Pest image annotation method based on the integration of vision and prior knowledge

Abstract

The present invention discloses a pest image annotation method based on the integration of vision and prior knowledge, which fuses prior knowledge with processed image data, and uses a feature annotation algorithm to realize the operation of extracting and annotating features of pest images in image data, and the present invention relates to the field of image data processing technology. The pest image annotation method based on the integration of vision and prior knowledge fuses prior knowledge with processed image data, and uses a feature annotation algorithm to realize the operation of extracting and annotating features of pest images in image data, and establishes a large model for pest detection, introduces the annotated image data into the large model for pest detection for training and optimizes model parameters, thereby not only developing a high-precision large model for pest detection and improving recognition efficiency and accuracy, but also providing rich training resources for the pest recognition algorithm, so that the operation effect of pest image annotation is better.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and in particular to a pest image labeling method based on the integration of vision and prior knowledge. Background Art

[0002] With the continuous advancement of agricultural modernization, the rapid and accurate identification of pests has become the key to improving crop yields and quality. Traditional pest identification relies on manual labor, which is inefficient and difficult to meet the needs of large-scale agricultural production. In addition, manual identification is easily affected by subjective factors, and accuracy is difficult to guarantee. At the same time, there is a lack of rapid response mechanisms, and pest problems cannot be dealt with in a timely manner, affecting the effectiveness of prevention and control.

[0003] The reference patent name is: A pest monitoring method based on machine vision (patent publication number: CN108040997A, patent publication date: 2018-05-18), and its steps include: installing an insect trap at a place where pests gather, and setting an image acquisition device to collect images facing the insect trap; identifying pests in the collected image and obtaining the number of pests; if the number of pests is greater than or equal to a preset pest number threshold, extracting the areas where each identified pest is located in the image as multiple suspected pest images, and determining the recognition accuracy of each suspected pest image; calculating the pest prediction level based on the number of pests and the recognition accuracy of each suspected pest image, and automatically collecting pest images by using the image acquisition device facing the insect trap, thereby eliminating the time-consuming and labor-intensive disadvantages of manual visual inspection, and also enabling real-time monitoring of pests; combining the number of pests and the recognition accuracy of each suspected pest image to calculate the pest prediction level has higher accuracy, more meaningful results, and enhanced guidance for pest control.

[0004] Based on the description in the above document, in the process of image processing of existing pest image labeling algorithms, as the number of pest species increases and the complexity becomes higher, objects of different categories may have similarities, and the model fails to solve the problem of fine-grained recognition. The time spent on pest feature labeling also increases accordingly, and the final labeling result also has the problem of accuracy deviation. For this reason, the present invention provides a pest image labeling method based on the integration of vision and prior knowledge. Summary of the invention

[0005] In view of the shortcomings of the existing technology, the present invention provides a pest image labeling method based on the integration of vision and prior knowledge, which solves the problem that in the existing pest image labeling algorithm, as the number of pest species increases and the complexity becomes higher during image processing, objects of different types may have similarities, the model fails to solve the problem of fine-grained recognition, the time spent on pest feature labeling also increases accordingly, and the final labeling result also has the problem of accuracy deviation.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a pest image annotation method based on the integration of vision and prior knowledge, specifically comprising the following steps:

[0007] A1. Using acquisition equipment to perform initial video and image acquisition operations on areas with pests, and storing and transmitting initial image data;

[0008] A2. Preprocessing the initial video and image data to achieve the required data screening and then augmenting the image data;

[0009] A3. Fusing prior knowledge with processed image data, and using feature annotation algorithms to extract and annotate features of pest images in image data, and establishing a large pest detection model, introducing the annotated image data into the large pest detection model for training and optimizing model parameters;

[0010] A4. Apply the trained pest detection model to new pest image data to realize automatic labeling of pest images.

