Detection method of diseased citrus leaves for inoculation and preservation of citrus Huanglongbing pathogen

Abstract

The present application discloses a method for detecting diseased citrus leaves inoculated with and preserved citrus Huanglongbing pathogenic bacteria, which uses a camera to collect leaf images, and uses image recognition and analysis technology based on deep learning neural networks to extract and process the features of regions of interest in the leaf images, thereby automatically obtaining a judgment result of whether symptoms of Huanglongbing appear based on the state features of multiple local significant regions of interest in the leaf images over the entire leaf area. In this way, citrus diseased leaf images can be quickly processed and detection results can be given, reducing the need for manual detection, improving detection efficiency and intelligence, and thus providing a basis for further research and prevention and control.

Description

Technical Field

[0001] The present application relates to the field of intelligent detection, and more specifically, to a method for detecting diseased citrus leaves that are inoculated and preserved with citrus Huanglongbing pathogens. Background Art

[0002] Huanglongbing, also known as citrus cancer, is a disease that is extremely destructive to the citrus industry because of its fast transmission speed and long incubation period. The pathogen of Huanglongbing is a phloem-restricted bacterium that is difficult to culture in vitro, which poses a challenge to etiological research and disease prevention and control. Therefore, it is necessary to test citrus leaves to accurately identify and confirm the pathogen of citrus Huanglongbing, providing a basis for further research and prevention and control.

[0003] However, the traditional method of detecting citrus diseased leaves is to conduct fluorescence quantitative PCR detection to accurately determine whether the sweet orange seedlings have been successfully infected with the Huanglongbing pathogen. However, this detection method takes a long time to detect, has a low sampling rate, and is difficult to achieve rapid and high-throughput detection.

[0004] Therefore, a method for detecting diseased citrus leaves for inoculating and preserving the citrus Huanglongbing pathogen is desired. Summary of the invention

[0005] In order to solve the above-mentioned technical problems, the present application is proposed. The embodiment of the present application provides a method for detecting diseased citrus leaves inoculated with and preserved citrus Huanglongbing pathogens, which collects leaf images by a camera, and uses image recognition and analysis technology based on deep learning neural networks to extract and process the features of the region of interest of the leaf images, thereby automatically obtaining a judgment result of whether the symptoms of Huanglongbing appear based on the state characteristics of multiple local significant regions of interest of the leaf image on the entire leaf area. In this way, the citrus diseased leaf images can be quickly processed and the detection results can be given, which reduces the need for manual detection, improves the detection efficiency and intelligence, and thus provides a basis for further research and prevention and control.

[0006] According to one aspect of the present application, a method for detecting diseased citrus leaves for inoculating and preserving citrus Huanglongbing pathogens is provided, comprising:

[0007] Selecting diseased citrus leaves with symptoms of Huanglongbing, and cutting the diseased citrus leaves obliquely into small pieces of 1 cm square to obtain small squares of diseased leaves with symptoms of Huanglongbing;

[0008] Select healthy and non-toxic sweet orange seedlings, and peel the main stem of the sweet orange seedlings 5-10 cm below the first leaf with a grafting knife to obtain peeled sweet orange seedlings;

[0009] Inserting the small square of diseased leaves with symptoms of Huanglongbing into the cut of the peeled sweet orange seedling, and wrapping and fixing it with Parafilm to obtain a fixed sweet orange seedling;

[0010] The fixed sweet orange seedlings are placed in a greenhouse for cultivation, and after 1-2 weeks, the Parafim is uncovered and the cultivation is continued for 2-3 weeks, and the leaves of the fixed sweet orange seedlings are observed to see whether symptoms of Huanglongbing appear, and the sweet orange seedlings that are successfully infected are selected;

[0011] The successfully infected sweet orange seedlings are placed in a greenhouse free of Huanglongbing disease for cultivation.

[0012] In the above-mentioned citrus diseased leaf detection method for inoculating and preserving citrus Huanglongbing pathogen, the fixed sweet orange seedlings are placed in a greenhouse for cultivation, and after 1-2 weeks, the Parafim is uncovered and the cultivation is continued for 2-3 weeks to observe whether the leaves of the fixed sweet orange seedlings show symptoms of Huanglongbing, and the sweet orange seedlings that are successfully infected are screened out, including: obtaining leaf images captured by a camera; performing local region of interest feature extraction on the leaf images to obtain a plurality of leaf local region of interest feature maps; inputting each of the plurality of leaf local region of interest feature maps into a foreground region transfer network based on a spatial attention map to obtain a plurality of foreground salient leaf local region of interest feature maps; and extracting the plurality of foreground salient leaf local region of interest feature maps from the foreground region transfer network based on a spatial attention map. Each foreground-significant leaf local region of interest feature map in the foreground-significant leaf local region of interest feature map is respectively input into a multi-scale semantic feature extraction module to obtain a plurality of foreground-significant leaf local region of interest feature vectors; the plurality of foreground-significant leaf local region of interest feature vectors are input into a feature optimizer based on semantic association and semantic span to obtain a plurality of optimized foreground-significant leaf local region of interest feature vectors; the position-wise mean vectors of the plurality of optimized foreground-significant leaf local region of interest feature vectors are calculated to obtain a leaf global state feature representation vector as a leaf global state feature representation; and, based on the leaf global state feature representation, a detection result is obtained, and the detection result is used to indicate whether the symptoms of Huanglongbing appear.

[0013] In the above-mentioned method for detecting diseased citrus leaves inoculated and preserved with the citrus Huanglongbing pathogen, local region of interest feature extraction is performed on the leaf image to obtain multiple leaf local region of interest feature maps, including: passing the leaf image through a region of interest feature extractor based on an R-CNN network to obtain the multiple leaf local region of interest feature maps.

