
- 2043-01-16
AI Technical Summary
Existing deep learning-based methods for detecting crop pests fail to fully utilize feature information at different levels in image super-resolution reconstruction, resulting in insufficient network reconstruction capabilities and affecting detection accuracy.
An image super-resolution reconstruction network is employed, comprising a generator network and a discriminator network. It utilizes hierarchical residual feature aggregation blocks and a hybrid attention module to perform deep feature extraction and upsampling through the generator network, and combines it with a target detection network for pest detection.
It improved the accuracy of crop pest detection, reduced the number of cameras deployed in farmland, and lowered costs.
Smart Images
Figure CN116030027B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of image super-resolution reconstruction and target detection, and relates to a crop pest detection method and system, in particular to a crop pest detection method and system based on image super-resolution reconstruction. BACKGROUND
[0002] The production of crops is related to many factors, such as climate change, plant diseases and pests. According to recent research, about half of the world's crop production is lost each year due to pests and crop diseases, so it is of great significance to monitor, identify and control pests in crops.
[0003] In the research of image-based crop pest detection, according to the different feature extraction methods, it can be divided into traditional pattern recognition-based method and deep learning-based method. The pattern recognition-based method often relies on multiple analysis and experiments of professionals to design feature vectors for classifier training, which has the problem of low generalization; the pest detection technology based on deep learning uses convolutional neural network to extract pest features end-to-end, which has potential connections and strong expression ability that cannot be found by human eyes, so the method based on deep learning has better generalization.
[0004] In the actual monitoring and identification of crop pests, the cameras in the farmland are relatively sparse, and even if these cameras capture pests, their pixels are very small and the resolution is low, so pest detection is essentially a small target detection problem. Small target refers to the number of target pixels being less than 32*32. In the field of computer vision, due to the low image coverage and less appearance clues, small target detection generally faces more difficulties than general target detection. In this regard, we can perform super-resolution reconstruction on the crop pest image to improve the resolution of the small target pest, so that the target detection network can obtain and learn more details and features of the pest, thereby improving the detection accuracy of the target detection network for crop pests.
[0005] Image super-resolution refers to restoring a low-resolution image that is not clear into a clear high-resolution image. Image super-resolution has good application scenarios in medical images, photo restoration, remote sensing images and many other fields. Image super-resolution reconstruction methods can be divided into three categories: interpolation method, reconstruction method and learning method. The first two methods are generally difficult to recover high-frequency detail information, resulting in blurred reconstructed images. The learning-based method, especially the deep learning-based method, has better effect on the two evaluation indexes of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). In recent years, with the rapid development of deep learning and the continuous improvement of people's quality of generated images, deep learning method has become an important branch in the field of image super-resolution reconstruction.
[0006] Although the current deep learning-based image super-resolution method for crop pests has been significantly improved, most of the existing methods still face the following limitations: many deep models ignore the irrelevance of different levels of features, which weakens the network presentation ability. Different levels often contain various feature information, and fully utilizing them can further improve the network reconstruction ability. SUMMARY
[0007] In order to solve the above technical problems, the present application provides a crop pest detection method and system based on image super-resolution reconstruction.
[0008] The technical scheme adopted by the method of the present application is: a crop pest detection method based on an image super-resolution reconstruction network, comprising the following steps:
[0009] Step 1: obtaining a crop pest image;
[0010] Step 2: inputting the crop pest image into an image super-resolution reconstruction network to perform super-resolution reconstruction on the crop pest image and obtaining a high-resolution image;
[0011] The image super-resolution reconstruction network is composed of a generator network and a discriminator network; wherein the discriminator is composed of VGG19, and the generator network is composed of three parts: head, trunk and reconstruction part; inputting a crop pest image, first using a 3*3 convolution as the head to perform initial feature extraction, then using ten consecutive hierarchical residual feature aggregation blocks to form the trunk to perform deep feature extraction, and finally using a sub-pixel convolution and a 3*3 convolution as the reconstruction part to perform up-sampling and decoding on the image feature information, and finally obtaining a high-resolution crop pest image;
[0012] The hierarchical residual feature aggregation block comprises a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a fifth convolution block, a sixth convolution block, a seventh convolution block, a first 1*1 convolution fusion block, a second 1*1 convolution fusion block, a third 1*1 convolution fusion block, a fourth 1*1 convolution fusion block, a fifth 1*1 convolution fusion block, a sixth 1*1 convolution fusion block, a first residual addition operation block, a second residual addition operation block and a third residual addition operation block;
[0013] The first convolution block, the second convolution block, the third convolution block, the fourth convolution block, the fifth convolution block, the sixth convolution block and the seventh convolution block are each composed of a 3*3 convolution layer, an LReLU activation layer, a 3*3 convolution layer and a hybrid attention block connected in sequence.