[0011] Preferably, the step of performing preprocessing operations on the initial video and image data in A2 is:

[0012] B1. Segment the initial video data into image data according to the set number of frames within the period, aggregate the segmented image data and the collected initial image data, and perform a sorting operation according to the corresponding collected time nodes;

[0013] B2. Starting from the first image data after sorting, adjacent image data are selected for feature comparison, and the corners of two adjacent image data are correspondingly covered, and it is determined whether the feature part of the image data has changed according to the corresponding feature boundaries of the image data, and the next comparison is performed based on the subsequent image data and the backward adjacent image data;

[0014] B3. If the characteristic part of the image data after comparison changes, the two adjacent image data are retained. If the characteristic part of the image data after comparison does not change, the preceding image data is removed and the succeeding image data is retained. After the image data comparison is completed, the retained image data are aggregated to form the original image data set marked as M.

[0015] Preferably, the operation of summarizing the segmented image data and the collected initial image data in B1 is:

[0016] b11, first compare the total number of segmented image data with the total number of initial image data, and perform processing operations on the image data with a smaller total number;

[0017] b12, by performing pixel adjustment operation on the image data with a smaller total amount, so that the pixel value and size of the smaller image data are consistent with the pixel value and size of the larger image data;

[0018] b13. Implement the direction adjustment operation of the image data to adjust the direction of all the image data to the same, and then implement the feature comparison operation.

[0019] Preferably, the steps of augmenting the image data in A2 are:

[0020] a21, performing feature checks on the image data subsets in the original image data set in turn, wherein the image data subsets having pest features are extracted;

[0021] a22, then randomly select a subset of image data with pest characteristics for rotation operation, and each rotation is 90 degrees and the rotated image data is retained, and the color tone of the image data is adjusted and the data is retained, and the image data is classified to form a diverse image data group;

[0022] a23, and complete the same processing operation of a22 for each image data subset to form a new image data set labeled N.

[0023] Preferably, the feature annotation algorithm in A3 implements the operation of extracting and annotating features of pest images in image data as follows:

[0024] C1. Extract the pest-related knowledge information from the repository and convert the pest-related knowledge information into a knowledge graph dataset labeled P;

[0025] C2, select an image data subset from the new image dataset N, perform feature segmentation on the image data subset, and use known triples to predict the required results in a format corresponding to the knowledge graph dataset;

[0026] C3. Then, a scoring operation is performed based on each predicted result to predict the pest features that the image data subset ultimately corresponds to.

[0027] Preferably, the operation of converting pest-related knowledge information into the knowledge graph dataset P in C1 is:

[0028] c11. Define the known head entity category name of the data according to the physical object related to the pest, and mark the known head entity category name as H;

[0029] c12. Set the corresponding relationship feature based on the known head entity category name H, and mark the relationship feature as R;

[0030] c13. Based on the known head entity category name H and the relationship feature R, the corresponding known tail entity result name is extracted and marked as T, and finally a knowledge graph dataset P (H, R, T) related to pests is formed.

[0031] Preferably, the operation of performing feature segmentation on the image data subset in C2 is:

[0032] c21. Grayscale the image data subset, set pixels at equal distances in the image, and compare the values ​​of the pixels;

[0033] c22, and mark the pixel points where the value changes, and connect the adjacent pixel points that have changed, and take the midpoint of the connected line as the boundary point;

[0034] c23. Then, the boundary points are connected in sequence to realize a closed-loop operation to form a feature boundary, and then the segmentation and marking operations of each feature are completed according to the feature boundary.

[0035] Preferably, the step of using known triples to predict the required result in C2 is:

[0036] D1, based on the segmented feature data combined with the knowledge graph dataset P (H, R, T), and matching the feature data according to the known head entity category name and relationship feature prediction in the knowledge graph dataset P;

[0037] D2. And according to the feature data analysis, the sorting operation of each predicted tail entity result is realized, and the front result data has a higher accuracy rate;

[0038] D3, thereby obtaining a ranked list of predicted tail entity results;

[0039] And the feature prediction method is Q (h, r, t), h is the known head entity category name in the feature data, r is the corresponding relationship feature in the feature data, and t is the tail entity result predicted by the corresponding head entity category name h combined with the relationship feature r.

[0040] Preferably, the operation of scoring each predicted result in C3 is:

[0041] c31. Setting according to the result of each predicted tail entity result sorting list, marking according to the corresponding pest characteristics, so that the content in each sorting list is marked with the corresponding pest characteristic information;

[0042] c32, extract the top three items in each sorted list to form a data group, and perform a scoring operation on the content in the data group. Each data group will receive one point if the same content appears once. After all data groups have been traversed, the scoring results corresponding to each pest feature are counted;

[0043] c33. Extract the pest feature information with the highest score, which is the marked pest feature result.