[0014] In the above-mentioned citrus diseased leaf detection method for inoculating and preserving citrus Huanglongbing pathogen, each of the multiple leaf local region of interest feature maps is respectively input into a foreground region transfer network based on a spatial attention map to obtain multiple foreground salient leaf local region of interest feature maps, including: performing convolution encoding and nonlinear activation processing on the leaf local region of interest feature map to obtain a spatially adaptive leaf local region of interest weight information feature map; along the channel dimension of the spatially adaptive leaf local region of interest weight information feature map, calculating the cross-channel global mean of each pixel position in the spatially adaptive leaf local region of interest weight information feature map to obtain a spatially adaptive leaf local region of interest weight information representation matrix matrix; in response to each position feature value in the spatial adaptive leaf local region of interest weight information representation matrix being greater than or equal to a predetermined threshold, taking the original value of the position feature value, otherwise setting it to zero to obtain a masked spatial adaptive leaf local region of interest weight information representation matrix; upsampling the masked spatial adaptive leaf local region of interest weight information representation matrix to obtain a dimension-aligned spatial adaptive leaf local region of interest weight information representation matrix; and using the dimension-aligned spatial adaptive leaf local region of interest weight information representation matrix as a weight matrix, respectively calculating the product between it and each feature matrix along the channel dimension of the leaf local region of interest feature map to obtain the foreground-salient leaf local region of interest feature map.

[0015] In the above-mentioned citrus diseased leaf detection method for inoculating and preserving citrus Huanglongbing pathogen, each of the multiple foreground significant leaf local region of interest feature maps is respectively input into a multi-scale semantic feature extraction module to obtain multiple foreground significant leaf local region of interest feature vectors, including: upsampling the foreground significant leaf local region of interest feature map to obtain an upsampled foreground significant leaf local region of interest feature map; performing global mean pooling processing on each feature matrix along the channel dimension in the upsampled foreground significant leaf local region of interest feature map to obtain an upsampled foreground significant leaf local region of interest semantic channel feature vector; and performing global mean pooling processing on the upsampled foreground significant leaf local region of interest semantic channel feature vector. The invention relates to a method for obtaining a foreground-significant leaf local region of interest by performing point convolution coding on a semantic channel feature vector of a local region of interest of a leaf to obtain a global semantic feature vector of a local region of interest of a foreground-significant leaf; performing point convolution coding on the upsampled foreground-significant leaf local region of interest feature map to obtain a channel-modulated upsampled foreground-significant leaf local region of interest feature map; performing two-dimensional convolution coding on the channel-modulated upsampled foreground-significant leaf local region of interest feature map to obtain a local semantic feature vector of a local region of interest of a foreground-significant leaf; and fusing the global semantic feature vector of the local region of interest of a foreground-significant leaf and the local semantic feature vector of the local region of interest of a foreground-significant leaf to obtain a feature vector of the local region of interest of a foreground-significant leaf.

[0016] In the above-mentioned citrus diseased leaf detection method for inoculating and preserving citrus Huanglongbing pathogen, the multiple foreground significant leaf local region of interest feature vectors are input into a feature optimizer based on semantic association and semantic span to obtain multiple optimized foreground significant leaf local region of interest feature vectors, including: extracting a predetermined foreground significant leaf local region of interest feature vector from the multiple foreground significant leaf local region of interest feature vectors; calculating the semantic association between the predetermined foreground significant leaf local region of interest feature vector and all other foreground significant leaf local region of interest feature vectors to obtain a sequence of foreground significant leaf local region of interest semantic associations; calculating the global mean of the sequence of foreground significant leaf local region of interest semantic associations to obtain a foreground significant leaf local region of interest semantic global average association; and, using the foreground significant leaf local region of interest semantic global average association as a weight, weighted optimization is performed on the predetermined foreground significant leaf local region of interest feature vector to obtain an optimized foreground significant leaf local region of interest feature vector corresponding to the predetermined foreground significant leaf local region of interest feature vector.

[0017] In the above-mentioned citrus diseased leaf detection method for inoculating and preserving citrus Huanglongbing pathogen, the semantic association between the predetermined foreground significant leaf local region of interest feature vector and all other foreground significant leaf local region of interest feature vectors is calculated to obtain a sequence of foreground significant leaf local region of interest semantic associations, including: calculating the semantic similarity between the predetermined foreground significant leaf local region of interest feature vector and all other foreground significant leaf local region of interest feature vectors to obtain a sequence of foreground significant leaf local region of interest semantic similarities; calculating the number of feature vectors between the predetermined foreground significant leaf local region of interest feature vector and all other foreground significant leaf local region of interest feature vectors to obtain a sequence of foreground significant leaf local region of interest distance meter numbers; and dividing the sequence of foreground significant leaf local region of interest semantic similarities by the sequence of foreground significant leaf local region of interest distance meter numbers by position to obtain a sequence of foreground significant leaf local region of interest semantic associations.

[0018] In the above-mentioned method for detecting diseased citrus leaves inoculated and preserved with the citrus Huanglongbing pathogen, a detection result is obtained based on the leaf global state feature representation, including: passing the leaf global state feature representation vector through a classifier-based Huanglongbing detection module to obtain the detection result.

[0019] In the above-mentioned method for detecting citrus diseased leaves inoculated and preserved with citrus Huanglongbing pathogens, a training step is also included: for training the R-CNN network-based region of interest feature extractor, the spatial attention map-based foreground region transfer network, the multi-scale semantic feature extraction module, the feature optimizer based on semantic association and semantic span, and the classifier-based Huanglongbing detection module.

[0020] In the above-mentioned citrus diseased leaf detection method for inoculating and preserving citrus Huanglongbing pathogen, the training step includes: obtaining training data, the training data including training leaf images collected by a camera, and real detection results, the real detection results being whether symptoms of Huanglongbing appear; passing the training leaf images through the R-CNN network-based region of interest feature extractor to obtain a plurality of training leaf local region of interest feature maps; inputting each of the multiple training leaf local region of interest feature maps into the spatial attention map-based foreground region transfer network to obtain a plurality of training foreground significant leaf local region of interest feature maps; inputting each of the multiple training foreground significant leaf local region of interest feature maps into the multi-scale semantic feature extraction module to obtain a plurality of training foreground significant leaf local region of interest feature vectors; and inputting each of the multiple training foreground significant leaf local region of interest feature maps into the multi-scale semantic feature extraction module to obtain a plurality of training foreground significant leaf local region of interest feature vectors. The region of interest feature vector is input into the feature optimizer based on semantic association and semantic span to obtain multiple training optimized foreground salient leaf local region of interest feature vectors; the positional mean vector of the multiple training optimized foreground salient leaf local region of interest feature vectors is calculated to obtain a training leaf global state feature representation vector; the training leaf global state feature representation vector is passed through the classifier-based Huanglongbing detection module to obtain a classification loss function value; based on the training leaf global state feature representation vector, the leaf global state loss function value is calculated; and, the weighted sum of the classification loss function value and the leaf global state loss function value is used as the final loss function value and the R-CNN network-based region of interest feature extractor, the spatial attention map-based foreground region transfer network, the multi-scale semantic feature extraction module, the feature optimizer based on semantic association and semantic span, and the classifier-based Huanglongbing detection module are trained through gradient descent back propagation.