[0014] The hybrid attention block comprises a spatial attention block, a channel attention block and a self-attention block arranged in parallel; the outputs of the spatial attention block, the channel attention block and the self-attention block and the original input are output after passing through the matrix addition operation layer.
[0015] When the crop pest image feature information is input into the hierarchical residual feature aggregation block, the crop pest image feature information passes through the first convolution block, the second convolution block, the third convolution block, the fourth convolution block, the fifth convolution block and the sixth convolution block in sequence; the output of the first convolution block and the crop pest image feature information are fused through the first 1*1 convolution fusion block; the output of the second convolution block and the output of the first 1*1 convolution fusion block are fused through the second 1*1 convolution fusion block; the output of the third convolution block and the output of the second 1*1 convolution fusion block are fused through the third 1*1 convolution fusion block; the output of the fourth convolution block and the output of the third 1*1 convolution fusion block are fused through the fourth 1*1 convolution fusion block; the output of the fifth convolution block and the output of the fourth 1*1 convolution fusion block are fused through the fifth 1*1 convolution fusion block; the output of the sixth convolution block and the output of the fifth 1*1 convolution fusion block are fused through the sixth 1*1 convolution fusion block; the outputs of the first convolution block and the second convolution block are input into the third convolution block after passing through the first residual addition operation block, and the outputs of the fourth convolution block and the fifth convolution block are input into the sixth convolution block after passing through the second residual addition operation block; the output of the sixth 1*1 convolution fusion block is output through the output of the seventh convolution block and the crop pest image feature information after passing through the third residual addition operation block.
[0016] Step 3: input the reconstructed high-resolution image into a target detection network to detect the pests.
[0017] The technical scheme of the system of the present application is as follows: a crop pest detection system based on an image super-resolution reconstruction network, comprising the following modules:
[0018] Module 1, for acquiring crop pest images;
[0019] Module 2, for inputting the crop pest images into an image super-resolution reconstruction network, performing super-resolution reconstruction on the crop pest images, and obtaining high-resolution images;
[0020] The image super-resolution reconstruction network is composed of a generator network and a discriminator network; wherein the discriminator is composed of VGG19, and the generator network is composed of three parts: a head, a trunk and a reconstruction part; inputting a crop pest image, first using a 3*3 convolution as the head to perform initial feature extraction, then using ten consecutive hierarchical residual feature aggregation blocks to constitute the trunk to perform deep feature extraction, and finally using a sub-pixel convolution and a 3*3 convolution as the reconstruction part to perform up-sampling and decoding on the image feature information, and finally obtaining a high-resolution crop pest image;
[0021] The hierarchical residual feature aggregation block comprises a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a fifth convolution block, a sixth convolution block, a seventh convolution block, a first 1*1 convolution fusion block, a second 1*1 convolution fusion block, a third 1*1 convolution fusion block, a fourth 1*1 convolution fusion block, a fifth 1*1 convolution fusion block, a sixth 1*1 convolution fusion block, a first residual addition operation block, a second residual addition operation block and a third residual addition operation block;
[0022] The first convolution block, the second convolution block, the third convolution block, the fourth convolution block, the fifth convolution block, the sixth convolution block and the seventh convolution block are each composed of a 3*3 convolution layer, an LReLU activation layer, a 3*3 convolution layer and a hybrid attention block connected in sequence;
[0023] The hybrid attention block comprises a spatial attention block, a channel attention block and a self-attention block arranged in parallel; the outputs of the spatial attention block, the channel attention block and the self-attention block and the original input are output after passing through the matrix addition operation layer;