[0044] Preferably, the expression for the scoring operation of C3 is:

[0045] ;

[0046] F represents the integral result of a single pest feature, Y represents a single pest feature, n represents the number of single pest features in all data groups, m represents the number of data groups, and k represents the set pest feature ratio threshold.

[0047] The present invention provides a pest image annotation method based on the integration of vision and prior knowledge. Compared with the prior art, it has the following beneficial effects:

[0048] (1) This pest image labeling method based on the integration of vision and prior knowledge integrates prior knowledge with processed image data, and uses a feature labeling algorithm to extract and label pest image features in image data, and establishes a large pest detection model. The labeled image data is introduced into the pest detection large model for training and optimizing model parameters. In this way, not only a high-precision pest detection large model is developed to improve recognition efficiency and accuracy, but also rich training resources are provided for the pest recognition algorithm, making the operation effect of pest image labeling better.

[0049] (2) The pest image annotation method based on the integration of vision and prior knowledge combines the segmented feature data with the knowledge graph dataset P, and matches the feature data according to the known head entity category name and relationship feature prediction in the knowledge graph dataset P, and implements the sorting operation of each predicted tail entity result according to the feature data analysis to obtain a sorted list of predicted tail entity results, thereby improving the recognition efficiency, realizing the prediction of the results according to multiple feature points, giving the fault tolerance rate and being able to effectively avoid the errors generated, thereby improving the accuracy of annotation.

[0050] (3) The pest image labeling method based on the integration of vision and prior knowledge extracts the top three items in each sorted list to form a data group, and performs a scoring operation on the content in the data group. Each data group accumulates one point when the same content appears once. After all data groups have been traversed, the scoring results corresponding to each pest feature are counted, and the pest feature information with the highest score is extracted as the labeled pest feature result, thereby reducing the judgment time of unnecessary data and making the extracted pest feature result more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1is an operation flow chart of the pest image annotation algorithm of the present invention;

[0052] Figure 2 The operation flow chart of data preprocessing of the present invention;

[0053] Figure 3 It is a schematic diagram of the knowledge graph of the tea geometrid of the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] See also Figure 1-Figure 3 , the present invention provides two technical solutions:

[0056] Embodiment 1: A pest image labeling method based on the integration of vision and prior knowledge, specifically comprising the following steps:

[0057] A1. Using acquisition equipment to perform initial video and image acquisition operations on areas with pests, and storing and transmitting initial image data;

[0058] A2. Preprocessing the initial video and image data to achieve the required data screening and then augmenting the image data;

[0059] A3. Fusing prior knowledge with processed image data, and using feature annotation algorithms to extract and annotate features of pest images in image data, and establishing a large pest detection model, introducing the annotated image data into the large pest detection model for training and optimizing model parameters;

[0060] A4. Apply the trained pest detection model to new pest image data to realize automatic labeling of pest images.

[0061] Among them, by fusing prior knowledge with processed image data, and using feature labeling algorithms to realize the operation of extracting and labeling pest image features in image data, and establishing a large pest detection model, the labeled image data is introduced into the pest detection large model for training and optimizing model parameters. In this way, not only a high-precision pest detection large model is developed to improve the efficiency and accuracy of recognition, but also rich training resources are provided for the pest recognition algorithm, making the operation effect of pest image labeling better.

[0062] In the embodiment of the present invention, the steps of performing preprocessing operations on the initial video and image data in A2 are:

[0063] B1. Segment the initial video data into image data according to the set number of frames within the period, aggregate the segmented image data and the collected initial image data, and perform a sorting operation according to the corresponding collected time nodes;

[0064] B2. Starting from the first image data after sorting, adjacent image data are selected for feature comparison, and the corners of two adjacent image data are correspondingly covered, and it is determined whether the feature part of the image data has changed according to the corresponding feature boundaries of the image data, and the next comparison is performed based on the subsequent image data and the backward adjacent image data;

[0065] B3. If the characteristic part of the image data after comparison changes, the two adjacent image data are retained. If the characteristic part of the image data after comparison does not change, the preceding image data is removed and the succeeding image data is retained. After the image data comparison is completed, the retained image data are aggregated to form the original image data set marked as M.