[0021] Compared with the prior art, the present application provides a method for detecting citrus diseased leaves inoculated with and preserved with the pathogen of citrus Huanglongbing. The method collects leaf images through a camera, and uses image recognition and analysis technology based on deep learning neural networks to extract and process the features of the regions of interest of the leaf images, thereby automatically obtaining the judgment result of whether the symptoms of Huanglongbing appear based on the state characteristics of multiple local significant regions of interest of the leaf images on the entire leaf area. In this way, the citrus diseased leaf images can be quickly processed and the detection results can be given, which reduces the need for manual detection, improves the detection efficiency and intelligence, and thus provides a basis for further research and prevention and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0023] Figure 1 The present invention is a flow chart of a method for detecting diseased citrus leaves inoculated with and preserved by the citrus Huanglongbing pathogen according to an embodiment of the present application.

[0024] Figure 2 The present invention is a flow chart of observing whether the leaves of the fixed sweet orange seedlings show symptoms of Huanglongbing in the citrus diseased leaf detection method for inoculating and preserving the citrus Huanglongbing pathogen according to an embodiment of the present application.

[0025] Figure 3 This is a schematic diagram of the structure of observing whether the leaves of the fixed sweet orange seedlings show symptoms of Huanglongbing in the citrus diseased leaf detection method for inoculating and preserving the citrus Huanglongbing pathogen according to an embodiment of the present application.

[0026] Figure 4 The present invention is a flowchart for training the R-CNN network-based region of interest feature extractor, the spatial attention map-based foreground region transfer network, the multi-scale semantic feature extraction module, the feature optimizer based on semantic association and semantic span, and the classifier-based Huanglongbing detection module in the mango trunk phloem damage repair method according to an embodiment of the present application. DETAILED DESCRIPTION

[0027] The embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein, but rather these embodiments are provided for a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0028] It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0029] In the description of the embodiments of the present disclosure, the term "including" and similar terms should be understood as open inclusion, that is, "including but not limited to". The term "based on" should be understood as "based at least in part on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc. may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0030] It should be noted that the modifications of "one" and "plurality" mentioned in the present disclosure are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0031] Huanglongbing, also known as citrus cancer, is a disease that is extremely destructive to the citrus industry because of its fast transmission speed and long incubation period. The pathogen of Huanglongbing is a phloem-restricted bacterium that is difficult to culture in vitro, which poses a challenge to etiological research and disease prevention and control. Therefore, it is necessary to test citrus leaves to accurately identify and confirm the pathogen of citrus Huanglongbing, providing a basis for further research and prevention and control.

[0032] Based on this, the present application proposes a citrus diseased leaf detection method for inoculating and preserving citrus Huanglongbing pathogens, which comprises: selecting citrus diseased leaves with Huanglongbing symptoms, and obliquely cutting the citrus diseased leaves into small pieces of 1 cm square to obtain small squares of diseased leaves containing Huanglongbing symptoms; selecting healthy and non-toxic sweet orange seedlings, and using a grafting knife to peel the outer skin on the main stem of the sweet orange seedling 5-10 cm below the first leaf to obtain peeled sweet orange seedlings; inserting the small squares of diseased leaves containing Huanglongbing symptoms into the incision of the peeled sweet orange seedlings, and wrapping and fixing them with Parafilm to obtain fixed sweet orange seedlings; placing the fixed sweet orange seedlings in a greenhouse for cultivation, uncovering the Parafilm after 1-2 weeks and continuing to cultivate for 2-3 weeks, observing whether the leaves of the fixed sweet orange seedlings show symptoms of Huanglongbing, and screening out the successfully infected sweet orange seedlings; and placing the successfully infected sweet orange seedlings in a greenhouse without Huanglongbing for cultivation.

[0033] However, the traditional method of detecting citrus diseased leaves is to conduct fluorescence quantitative PCR detection to accurately determine whether the sweet orange seedlings have been successfully infected with the Huanglongbing pathogen. However, this detection method takes a long time to detect, has a low sampling rate, and is difficult to achieve rapid and high-throughput detection.

[0034] Therefore, in response to the above technical problems, the technical concept of this application is to collect leaf images through a camera, and use image recognition and analysis technology based on deep learning neural networks to extract and process the features of the region of interest of the leaf images, so as to automatically obtain the judgment result of whether the symptoms of Huanglongbing appear based on the state characteristics of multiple local significant regions of interest of the leaf images on the entire leaf area. In this way, the citrus diseased leaf images can be quickly processed and the detection results can be given, which reduces the need for manual detection, improves the detection efficiency and intelligence, and thus provides a basis for further research and prevention and control.

[0035] Figure 1 Flow chart of the method for detecting diseased citrus leaves inoculated with and preserved with the pathogen of citrus Huanglongbing according to an embodiment of the present application. Figure 1 As shown, according to the embodiment of the present application, the citrus diseased leaf detection method for inoculating and preserving the citrus Huanglongbing pathogen comprises: S110, selecting citrus diseased leaves with Huanglongbing symptoms, and obliquely cutting the citrus diseased leaves into small blocks of 1 cm square to obtain diseased leaf small blocks containing Huanglongbing symptoms; S120, selecting healthy and non-toxic sweet orange seedlings, and using a grafting knife to peel the outer skin of the main stem of the sweet orange seedling 5-10 cm below the first leaf to obtain a peeled sweet orange seedling; S130, inserting the diseased leaf small blocks containing Huanglongbing symptoms into the incision of the peeled sweet orange seedling, and using Parafilm Wrapping and fixing to obtain fixed sweet orange seedlings; S140, placing the fixed sweet orange seedlings in a greenhouse for cultivation, uncovering the Parafim after 1-2 weeks and continuing to cultivate for 2-3 weeks, observing whether the leaves of the fixed sweet orange seedlings have symptoms of Huanglongbing, and screening the sweet orange seedlings that are successfully infected; and, S150, placing the successfully infected sweet orange seedlings in a greenhouse without Huanglongbing for cultivation.