[0024] The crop pest image feature information sequentially passes through the first convolutional block, the second convolutional block, the third convolutional block, the fourth convolutional block, the fifth convolutional block and the sixth convolutional block when inputting the hierarchical residual feature aggregation block; the crop pest image feature information is fused with the first convolutional block output through the first 1*1 convolutional fusion block; the second convolutional block output is fused with the first 1*1 convolutional fusion block output through the second 1*1 convolutional fusion block; the third convolutional block output is fused with the second 1*1 convolutional fusion block output through the third 1*1 convolutional fusion block; the fourth convolutional block output is fused with the third 1*1 convolutional fusion block output through the fourth 1*1 convolutional fusion block; the fifth convolutional block output is fused with the fourth 1*1 convolutional fusion block output through the fifth 1*1 convolutional fusion block; the sixth convolutional block output is fused with the fifth 1*1 convolutional fusion block output through the sixth 1*1 convolutional fusion block; the outputs of the first convolutional block and the second convolutional block are input into the third convolutional block after passing through the first residual addition operation block, and the outputs of the fourth convolutional block and the fifth convolutional block are input into the sixth convolutional block after passing through the second residual addition operation block; the output of the sixth 1*1 convolutional fusion block is output after passing through the third residual addition operation block together with the crop pest image feature information through the output of the seventh convolutional block;
[0025] Module 3 is used for inputting the reconstructed high-resolution image into a target detection network to detect pests.
[0026] Compared with the prior art, the present application has the following advantages:
[0027] 1) The hierarchical residual feature aggregation block can effectively learn and fuse pest information features of different levels, and better utilize local residual features.
[0028] 2) The hybrid attention module composed of a spatial attention module, a channel attention module and a self-attention module can make the feature information concentrate on important spatial content, help utilize the mutual dependence between feature channels, and make the features at each position more coordinated with the surrounding parts.
[0029] 3) The super-resolution reconstruction processing of the crop pest image by the present application can effectively improve the detection effect of the crop pest image, thereby helping to further reduce the deployment of cameras in farmland and save costs, which has practical significance. BRIEF DESCRIPTION OF DRAWINGS
[0030] Figure 1 The network structure diagram of the hierarchical residual feature aggregation block in the generator of the embodiment of the present application.
[0031] Figure 2Network structure diagram of a convolution block in a hierarchical residual feature aggregation block of an embodiment.
[0032] Figure 3 Overall structure diagram of a hybrid attention block of an embodiment.
[0033] Figure 4 Network structure diagram of a spatial attention block of an embodiment.
[0034] Figure 5 Network structure diagram of a channel attention block of an embodiment.
[0035] Figure 6 Network structure diagram of a self-attention block of an embodiment.
[0036] Figure 7 Flowchart of an image super-resolution reconstruction network training process of an embodiment of the application.
[0037] Figure 8 Images of four types of crop pest in the AgriPest dataset, where (a) is Sitobion avenae (SA for short), (b) is Rice planthopper (RPH for short), (c) is Rhopalosiphum maidis (RM for short), and (d) is Cruciferae padi (CP for short).
[0038] Figure 9 Comparison of visual effects of the application and previous methods on super-resolution reconstruction. DETAILED DESCRIPTION
[0039] In order to facilitate those skilled in the art to understand and implement the present application, the present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments described herein are only used to illustrate and explain the present application, and are not intended to limit the present application.
[0040] Although the current deep learning-based image super-resolution method for crop pests has made significant improvements, most existing methods still face the following limitations: many deep models ignore the irrelevance of different levels of features, which weakens the network presentation ability, and different levels often contain various feature information, and fully utilizing them can further improve the network reconstruction ability.
[0041] The present application aims at the deficiencies of the prior art, and provides a crop pest detection method and system based on an image super-resolution reconstruction network. In order to make full use of feature information at different levels, the present application designs the image super-resolution reconstruction network to be composed of a generator network and a discriminator network; the discriminator is composed of VGG19, and the generator network is composed of three parts: a head, a trunk and a reconstruction part; an input crop pest image is firstly subjected to initial feature extraction using a 3*3 convolution as the head, then subjected to deep feature extraction using ten hierarchical residual feature aggregation blocks in succession to constitute the trunk, and finally subjected to up-sampling and decoding of image feature information using a sub-pixel convolution and a 3*3 convolution as the reconstruction part, so as to obtain a high-resolution crop pest image.