[0066] In the embodiment of the present invention, the operation of summarizing the segmented image data and the collected initial image data in B1 is:

[0067] b11, first compare the total number of segmented image data with the total number of initial image data, and perform processing operations on the image data with a smaller total number;

[0068] b12, by performing pixel adjustment operation on the image data with a smaller total amount, so that the pixel value and size of the smaller image data are consistent with the pixel value and size of the larger image data;

[0069] b13. Implement the direction adjustment operation of the image data to adjust the direction of all the image data to the same, and then implement the feature comparison operation.

[0070] In the embodiment of the present invention, the steps of performing augmentation processing on the image data in A2 are:

[0071] a21, performing feature checks on the image data subsets in the original image data set in turn, wherein the image data subsets having pest features are extracted;

[0072] a22, then randomly select a subset of image data with pest characteristics for rotation operation, and each rotation is 90 degrees and the rotated image data is retained, and the color tone of the image data is adjusted and the data is retained, and the image data is classified to form a diverse image data group;

[0073] a23, and complete the same processing operation of a22 for each image data subset to form a new image data set labeled N.

[0074] In the embodiment of the present invention, the feature annotation algorithm in A3 implements the operation of extracting and annotating features of pest images in image data as follows:

[0075] C1. Extract the pest-related knowledge information from the repository and convert the pest-related knowledge information into a knowledge graph dataset labeled P;

[0076] C2, select an image data subset from the new image dataset N, perform feature segmentation on the image data subset, and use known triples to predict the required results in a format corresponding to the knowledge graph dataset;

[0077] C3. Then, a scoring operation is performed based on each predicted result to predict the pest features that the image data subset ultimately corresponds to.

[0078] In the embodiment of the present invention, the operation of converting pest-related knowledge information into the knowledge graph dataset P in C1 is:

[0079] c11. Define the known head entity category name of the data according to the physical object related to the pest, and mark the known head entity category name as H;

[0080] c12. Set the corresponding relationship feature based on the known head entity category name H, and mark the relationship feature as R;

[0081] c13. Based on the known head entity category name H and the relationship feature R, the corresponding known tail entity result name is extracted and marked as T, and finally a knowledge graph dataset P (H, R, T) related to pests is formed.

[0082] In the embodiment of the present invention, the operation of performing feature segmentation on the image data subset in C2 is:

[0083] c21. Grayscale the image data subset, set pixels at equal distances in the image, and compare the values ​​of the pixels;

[0084] c22, and mark the pixel points where the value changes, and connect the adjacent pixel points that have changed, and take the midpoint of the connected line as the boundary point;

[0085] c23. Then, the boundary points are connected in sequence to realize a closed-loop operation to form a feature boundary, and then the segmentation and marking operations of each feature are completed according to the feature boundary.

[0086] In the embodiment of the present invention, the step of using known triples to predict the required result in C2 is:

[0087] D1, based on the segmented feature data combined with the knowledge graph dataset P (H, R, T), and matching the feature data according to the known head entity category name and relationship feature prediction in the knowledge graph dataset P;

[0088] D2. And according to the feature data analysis, the sorting operation of each predicted tail entity result is realized, and the front result data has a higher accuracy rate;

[0089] D3, thereby obtaining a ranked list of predicted tail entity results;

[0090] And the feature prediction method is Q (h, r, t), h is the known head entity category name in the feature data, r is the corresponding relationship feature in the feature data, and t is the tail entity result predicted by the corresponding head entity category name h combined with the relationship feature r.

[0091] Among them, by combining the segmented feature data with the knowledge graph dataset P, and matching the feature data according to the known head entity category name and relationship feature prediction in the knowledge graph dataset P, and implementing the sorting operation of each predicted tail entity result according to the feature data analysis, a sorted list of predicted tail entity results is obtained, thereby improving the recognition efficiency, realizing the prediction of results based on multiple feature points, giving fault tolerance and being able to effectively avoid the errors generated, thereby improving the accuracy of labeling.