[0036] Figure 2 The present invention is a flow chart of observing whether the leaves of the fixed sweet orange seedlings show symptoms of Huanglongbing in the citrus diseased leaf detection method for inoculating and preserving the citrus Huanglongbing pathogen according to an embodiment of the present application. Figure 3 This is a schematic diagram of the structure of observing whether the leaves of the fixed sweet orange seedlings have symptoms of Huanglongbing in the citrus diseased leaf detection method for inoculating and preserving the citrus Huanglongbing pathogen according to the embodiment of the present application. Figure 2 and Figure 3As shown, the fixed sweet orange seedlings are placed in a greenhouse for cultivation, the Parafim is uncovered after 1-2 weeks and the cultivation is continued for 2-3 weeks, the leaves of the fixed sweet orange seedlings are observed to see if there are symptoms of Huanglongbing, and the sweet orange seedlings that are successfully infected are screened out, including: S210, acquiring a leaf image captured by a camera; S220, extracting local regions of interest features from the leaf image to obtain a plurality of leaf local regions of interest feature maps; S230, inputting each of the plurality of leaf local regions of interest feature maps into a foreground region transfer network based on a spatial attention map to obtain a plurality of foreground salient leaf local regions of interest feature maps; S240, extracting the plurality of foreground salient leaf local regions of interest feature maps from the plurality of leaf local regions of interest feature maps. Each foreground-significant leaf local region of interest feature map in the region of interest feature map is respectively input into a multi-scale semantic feature extraction module to obtain a plurality of foreground-significant leaf local region of interest feature vectors; S250, the plurality of foreground-significant leaf local region of interest feature vectors are input into a feature optimizer based on semantic association and semantic span to obtain a plurality of optimized foreground-significant leaf local region of interest feature vectors; S260, the position-wise mean vectors of the plurality of optimized foreground-significant leaf local region of interest feature vectors are calculated to obtain a leaf global state feature representation vector as a leaf global state feature representation; and, S270, based on the leaf global state feature representation, a detection result is obtained, and the detection result is used to indicate whether the symptoms of Huanglongbing appear.

[0037] In step S210, a leaf image captured by a camera is obtained. It should be understood that the leaf image can observe visible symptoms of the leaf, such as yellowing, mottled or other typical lesions caused by Huanglongbing. Therefore, in the technical solution of the present application, by acquiring the leaf image captured by the camera, rapid and automated diagnosis can be achieved to improve the efficiency and accuracy of disease detection.

[0038] In step S220, the local region of interest feature extraction is performed on the leaf image to obtain a plurality of leaf local region of interest feature maps. Specifically, in an embodiment of the present application, the local region of interest feature extraction is performed on the leaf image to obtain a plurality of leaf local region of interest feature maps, including: passing the leaf image through a region of interest feature extractor based on an R-CNN network to obtain the plurality of leaf local region of interest feature maps. Accordingly, considering that the surface features of the leaves shown in different regions of the leaf image are different, for example, some local regions are in a normal state, and some local regions have yellowing symptoms, etc. The R-CNN network can accurately identify different regions in the image and generate corresponding category labels and bounding boxes for each region, and this accuracy is very important for detecting the local region of interest features on the leaf, and can help identify the specific location of the diseased leaf. Based on this, in the technical solution of the present application, the leaf image is passed through a region of interest feature extractor based on an R-CNN network to capture and extract regions of interest of different sizes and shapes in the leaf image, thereby obtaining a plurality of leaf local region of interest feature maps.

[0039] In step S230, each of the plurality of leaf local region of interest feature maps is input into a foreground region transfer network based on a spatial attention map to obtain a plurality of foreground salient leaf local region of interest feature maps. It should be understood that, considering that each of the plurality of leaf local region of interest feature maps represents the surface feature information of the leaf reflected by a local region, wherein the surface feature information of the leaf contains the foreground features of the disease of Huanglongbing, but also contains background feature information that is not related to the disease detection. Therefore, in order to focus more on the significant foreground features in the plurality of leaf local region of interest feature maps, in the technical solution of the present application, each of the plurality of leaf local region of interest feature maps is input into a foreground region transfer network based on a spatial attention map to obtain a plurality of foreground salient leaf local region of interest feature maps. Specifically, the foreground region transfer network based on a spatial attention map first performs convolutional coding and nonlinear activation on the leaf local region of interest feature map to extract high-level abstract leaf local region of interest spatial features, so that the content and structure in the local region of the leaf image can be better understood to obtain spatial adaptive features. Next, the spatial adaptive features are subjected to cross-channel global mean calculation, thereby introducing global contextual information, which helps the features better understand the background and structure of the entire local leaf image to obtain multi-channel weight information. The multi-channel weight information is then masked and upsampled in turn, making the local features of interest on the leaves more discriminative and expressive, while highlighting the important and significant features in the leaf image. Finally, the features obtained after the upsampling process are used as weights to perform feature weighting on the feature map of the local region of interest on the leaves, so as to enhance the features of the local region of interest on the leaves, making the foreground features of the diseased area more prominent and clear, thereby better supporting the detection and diagnosis of citrus leaf Huanglongbing.

[0040] Specifically, in an embodiment of the present application, each of the plurality of leaf local region of interest feature maps is respectively input into a foreground region transfer network based on a spatial attention map to obtain a plurality of foreground salient leaf local region of interest feature maps, including: performing convolution encoding and nonlinear activation processing on the leaf local region of interest feature map to obtain a spatially adaptive leaf local region of interest weight information feature map; along the channel dimension of the spatially adaptive leaf local region of interest weight information feature map, calculating the cross-channel global mean of each pixel position in the spatially adaptive leaf local region of interest weight information feature map to obtain a spatially adaptive leaf local region of interest weight information representation matrix; in response to the If each position feature value in the spatially adaptive leaf local region of interest weight information representation matrix is ​​greater than or equal to a predetermined threshold, the original value of the position feature value is taken, otherwise it is set to zero to obtain a masked spatially adaptive leaf local region of interest weight information representation matrix; the masked spatially adaptive leaf local region of interest weight information representation matrix is ​​upsampled to obtain a dimensionally aligned spatially adaptive leaf local region of interest weight information representation matrix; and the dimensionally aligned spatially adaptive leaf local region of interest weight information representation matrix is ​​used as a weight matrix, and the product between it and each feature matrix along the channel dimension of the leaf local region of interest feature map is calculated respectively to obtain the foreground salient leaf local region of interest feature map.