[0042] The hierarchical residual feature aggregation block has a structure as shown in Figure 1 The hierarchical residual feature aggregation block has a structure as shown in Figure 1 The first six convolution blocks are used for extracting hierarchical features, the output of each convolution block is connected with the hierarchical feature of the previous layer, and 1*1 convolution is used to fuse the features, so that the hierarchical crop pest information features can be more fully learned. Meanwhile, in order to avoid the problems of gradient vanishing or gradient explosion, residual addition operation is performed every two convolution blocks, and finally all feature details are refined through a convolution block.
[0043] The crop pest image feature information sequentially passes through the first convolutional block, the second convolutional block, the third convolutional block, the fourth convolutional block, the fifth convolutional block and the sixth convolutional block when inputting the hierarchical residual feature aggregation block; the crop pest image feature information and the first convolutional block output are fused through the first 1*1 convolutional fusion block; the second convolutional block output and the first 1*1 convolutional fusion block output are fused through the second 1*1 convolutional fusion block; the third convolutional block output and the second 1*1 convolutional fusion block output are fused through the third 1*1 convolutional fusion block; the fourth convolutional block output and the third 1*1 convolutional fusion block output are fused through the fourth 1*1 convolutional fusion block; the fifth convolutional block output and the fourth 1*1 convolutional fusion block output are fused through the fifth 1*1 convolutional fusion block; the sixth convolutional block output and the fifth 1*1 convolutional fusion block output are fused through the sixth 1*1 convolutional fusion block; the outputs of the first convolutional block and the second convolutional block are input into the third convolutional block after the first residual addition operation block; the outputs of the fourth convolutional block and the fifth convolutional block are input into the sixth convolutional block after the second residual addition operation block; the output of the sixth 1*1 convolutional fusion block is output through the output of the seventh convolutional block, and the crop pest image feature information is output after the third residual addition operation block;
[0044] The structures of the first convolutional block, the second convolutional block, the third convolutional block, the fourth convolutional block, the fifth convolutional block, the sixth convolutional block and the seventh convolutional block of the embodiment are shown in Figure 2 respectively, and each is composed of sequentially connected 3*3 convolutional layers, LReLU activation layers, 3*3 convolutional layers and hybrid attention blocks, and the hybrid attention block can maximize the feature extraction capability of the convolutional block in the hierarchical residual feature aggregation block.
[0045] The structures of the hybrid attention blocks of the embodiment are shown in Figure 3 respectively. The hybrid attention block includes a spatial attention block, a channel attention block and a self-attention block arranged in parallel; the outputs of the spatial attention block, the channel attention block and the self-attention block and the original input are output after a matrix addition operation layer. The spatial attention block helps to concentrate feature information on important spatial content, the channel attention block helps to utilize the mutual dependence between feature channels, so that the network pays attention to more informative features, and the self-attention block helps to establish long-distance and multi-level dependencies across image regions, so that the fine details of the crop pest image reconstructed by super-resolution at each position are coordinated with the fine details of the remote part of the image.
[0046] The spatial attention block helps to concentrate feature information on important spatial content, and its network structure is shown in Figure 4As shown in the figure; the spatial attention block, the original input sequentially passes through a 1*1 convolution layer, a 3*3 convolution layer, a 7*7 pooling layer, a 3*3 convolution layer, a 3*3 convolution layer, a 3*3 convolution layer, and an up-sampling layer, and the output is added to the output of the 1*1 convolution layer after the residual addition operation block, and then the output sequentially passes through a 1*1 convolution layer and a sigmoid layer, and then the output is output after the matrix point multiplication operation layer with the original input.
[0047] The channel attention block helps to utilize the mutual dependence between feature channels, so that the network pays attention to more informative features, and its network structure is as shown in the figure Figure 5 ; the channel attention block, the original input sequentially passes through a global average pooling layer, a 1*1 convolution layer, an LReLU activation layer, a 1*1 convolution layer, and a sigmoid layer, and then the output is output after the matrix point multiplication operation layer with the original input.