[0092] Among them, knowledge graphs usually represent knowledge in triples, and the general format is (head entity, relationship, tail entity). Taking the knowledge graph of tea geometrid pests as an example, Figure 3 As shown, for example, the knowledge graph dataset P (H, R, T) is P (head, color, gray-brown), and the prediction operation is implemented through Q (h, r, t), and (head, color,?) is obtained, and t forms a sorted list, as shown in Table 1:

[0093] Table 1 Sorted list

[0094]

[0095] Among them, each color corresponds to the characteristics of each pest.

[0096] In the embodiment of the present invention, the operation of scoring according to each predicted result in C3 is:

[0097] c31. Setting according to the result of each predicted tail entity result sorting list, marking according to the corresponding pest characteristics, so that the content in each sorting list is marked with the corresponding pest characteristic information;

[0098] c32, extract the top three items in each sorted list to form a data group, and perform a scoring operation on the content in the data group. Each data group will receive one point if the same content appears once. After all data groups have been traversed, the scoring results corresponding to each pest feature are counted;

[0099] c33. Extract the pest feature information with the highest score, which is the marked pest feature result.

[0100] Among them, the top three items in each sorted list are extracted to form a data group, and a scoring operation is performed corresponding to the content in the data group. One point is accumulated for each data group where the same content appears once. After all data groups have been traversed, the scoring results corresponding to each pest feature are counted, and the pest feature information with the highest score is extracted as the marked pest feature result, which reduces the judgment time for unnecessary data and makes the extracted pest feature results more accurate.

[0101] In the embodiment of the present invention, the expression for C3 to perform the scoring operation is:

[0102] ;

[0103] F represents the integral result of a single pest feature, Y represents a single pest feature, n represents the number of single pest features in all data groups, m represents the number of data groups, and k represents the set pest feature ratio threshold.

[0104] Embodiment 2: The difference from Embodiment 1 is that the present invention further designs a comparative experiment, by applying the existing pest image annotation algorithm integrating vision and prior knowledge and the pest image annotation algorithm integrating vision and prior knowledge of the present invention to known pest feature areas for detection, and recording the time used for annotation and the accuracy of the final result. The specific results are shown in Table 2:

[0105] Table 2 Comparison results

[0106]

[0107] In summary, by adopting the pest image standard algorithm of the present invention to perform the labeling operation of the pest features in the area, the time taken to label all features is the shortest, and the accuracy of the labeled results is higher. Therefore, the effect of using the pest image standard algorithm of the present invention in actual operation is better.

[0108] Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0109] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0110] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A pest image annotation method based on the integration of vision and prior knowledge, characterized by: The specific steps include: A1. Using acquisition equipment to perform initial video and image acquisition operations on areas with pests, and storing and transmitting initial image data; A2. Preprocessing the initial video and image data to achieve the required data screening and then augmenting the image data; A3. Fusing prior knowledge with processed image data, and using feature annotation algorithms to extract and annotate features of pest images in image data, and establishing a large pest detection model, introducing the annotated image data into the large pest detection model for training and optimizing model parameters; A4. Apply the trained pest detection model to new pest image data to realize automatic annotation of pest images; The feature annotation algorithm in A3 implements the operation of extracting and annotating features of pest images in image data as follows: C1. Extract the pest-related knowledge information from the repository and convert the pest-related knowledge information into a knowledge graph dataset labeled P; C2, select an image data subset from the new image dataset N, perform feature segmentation on the image data subset, and use known triples to predict the required results in a format corresponding to the knowledge graph dataset; C3. Then, a scoring operation is performed based on each predicted result to predict the pest features that the image data subset ultimately corresponds to.

2. The pest image labeling method based on the integration of vision and prior knowledge according to claim 1 is characterized by: The steps of preprocessing the initial video and image data in A2 are: B1. Segment the initial video data into image data according to the set number of frames within the period, aggregate the segmented image data and the collected initial image data, and perform a sorting operation according to the corresponding collected time nodes; B2. Starting from the first image data after sorting, adjacent image data are selected for feature comparison, and the corners of two adjacent image data are correspondingly covered, and it is determined whether the feature part of the image data has changed according to the corresponding feature boundaries of the image data, and the next comparison is performed based on the subsequent image data and the backward adjacent image data; B3. If the characteristic part of the image data after comparison changes, the two adjacent image data are retained. If the characteristic part of the image data after comparison does not change, the preceding image data is removed and the succeeding image data is retained. After the image data comparison is completed, the retained image data are aggregated to form the original image data set marked as M.