[0041] In an embodiment of the present application, specifically, each of the plurality of leaf local region of interest feature maps is input into a foreground region transfer network based on a spatial attention map to obtain a plurality of foreground salient leaf local region of interest feature maps, including: using the foreground region transfer network based on the spatial attention map to process the leaf local region of interest feature map with the following salient formula to obtain the foreground salient leaf local region of interest feature map; wherein the salient formula is:

[0042]

[0043]

[0044]

[0045]

[0046]

[0047] in, represents the feature map of the local region of interest of the leaf, Indicates that the feature map is based on Convolution processing of convolution kernel, yes Activation function, is the weight information feature map of the local area of ​​interest of the spatial adaptive leaf, represents the number of channels of the spatial adaptive leaf local region of interest weight information feature map, represents the weight information feature map of the local area of ​​interest of the spatial adaptive leaf Eigenvalues ​​of positions is the weight information representation matrix of the local area of ​​interest of the spatial adaptive leaf The eigenvalues ​​of the positions, is the weight information representation matrix of the local area of ​​interest of the spatial adaptive leaf, Indicates masking processing. is the predetermined threshold, represents upsampling processing, represents the dimensionally aligned spatially adaptive leaf local region of interest weight information representation matrix, is the local region of interest feature map of the foreground salient leaf, Represents positional multiplication.

[0048] In step S240, each of the plurality of foreground-significant leaf local region of interest feature maps is input into a multi-scale semantic feature extraction module to obtain a plurality of foreground-significant leaf local region of interest feature vectors. Accordingly, it is considered that each of the plurality of foreground-significant leaf local region of interest feature maps has overall background information based on the entire local region and local detail features. For example, the distribution pattern of the disease on the leaves in the entire local region of interest, such as random distribution; the microscopic detail structure of the disease on the upper surface of the leaves in the local region of interest, such as the texture features of spots and patches. Based on this, in order to more comprehensively and meticulously capture the feature information of different scales implied in each foreground-significant leaf local region of interest feature map, in the technical solution of the present application, each of the plurality of foreground-significant leaf local region of interest feature maps is input into a multi-scale semantic feature extraction module for multi-scale semantic feature extraction to obtain a plurality of foreground-significant leaf local region of interest feature vectors. That is to say, the multi-scale semantic feature extraction module captures comprehensive information from the overall to the details of the feature map of the local region of interest of each foreground significant leaf, making the features of the local region of interest of the leaf richer and the feature expression ability stronger, which helps to more comprehensively understand and identify diseased leaves and improve the discriminability and accuracy of subsequent lesion features.

[0049] Specifically, in an embodiment of the present application, each foreground significant leaf local area of ​​interest feature map in the multiple foreground significant leaf local area of ​​interest feature maps is respectively input into a multi-scale semantic feature extraction module to obtain multiple foreground significant leaf local area of ​​interest feature vectors, including: upsampling the foreground significant leaf local area of ​​interest feature map to obtain an upsampled foreground significant leaf local area of ​​interest feature map; performing global mean pooling processing on each feature matrix along the channel dimension in the upsampled foreground significant leaf local area of ​​interest feature map to obtain an upsampled foreground significant leaf local area of ​​interest semantic channel feature vector; performing global mean pooling processing on each feature matrix along the channel dimension in the upsampled foreground significant leaf local area of ​​interest feature map to obtain an upsampled foreground significant leaf local area of ​​interest semantic channel feature vector; performing global mean pooling processing on the .... The regional semantic channel feature vector is point convolutionally encoded to obtain a global semantic feature vector of a foreground-significant leaf local region of interest; the upsampled foreground-significant leaf local region of interest feature map is point convolutionally encoded to obtain a channel-modulated upsampled foreground-significant leaf local region of interest feature map; the channel-modulated upsampled foreground-significant leaf local region of interest feature map is two-dimensionally convolutionally encoded to obtain a local semantic feature vector of a foreground-significant leaf local region of interest; and the global semantic feature vector of the foreground-significant leaf local region of interest and the local semantic feature vector of the foreground-significant leaf local region of interest are fused to obtain the feature vector of the foreground-significant leaf local region of interest.

[0050] In step S250, the multiple foreground-significant leaf local region of interest feature vectors are input into a feature optimizer based on semantic association and semantic span to obtain multiple optimized foreground-significant leaf local region of interest feature vectors. It should be understood that, considering that there are mutually similar correlation influence relationships between different regional spans between the different foreground-significant leaf local region of interest feature vectors, for example, certain disease symptoms may appear simultaneously in different regions of the leaf. Based on this, in order to more accurately improve the association between features of different regions to improve the performance in the citrus leaf Huanglongbing detection task, in the technical solution of the present application, the multiple foreground-significant leaf local region of interest feature vectors are input into a feature optimizer based on semantic association and semantic span to obtain multiple optimized foreground-significant leaf local region of interest feature vectors. It should be understood that the feature optimizer based on semantic association and semantic span makes the features more consistent and correlated by respectively calculating the feature similarity and feature distance between each two feature vectors of the foreground-significant leaf local regions of interest, so as to better capture the semantic information between different local regions in the leaf image, and performs global average calculation on the feature sequence obtained by dividing the feature similarity and feature distance by position to integrate the local feature information of leaves of different positions to obtain an overall feature representation, and finally uses the obtained overall feature to perform weighted optimization on the feature vector of each foreground-significant leaf local region of interest to help enhance the expression performance of the features of each foreground-significant leaf region, thereby more accurately distinguishing the foreground-significant leaf features of different regions and improving the performance in the citrus leaf Huanglongbing detection task.

[0051] Specifically, in an embodiment of the present application, the multiple foreground significant leaf local region of interest feature vectors are input into a feature optimizer based on semantic association and semantic span to obtain multiple optimized foreground significant leaf local region of interest feature vectors, including: extracting a predetermined foreground significant leaf local region of interest feature vector from the multiple foreground significant leaf local region of interest feature vectors; calculating the semantic association between the predetermined foreground significant leaf local region of interest feature vector and all other foreground significant leaf local region of interest feature vectors to obtain a sequence of foreground significant leaf local region of interest semantic associations; calculating the global mean of the sequence of foreground significant leaf local region of interest semantic associations to obtain a global average semantic association of foreground significant leaf local region of interest; and, using the global average semantic association of foreground significant leaf local region of interest as a weight, weighted optimization is performed on the predetermined foreground significant leaf local region of interest feature vector to obtain an optimized foreground significant leaf local region of interest feature vector corresponding to the predetermined foreground significant leaf local region of interest feature vector.