[0048] And the self-attention block helps to establish long-distance and multi-level dependencies across image regions, so that the fine details of the crop pest image reconstructed by super-resolution at each position are coordinated with the fine details of the remote part of the image, and its network structure is as shown in the figure Figure 6 . The self-attention block, after the 2*2 convolution pooling layer, the input is three channels; the first channel is a 1*1 convolution layer, a matrix reshaping layer, and a matrix transpose layer connected in sequence; the second channel is a 1*1 convolution layer and a matrix reshaping layer connected in sequence, and the output is output after the matrix multiplication layer with the matrix transpose layer output of the first channel, and then sequentially passes through a Softmax layer and a matrix transpose layer and is output; the third channel is a 1*1 convolution layer and a matrix reshaping layer connected in sequence, and the output is output after the matrix multiplication layer with the matrix transpose layer output of the second channel, and then sequentially passes through a 1*1 convolution layer and an up-sampling layer and is output.
[0049] The image super-resolution reconstruction network of the embodiment is a trained image super-resolution reconstruction network; see Figure 7 , and its training process includes the following steps:
[0050] Step 1, obtain a plurality of crop pest data sets, and divide the data set images into a training set and a test set; wherein the data set uses crop pest images AgriPest in a real agricultural scene, and see Figure 8 ; and randomly selected 772 images as the training set and 148 images as the test set.
[0051] Step 2, perform bicubic down-sampling processing on the data set to obtain a low-resolution image (LR) corresponding to a high-resolution image (HR), and construct a high-resolution and low-resolution image pair (LR-HR image pair);
[0052] Step 3, input the LR-HR image pair to train the image super-resolution reconstruction network.
[0053] After the LR-HR image pair input image super-resolution reconstruction network, the head of the generator network will first extract the initial features of the LR image, and the trunk part of the generator network will deeply extract and learn the features of the LR image, and finally the reconstruction part of the generator network will upsample and decode the feature information of the LR image.
[0054] Step 4, output the super-resolution reconstructed image SR.
[0055] Step 5, compare the output SR image with the HR image, calculate the pixel loss and the content-aware loss.
[0056] The pixel loss function is:
[0057]
[0058] In the formula, h, w, and c are the height, width, and channel number of the image, respectively. i,j,k is the pixel value in the real high-definition image with height i, width j, and width k, is the pixel value in the generated high-definition image with height i, width j, and width k;
[0059] The content-aware loss function is:
[0060]
[0061] In the formula, φ i,j,k (I) represents the feature value in the real high-definition image after being recognized by the VGG network with height i, width j, and width k; represents the feature value in the generated high-definition image after being recognized by the VGG network with height i, width j, and width k;
[0062] (6) The discriminator network discriminates between the SR image and the HR image, and calculates the generative adversarial loss.
[0063] The generative adversarial loss function is:
[0064] D Ra (x r ,x f )=σ(C(x r )-E[C(x f )]);
[0065] D Ra (x f ,x r )=σ(C(x f )-E[C(x r )]);
[0066]
[0067] where σ(·) is a sigmoid function, C(·) is the output of the discriminator, E[·] denotes the average, x f is the generated high-definition image, and x r is the real high-definition image. During the training of the discriminator network, the overall loss function is:
[0068]
[0069] (7) Calculate and back-propagate the total loss;
[0070] During the training of the generator network, the overall loss function is composed of the pixel loss function, the content-aware loss function, and the generative adversarial loss function, specifically:
[0071]
[0072] (8) Determine whether the training is complete. If not, return to (2) to continue training. If yes, end the training and obtain the trained image super-resolution reconstruction network.