3. The pest image annotation method based on the integration of vision and prior knowledge according to claim 2 is characterized by: The operation of summarizing the segmented image data and the collected initial image data in B1 is: b11, first compare the total number of segmented image data with the total number of initial image data, and perform processing operations on the image data with a smaller total number; b12, by performing pixel adjustment operation on the image data with a smaller total amount, so that the pixel value and size of the smaller image data are consistent with the pixel value and size of the larger image data; b13. Implement the direction adjustment operation of the image data to adjust the direction of all the image data to the same, and then implement the feature comparison operation.

4. The pest image labeling method based on the integration of vision and prior knowledge according to claim 2 is characterized by: The steps of augmenting the image data in A2 are as follows: a21, performing feature checks on the image data subsets in the original image data set in turn, wherein the image data subsets having pest features are extracted; a22, then randomly select a subset of image data with pest characteristics for rotation operation, and each rotation is 90 degrees and the rotated image data is retained, and the color tone of the image data is adjusted and the data is retained, and the image data is classified to form a diverse image data group; a23, and complete the same processing operation of a22 for each image data subset to form a new image data set labeled N.

5. The pest image labeling method based on the integration of vision and prior knowledge according to claim 1 is characterized by: The operation of converting pest-related knowledge information into the knowledge graph dataset P in C1 is: c11. Define the known head entity category name of the data according to the physical object related to the pest, and mark the known head entity category name as H; c12. Set the corresponding relationship feature based on the known head entity category name H, and mark the relationship feature as R; c13. Based on the known head entity category name H and the relationship feature R, the corresponding known tail entity result name is extracted and marked as T, and finally a knowledge graph dataset P (H, R, T) related to pests is formed.

6. The pest image labeling method based on the integration of vision and prior knowledge according to claim 1 is characterized by: The operation of performing feature segmentation on the image data subset in C2 is: c21. Grayscale the image data subset, set pixels at equal distances in the image, and compare the values ​​of the pixels; c22, and mark the pixel points where the value changes, and connect the adjacent pixel points that have changed, and take the midpoint of the connected line as the boundary point; c23. Then, the boundary points are connected in sequence to realize a closed-loop operation to form a feature boundary, and then the segmentation and marking operations of each feature are completed according to the feature boundary.

7. The pest image labeling method based on the integration of vision and prior knowledge according to claim 5 is characterized by: The steps of using known triples to predict the desired results in C2 are: D1, based on the segmented feature data combined with the knowledge graph dataset P (H, R, T), and matching the feature data according to the known head entity category name and relationship feature prediction in the knowledge graph dataset P; D2. And according to the feature data analysis, the sorting operation of each predicted tail entity result is realized, and the front result data has a higher accuracy rate; D3, thereby obtaining a ranked list of predicted tail entity results; And the feature prediction method is Q (h, r, t), h is the known head entity category name in the feature data, r is the corresponding relationship feature in the feature data, and t is the tail entity result predicted by the corresponding head entity category name h combined with the relationship feature r.

8. The pest image labeling method based on the integration of vision and prior knowledge according to claim 7 is characterized by: The operation of scoring each predicted result in C3 is: c31. Setting according to the result of each predicted tail entity result sorting list, marking according to the corresponding pest characteristics, so that the content in each sorting list is marked with the corresponding pest characteristic information; c32, extract the top three items in each sorted list to form a data group, and perform a scoring operation on the content in the data group. Each data group will receive one point if the same content appears once. After all data groups have been traversed, the scoring results corresponding to each pest feature are counted; c33. Extract the pest feature information with the highest score, which is the marked pest feature result.

9. The pest image labeling method based on the integration of vision and prior knowledge according to claim 1 is characterized by: The expression for the scoring operation of C3 is: ； F represents the integral result of a single pest feature, Y represents a single pest feature, n represents the number of single pest features in all data groups, m represents the number of data groups, and k represents the set pest feature ratio threshold.

Images