[0052] More specifically, in an embodiment of the present application, the semantic association between the predetermined foreground significant leaf local region of interest feature vector and all other foreground significant leaf local region of interest feature vectors is calculated to obtain a sequence of foreground significant leaf local region of interest semantic associations, including: calculating the semantic similarity between the predetermined foreground significant leaf local region of interest feature vector and all other foreground significant leaf local region of interest feature vectors to obtain a sequence of foreground significant leaf local region of interest semantic similarities; calculating the number of feature vectors between the predetermined foreground significant leaf local region of interest feature vector and all other foreground significant leaf local region of interest feature vectors to obtain a sequence of foreground significant leaf local region of interest distance meter numbers; and dividing the sequence of foreground significant leaf local region of interest semantic similarities by the sequence of foreground significant leaf local region of interest distance meter numbers by position to obtain a sequence of foreground significant leaf local region of interest semantic associations.

[0053] In an embodiment of the present application, specifically, the multiple foreground significant leaf local interest region feature vectors are input into a feature optimizer based on semantic association and semantic span to obtain multiple optimized foreground significant leaf local interest region feature vectors, including: inputting the multiple foreground significant leaf local interest region feature vectors into the feature optimizer based on semantic association and semantic span, and processing them with the following optimization formula to obtain the multiple optimized foreground significant leaf local interest region feature vectors; wherein the optimization formula is:

[0054]

[0055] in, and respectively represent the first of the plurality of foreground salient leaf local interest region feature vectors and The feature vector of the local region of interest of the foreground salient leaf, express and The number of eigenvectors between express and The similarity between represents the total number of feature vectors in the plurality of foreground salient leaf local region of interest feature vectors minus one, and The values ​​of are equal, represents the first of the plurality of optimized foreground salient leaf local region of interest feature vectors An optimized foreground saliency leaf local region of interest feature vector.

[0056] In step S260, the positional mean vectors of the feature vectors of the multiple optimized foreground-significant leaf local regions of interest are calculated to obtain a leaf global state feature representation vector as a leaf global state feature representation. Accordingly, in order to integrate the feature information of each optimized foreground-significant leaf local regions of interest, to comprehensively consider the local detail features of the entire leaf image, and thus better capture the global state features of the leaf, in the technical solution of the present application, the positional mean vectors of the feature vectors of the multiple optimized foreground-significant leaf local regions of interest are calculated to obtain a leaf global state feature representation vector, so that the features of each local region of interest can be integrated to provide a comprehensive leaf state representation, which helps to capture the overall disease condition of the leaf and provides a basis and support for the subsequent detection of leaf Huanglongbing.

[0057] In step S270, based on the global state feature representation of the leaves, a detection result is obtained, and the detection result is used to indicate whether symptoms of Huanglongbing appear. Specifically, in an embodiment of the present application, based on the global state feature representation of the leaves, a detection result is obtained, including: passing the global state feature representation vector of the leaves through a Huanglongbing detection module based on a classifier to obtain the detection result. That is, the global state feature representation of the leaves obtained by position mean using the multiple optimized foreground-significant leaf local region of interest feature vectors is classified and processed to automatically obtain a judgment result on whether symptoms of Huanglongbing appear. In this way, citrus diseased leaf images can be quickly processed and detection results can be given, reducing the need for manual detection, improving detection efficiency and intelligence, and thus providing a basis for further research and prevention and control.

[0058] It is worth mentioning that those skilled in the art should be aware that before applying a deep neural network model for inference, the deep neural network model must be trained so that the deep neural network can implement specific functional capabilities.

[0059] Specifically, in an embodiment of the present application, a training step is also included: for training the R-CNN network-based region of interest feature extractor, the spatial attention map-based foreground region transfer network, the multi-scale semantic feature extraction module, the feature optimizer based on semantic association and semantic span, and the classifier-based Huanglongbing detection module.

[0060] Figure 4The present invention is a flowchart of training the region of interest feature extractor based on the R-CNN network, the foreground region transfer network based on the spatial attention map, the multi-scale semantic feature extraction module, the feature optimizer based on semantic association and semantic span, and the Huanglongbing detection module based on the classifier in the mango trunk phloem damage repair method according to the embodiment of the present application. Figure 4 As shown, the training step includes: S310, obtaining training data, the training data including training leaf images collected by a camera, and real detection results, the real detection results being whether symptoms of Huanglongbing appear; S320, passing the training leaf images through the region of interest feature extractor based on the R-CNN network to obtain multiple training leaf local region of interest feature maps; S330, inputting each of the multiple training leaf local region of interest feature maps into the foreground region transfer network based on the spatial attention map to obtain multiple training foreground significant leaf local region of interest feature maps; S340, inputting each of the multiple training foreground significant leaf local region of interest feature maps into the multi-scale semantic feature extraction module to obtain multiple training foreground significant leaf local region of interest feature vectors; S350, inputting the multiple training foreground significant leaf local region of interest feature vectors into the multi-scale semantic feature extraction module to obtain multiple training foreground significant leaf local region of interest feature vectors; S360, calculate the positional mean vector of the multiple training optimized foreground-significant leaf local region of interest feature vectors to obtain the training leaf global state feature representation vector; S370, pass the training leaf global state feature representation vector through the classifier-based Huanglongbing detection module to obtain the classification loss function value; S380, calculate the leaf global state loss function value based on the training leaf global state feature representation vector; and, S390, use the weighted sum of the classification loss function value and the leaf global state loss function value as the final loss function value and train the R-CNN network-based region of interest feature extractor, the spatial attention map-based foreground region transfer network, the multi-scale semantic feature extraction module, the feature optimizer based on semantic association and semantic span, and the classifier-based Huanglongbing detection module through gradient descent back propagation.

[0061] It should be understood that in the embodiment of the present application, the multiple training leaf local interest region feature maps represent the image semantic features of the training leaf image in the local image semantic space domain. After spatial attention, multi-scale semantic feature extraction and feature optimization based on semantic association and semantic span, the differences in the distribution of image semantic features in each local image semantic space domain will become larger. The training leaf global state feature representation vector obtained by calculating the positional mean vector of the feature vectors of the local interest regions of the multiple training optimized foreground significant leaves will also have the image semantic distribution knowledge complexity corresponding to the image semantic feature distribution difference in the local image semantic space domain. Therefore, it is expected to enhance the accuracy of the classification result by targeting the complex feature distribution expression characteristics of the training leaf global state feature representation vector.