[0073] The magnification scale of the image super-resolution reconstruction network is 2. When training the overall network, the size of the data set image is first cropped to 64x64, and the data set is randomly flipped by 90 degrees, 180 degrees, 270 degrees, and horizontally flipped to achieve data augmentation. In each training mini-batch, 16 low-resolution images of size 32x32 are used, where the low-resolution images are obtained by bicubic down-sampling from high-resolution images of size 64x64. In addition, the evaluation results of crop pest image super-resolution reconstruction are evaluated by PNSR and SSIM on the Y channel of the transformed YcbCr space. Use Adam as the optimizer, train for 250 epochs, with an initial learning rate of 0.0001, and halve the learning rate at [0.125*epoch, 0.25*epoch, 0.5*epoch, 0.75*epoch].
[0074] To evaluate the effectiveness of the method of the present application, ablation experiments are first carried out: first, the method of the present application is used to train without spatial attention blocks (HRFABNet-SAM); second, the method of the present application is used to train without channel attention blocks (HRFABNet-CAM); third, the method of the present application is used to train without self-attention blocks (HRFABNet-SelfAM); finally, the method of the present application is implemented (HRFABNet). Then the method of the present application is compared with advanced image super-resolution reconstruction methods such as ESRGAN, ECBSR, RFANet, etc. in performance. Finally, the method of the present application is combined with the DynamicRCNN target detection method to detect crop pests, and compared with other advanced methods.
[0075] Table 1
[0076]
[0077] Table 1 is the comparison experiment results of the present application and HRFABNet-SAM, HRFABNet-CAM, HRFABNet-SelfAM on the AgriPest dataset. Through the comparison results, it can be seen that the method proposed in the present application is the highest in the SSIM and PSNR evaluation indexes.
[0078] Table 2
[0079]
[0080] Table 2 is the comparison of the present application and other methods on the AgriPest dataset, through the comparison results, it can be seen that the method proposed in the present application is the highest in the SSIM and PSNR evaluation indexes. In addition, from the visual comparison in Table 2, it can be seen that the present application is better than the previous methods. Figure 9
[0081] Table 3
[0082]
[0083]
[0084] Table 3 is the comparison of the present application and other methods in the improvement effect of crop pest detection after combining with the DynamicRCNN target detection method. Through the comparison results, it can be seen that the present application has the highest improvement in average accuracy.
[0085] Using the present application to perform super-resolution reconstruction processing on crop pest images can effectively improve the detection effect of crop pest images, thereby helping to further reduce the deployment of cameras in farmland, save costs, and have practical significance.
[0086] It should be understood that the above description is merely a detailed explanation of the preferred embodiments and is not intended to limit the patent protection scope of the present application. Based on the above disclosure, those skilled in the art can make substitutions or modifications without departing from the scope of the present application. The patent protection scope of the present application should be subject to the appended claims.
Claims
1. A crop pest detection method based on an image super-resolution reconstruction network, characterized in that, The method comprises the following steps: Step 1: obtaining a crop pest image; Step 2: inputting the crop pest image into an image super-resolution reconstruction network to perform super-resolution reconstruction on the crop pest image to obtain a high-resolution image; The image super-resolution reconstruction network is composed of a generator network and a discriminator network; The discriminator is composed of VGG19, and the generator network is composed of three parts: a head, a trunk and a reconstruction part; inputting a crop pest image, first using a 3*3 convolution as the head to perform initial feature extraction, then using ten hierarchical residual feature aggregation blocks to constitute the trunk to perform deep feature extraction, and finally using a sub-pixel convolution and a 3*3 convolution as the reconstruction part to perform up-sampling and decoding on the image feature information, and finally obtaining a high-resolution crop pest image; The hierarchical residual feature aggregation block comprises a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a fifth convolution block, a sixth convolution block, a seventh convolution block, a first 1*1 convolution fusion block, a second 1*1 convolution fusion block, a third 1*1 convolution fusion block, a fourth 1*1 convolution fusion block, a fifth 1*1 convolution fusion block, a sixth 1*1 convolution fusion block, a first residual addition operation block, a second residual addition operation block and a third residual addition operation block; The first convolution block, the second convolution block, the third convolution block, the fourth convolution block, the fifth convolution block, the sixth convolution block and the seventh convolution block are each composed of a 3*3 convolution layer, an LReLU activation layer, a 3*3 convolution layer and a hybrid attention block connected in sequence; The hybrid attention block comprises a spatial attention block, a channel attention block and a self-attention block arranged in parallel; the outputs of the spatial attention block, the channel attention block and the self-attention block and the original input are output after a matrix addition operation layer. The crop pest image feature information sequentially passes through the first convolution block, the second convolution block, the third convolution block, the fourth convolution block, the fifth convolution block and the sixth convolution block when inputting the hierarchical residual feature aggregation block; the crop pest image feature information is fused with the first convolution block output through the first 1*1 convolution fusion block; the second