[0062] Based on this, the present application further introduces a loss function for promoting classification understanding based on the complex feature expression of the training blade global state feature representation vector, that is, based on the training blade global state feature representation vector, the blade global state loss function value is calculated. Specifically, in an embodiment of the present application, based on the training blade global state feature representation vector, the blade global state loss function value is calculated, including: first, calculating the mean and variance between the eigenvalues ​​of any two positions of the blade global state feature representation vector to obtain a mean weight matrix and a variance weight matrix, then query-multiplying the blade global state feature representation vector with the mean weight matrix and the variance weight matrix respectively to obtain a mean eigenvector and a variance eigenvector, and then calculating the autocorrelation matrix of the obtained mean eigenvector and variance eigenvector. norm, and subtract the autocorrelation matrix of the leaf global state feature representation vector The value of the leaf global state loss function is obtained by multiplying the norm with the weight as a hyperparameter.

[0063] For example, represented as:

[0064]

[0065] in, represents the global state feature representation vector of the training leaf, is the mean weight matrix, is the variance weight matrix, represents the transpose of a vector, represents the matrix multiplication operation, is the weight as a hyperparameter, Represents the matrix norm, represents the global state loss function value of the leaf, and is as described above:

[0066]

[0067] in, The first eigenvalues, The first eigenvalues, represents the mean weight matrix The eigenvalues ​​of the positions, represents the variance weight matrix The eigenvalue of the location.

[0068] That is, the leaf global state loss function value, based on determining the overall distributed structural bottleneck of the training leaf global state feature representation vector through the multi-dimensional interaction of the detail group of the training leaf global state feature representation vector, simulates the correlation between the structural details of the training leaf global state feature representation vector and the macro-structured feature representation behavior through the low-rank structured reasoning of the training leaf global state feature representation vector. In this way, by training the classifier with the leaf global state loss function value, it is possible to model the structural detail dependency relative to the macro-complexity in the classification scenario, thereby promoting the understanding of the macro-complex feature expression based on the structural details of the training leaf global state feature representation vector and improving the accuracy of the classification results. In this way, citrus diseased leaf images can be quickly processed and detection results can be given, reducing the need for manual detection, improving detection efficiency and intelligence, and thus providing a basis for further research and prevention and control.

[0069] In summary, the citrus diseased leaf detection method for inoculating and preserving the citrus Huanglongbing pathogen based on the embodiment of the present application is explained, which collects leaf images by a camera, and uses image recognition and analysis technology based on deep learning neural networks to extract and process the features of the region of interest of the leaf image, so as to automatically obtain the judgment result of whether the symptoms of Huanglongbing appear based on the state characteristics of multiple local significant regions of interest of the leaf image on the entire leaf area. In this way, the citrus diseased leaf images can be quickly processed and the detection results can be given, which reduces the need for manual detection, improves the detection efficiency and intelligence, and thus provides a basis for further research and prevention and control.

[0070] The foregoing is merely an example of the principles of the present disclosure, and various modifications may be made by those skilled in the art without departing from the scope of the present disclosure. The above embodiments are presented for the purpose of illustration and not limitation. The present disclosure may also take many forms other than those explicitly described herein. Therefore, it is emphasized that the present disclosure is not limited to the methods, systems and devices explicitly disclosed, but is intended to include variations and modifications within the spirit and scope of the appended claims.