convolution block output is fused with the first 1*1 convolution fusion block output through the second 1*1 convolution fusion block; the third convolution block output is fused with the second 1*1 convolution fusion block output through the third 1*1 convolution fusion block; the fourth convolution block output is fused with the third 1*1 convolution fusion block output through the fourth 1*1 convolution fusion block; the fifth convolution block output is fused with the fourth 1*1 convolution fusion block output through the fifth 1*1 convolution fusion block; the sixth convolution block output is fused with the fifth 1*1 convolution fusion block output through the sixth 1*1 convolution fusion block; the outputs of the first convolution block and the second convolution block are input into the third convolution block after passing through the first residual addition operation block, and the outputs of the fourth convolution block and the fifth convolution block are input into the sixth convolution block after passing through the second residual addition operation block; the output of the sixth 1*1 convolution fusion block is output through the output of the seventh convolution block, and the crop pest image feature information is output after passing through the third residual addition operation block; Step 3: input the reconstructed high-resolution image into the target detection network to detect the pests.
2. The crop pest detection method based on the image super-resolution reconstruction network according to claim 1, characterized in that: The spatial attention block, the original input sequentially passes through a 1*1 convolution layer, a 3*3 convolution layer, a 7*7 pooling layer, a 3*3 convolution layer, a 3*3 convolution layer, a 3*3 convolution layer, and an up-sampling layer, and the output after the residual addition operation block of the output of the 1*1 convolution layer is sequentially passed through a 1*1 convolution layer and a sigmoid layer, and the output after the matrix point multiplication operation layer of the original input is output. 3.The crop pest detection method based on image super-resolution reconstruction network according to claim 1, characterized in that: The channel attention block, the original input sequentially passes through a global flat pooling layer, a 1*1 convolution layer, an LReLU activation layer, a 1*1 convolution layer, and a sigmoid layer, and the output after the matrix point multiplication operation layer of the original input is output.
4. The crop pest detection method based on the image super-resolution reconstruction network according to claim 1, characterized in that: The self-attention block, the original input is input into three channels after passing through a 2*2 convolution pooling layer; the first channel is sequentially connected with a 1*1 convolution layer, a matrix reshaping layer, and a matrix transposition layer; the second channel is sequentially connected with a 1*1 convolution layer and a matrix reshaping layer, and the output after the matrix multiplication layer of the matrix transposition layer output of the first channel is sequentially output after passing through a Softmax layer and a matrix transposition layer; The third channel is sequentially connected with a 1*1 convolution layer and a matrix reshaping layer, and the output after the matrix multiplication layer of the matrix transposition layer output of the second channel is sequentially output after passing through a 1*1 convolution layer and an up-sampling layer.
5. The crop pest detection method based on the image super-resolution reconstruction network according to any one of claims 1-4, characterized in that: The image super-resolution reconstruction network is a trained image super-resolution reconstruction network; The training process includes the following steps: (1) Obtain a plurality of crop pest data sets, and divide the data set images into a training set and a test set; (2) bicubic down-sampling the data set to obtain a low-resolution image LR corresponding to a high-resolution image HR, and constructing a high-resolution and low-resolution image pair LR-HR image pair; (3) inputting the LR-HR image pair to train the image super-resolution reconstruction network; After the LR-HR image pair is input into the image super-resolution reconstruction network, the head of the generator network first extracts the initial features of the LR image, and then the trunk of the generator network deeply extracts and learns the features of the LR image, and finally the reconstruction part of the generator network up-samples and decodes the feature information of the LR image; (4) outputting a super-resolution reconstructed image SR; (5) comparing the output SR image with the HR image to calculate the pixel loss and the content-aware loss; The pixel loss function is: In the formula , , respectively the height, width, and channel number of the image; the pixel value in the real high-definition image with height , width , width ; the pixel value in the generated high-definition image with height , width , width ; The content-aware loss function is: In the formula, represent the feature values in the VGG network after the real high-definition image is identified, with a height of , and a width of , and a width of ; represent the feature values in the VGG network after the generated high-definition image is identified, with a height of , and a width of , and a width of ; (6) the discriminator network discriminates the SR image and the HR image to calculate the generative adversarial loss; The generative adversarial loss function is: wherein is a sigmoid function, is an output of the discriminator, denotes averaging, is a generated high-definition image, is a real high-definition image; During the training of the discriminator network, the overall loss function is: (7) calculating and back-propagating the total loss; During the training of the generator network, the overall loss function is composed of the pixel loss function, the content-aware loss function and the generative adversarial loss function, specifically: (8) determining whether the training is completed; if not, returning to (2) to continue training; otherwise, ending the training to obtain the trained image super-resolution reconstruction network.
6. A crop pest detection system based on an image super-resolution reconstruction network, characterized in that, The method comprises the following modules: Module 1 is configured to acquire a crop pest image; Module 2 is configured to input the crop pest image into the image super-resolution reconstruction network to perform super-resolution reconstruction on the crop pest image to obtain a high-resolution image; The image super-resolution reconstruction network is composed of a generator network and a discriminator network; The discriminator is composed of VGG19, and the generator network is composed of three parts: a head, a trunk and a reconstruction part; when a crop pest image is input, an initial feature is extracted by using a 3*3 convolution as the head, then deep feature extraction is performed by using ten consecutive hierarchical residual feature aggregation blocks to constitute the trunk, finally, a sub-pixel convolution and a 3*3 convolution are used as the reconstruction part to up-sample and decode the image feature information, and finally a high-resolution crop pest image is obtained; The hierarchical residual feature aggregation block comprises a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a fifth convolution block, a sixth convolution block, a seventh convolution block, a first 1*1 convolution fusion block, a second 1*1 convolution fusion block, a third 1*1 convolution fusion block, a fourth 1*1 convolution fusion block, a fifth 1*1 convolution fusion block, a sixth 1*1 convolution fusion block, a first residual addition operation block, a second residual addition operation block and a third residual addition operation block; The first convolution block, the second convolution block, the third convolution block, the fourth convolution block, the fifth convolution block, the sixth convolution block and the seventh convolution block are each composed of sequentially connected 3*3 convolution layers, LReLU activation layers, 3*3 convolution layers and hybrid attention blocks; The mixed attention block includes a spatial attention block, a channel attention block and a self-attention block arranged in parallel; the outputs of the spatial attention block, the channel attention block and the self-attention block and the original input are output after a matrix addition operation layer; When the crop pest image feature information is input into the hierarchical residual feature aggregation block, the crop pest image feature information sequentially passes through the first convolution block, the second convolution block, the third convolution block, the fourth convolution block, the fifth convolution block and the sixth convolution block; the crop pest image feature information and the output of the first convolution block are fused through the first 1*1 convolution fusion block; the output of the second convolution block and the output of the first 1*1 convolution fusion block are fused through the second 1*1 convolution fusion block; the output of the third convolution block and the output of the second 1*1 convolution fusion block are fused through the third 1*1 convolution fusion block; the output of the fourth convolution block and the output of the third 1*1 convolution fusion block are fused through the fourth 1*1 convolution fusion block; the output of the fifth convolution block and the output of the fourth 1*1 convolution fusion block are fused through the fifth 1*1 convolution fusion block; the output of the sixth convolution block and the output of the fifth 1*1 convolution fusion block are fused through the sixth 1*1 convolution fusion block; the outputs of the first convolution block and the second convolution block are input into the third convolution block after the first residual addition operation block, and the outputs of the fourth convolution block and the fifth convolution block are input into the sixth convolution block after the second residual addition operation block; the output of the sixth 1*1 convolution fusion block is output through the output of the seventh convolution block, and the crop pest image feature information is output after the third residual addition operation block; Module 3 is configured to input the reconstructed high-resolution image into a target detection network to detect the pest.