Claims

1. A method for detecting diseased citrus leaves inoculated with and stored with the pathogen of citrus Huanglongbing, characterized in that: include: Selecting diseased citrus leaves with symptoms of Huanglongbing, and cutting the diseased citrus leaves obliquely into small pieces of 1 cm square to obtain small squares of diseased leaves with symptoms of Huanglongbing; Select healthy and non-toxic sweet orange seedlings, and peel the main stem of the sweet orange seedlings 5-10 cm below the first leaf with a grafting knife to obtain peeled sweet orange seedlings; Inserting the small square of diseased leaves with symptoms of Huanglongbing into the cut of the peeled sweet orange seedling, and wrapping and fixing it with Parafilm to obtain a fixed sweet orange seedling; The fixed sweet orange seedlings are placed in a greenhouse for cultivation, and after 1-2 weeks, the Parafilm is uncovered and the cultivation is continued for 2-3 weeks, and the leaves of the fixed sweet orange seedlings are observed to see whether symptoms of Huanglongbing appear, and the sweet orange seedlings that are successfully infected are selected; The successfully infected sweet orange seedlings are placed in a greenhouse free of Huanglongbing disease for cultivation; Observe whether the leaves of the fixed sweet orange seedlings show symptoms of Huanglongbing, including: Acquire a blade image captured by a camera; Extracting local regions of interest from the leaf image to obtain a plurality of leaf local regions of interest feature maps; Inputting each of the plurality of leaf local region of interest feature maps into a foreground region transfer network based on a spatial attention map to obtain a plurality of foreground salient leaf local region of interest feature maps; Inputting each of the plurality of foreground salient leaf local region of interest feature maps into a multi-scale semantic feature extraction module to obtain a plurality of foreground salient leaf local region of interest feature vectors; Inputting the plurality of foreground salient leaf local interest region feature vectors into a feature optimizer based on semantic association and semantic span to obtain a plurality of optimized foreground salient leaf local interest region feature vectors; Calculating the position mean vector of the plurality of optimized foreground-salient leaf local regions of interest feature vectors to obtain a leaf global state feature representation vector as the leaf global state feature representation; Based on the leaf global state feature representation, a detection result is obtained, wherein the detection result is used to indicate whether symptoms of Huanglongbing appear; Performing local region of interest feature extraction on the leaf image to obtain a plurality of leaf local region of interest feature maps, including: passing the leaf image through a region of interest feature extractor based on an R-CNN network to obtain the plurality of leaf local region of interest feature maps; Inputting each of the plurality of leaf local region of interest feature maps into a foreground region transfer network based on a spatial attention map to obtain a plurality of foreground salient leaf local region of interest feature maps, including: Performing convolution coding and nonlinear activation processing on the leaf local region of interest feature map to obtain a spatially adaptive leaf local region of interest weight information feature map; Along the channel dimension of the spatially adaptive leaf local region of interest weight information feature map, calculating the cross-channel global mean of each pixel position in the spatially adaptive leaf local region of interest weight information feature map to obtain a spatially adaptive leaf local region of interest weight information representation matrix; In response to each position feature value in the spatially adaptive leaf local region of interest weight information representation matrix being greater than or equal to a predetermined threshold, taking the original value of the position feature value, otherwise setting it to zero to obtain a masked spatially adaptive leaf local region of interest weight information representation matrix; Upsampling the masked spatially adaptive leaf local region of interest weight information representation matrix to obtain a dimensionally aligned spatially adaptive leaf local region of interest weight information representation matrix; The dimension-aligned spatially adaptive leaf local region of interest weight information representation matrix is ​​used as a weight matrix, and the product between the weight matrix and each feature matrix along the channel dimension of the leaf local region of interest feature map is calculated to obtain the foreground salient leaf local region of interest feature map; Inputting each of the plurality of foreground significant leaf local region of interest feature maps into a multi-scale semantic feature extraction module to obtain a plurality of foreground significant leaf local region of interest feature vectors, including: Upsampling the foreground salient leaf local region of interest feature map to obtain an upsampled foreground salient leaf local region of interest feature map; Performing global mean pooling processing on each feature matrix along the channel dimension in the upsampled foreground salient leaf local region of interest feature map to obtain an upsampled foreground salient leaf local region of interest semantic channel feature vector; Performing point convolution encoding on the semantic channel feature vector of the upsampled foreground salient leaf local region of interest to obtain a global semantic feature vector of the foreground salient leaf local region of interest; Performing point convolution coding on the upsampled foreground salient leaf local region of interest feature map to obtain a channel modulated upsampled foreground salient leaf local region of interest feature map; Performing two-dimensional convolution coding on the channel-modulated up-sampled foreground salient leaf local region of interest feature map to obtain a local semantic feature vector of the foreground salient leaf local region of interest; Fusing the global semantic feature vector of the foreground salient leaf local region of interest and the local semantic feature vector of the foreground salient leaf local region of interest to obtain the feature vector of the foreground salient leaf local region of interest; Inputting the plurality of foreground salient leaf local interest region feature vectors into a feature optimizer based on semantic association and semantic span to obtain a plurality of optimized foreground salient leaf local interest region feature vectors, comprising: Extracting a predetermined foreground salient leaf local interest region feature vector from the plurality of foreground salient leaf local interest region feature vectors; Calculating the semantic association between the predetermined foreground significant leaf local interest region feature vector and all other foreground significant leaf local interest region feature vectors to obtain a sequence of foreground significant leaf local interest region semantic associations; Calculating the global mean of the sequence of semantic associations of the foreground-salient leaf local regions of interest to obtain a global average semantic association of the foreground-salient leaf local regions of interest; The global average semantic association of the foreground significant leaf local region of interest is used as a weight, and the predetermined foreground significant leaf local region of interest feature vector is weighted optimized to obtain an optimized foreground significant leaf local region of interest feature vector corresponding to the predetermined foreground significant leaf local region of interest feature vector.

2. The method for detecting diseased citrus leaves for inoculating and preserving the citrus Huanglongbing pathogen according to claim 1, characterized in that: Calculating the semantic association between the predetermined foreground significant leaf local interest region feature vector and all other foreground significant leaf local interest region feature vectors to obtain a sequence of foreground significant leaf local interest region semantic associations, including: Calculating the semantic similarity between the predetermined foreground significant leaf local region of interest feature vector and all other foreground significant leaf local region of interest feature vectors to obtain a sequence of foreground significant leaf local region of interest semantic similarities; Calculating the number of feature vectors between the predetermined foreground significant leaf local interest region feature vector and all other foreground significant leaf local interest region feature vectors to obtain a sequence of foreground significant leaf local interest region distance meter numbers; The sequence of semantic similarities of the foreground significant leaf local interest regions and the sequence of distance counts of the foreground significant leaf local interest regions are divided by position to obtain a sequence of semantic associations of the foreground significant leaf local interest regions.

3. The method for detecting diseased citrus leaves for inoculating and preserving the citrus Huanglongbing pathogen according to claim 2, characterized in that: Based on the leaf global state feature representation, a detection result is obtained, including: passing the leaf global state feature representation vector through a classifier-based Huanglongbing detection module to obtain the detection result.

4. The method for detecting diseased citrus leaves for inoculating and preserving the citrus Huanglongbing pathogen according to claim 3, characterized in that: It also includes a training step: for training the region of interest feature extractor based on the R-CNN network, the foreground area transfer network based on the spatial attention map, the multi-scale semantic feature extraction module, the feature optimizer based on semantic association and semantic span, and the Huanglongbing detection module based on the classifier.

5. The method for detecting diseased citrus leaves for inoculating and preserving the citrus Huanglongbing pathogen according to claim 4, characterized in that: The training step comprises: Acquire training data, wherein the training data includes training leaf images captured by a camera, and real detection results, wherein the real detection results are whether symptoms of Huanglongbing appear; Passing the training leaf image through the region of interest feature extractor based on the R-CNN network to obtain a plurality of training leaf local region of interest feature maps; Inputting each of the plurality of training leaf local interest region feature maps into the foreground region transfer network based on the spatial attention map to obtain a plurality of training foreground salient leaf local interest region feature maps; Inputting each of the plurality of training foreground salient leaf local region of interest feature maps into the multi-scale semantic feature extraction module to obtain a plurality of training foreground salient leaf local region of interest feature vectors; Inputting the plurality of training foreground salient leaf local interest region feature vectors into the feature optimizer based on semantic association and semantic span to obtain a plurality of training optimized foreground salient leaf local interest region feature vectors; Calculating the position mean vector of the feature vectors of the local regions of interest of the plurality of training optimized foreground salient leaves to obtain a training leaf global state feature representation vector; Passing the training leaf global state feature representation vector through the classifier-based Huanglongbing detection module to obtain a classification loss function value; Calculating a blade global state loss function value based on the training blade global state feature representation vector; The weighted sum of the classification loss function value and the leaf global state loss function value is used as the final loss function value and the back propagation of gradient descent is used to train the region of interest feature extractor based on the R-CNN network, the foreground area transfer network based on the spatial attention map, the multi-scale semantic feature extraction module, the feature optimizer based on semantic association and semantic span, and the classifier-based Huanglongbing detection module.

Images