A Visual Language-Based Method for Identifying Crop Diseases and Pests

Abstract

This invention discloses a method for identifying crop diseases and pests based on visual language, comprising the following steps: S100, temporal image acquisition and temporal texture set construction; S200, bidirectional dynamic masking processing and mask temporal feature map generation; S300, visual language anchor point library construction and bimodal feature encoding; S400, temporal anchor point alignment and coarse identification of diseases and pests; S500, disease and pest evolution inference and confidence level determination; S600, structured text output and presentation of recognition results. This invention uses temporal image sequence acquisition and temporal texture evolution features as the core recognition basis, completely breaking through the technical limitations of traditional single-frame static recognition, effectively reducing the probability of confused recognition and misjudgment, and significantly improving the accuracy and discriminative power of disease and pest identification. Simultaneously, by guiding temporal feature alignment and disease and pest evolution confidence level determination through language anchor points, it can simultaneously output the disease and pest evolution stage and damage level, providing direct and reliable decision support for precision plant protection in the field.

Description

Technical Field

[0001] This invention relates to the field of crop pest and disease identification technology, and in particular to a method for identifying crop pests and diseases based on visual language. Background Technology

[0002] Against the backdrop of the rapid development of smart agriculture, crop pest and disease identification technology based on computer vision and visual language multimodal methods has become an important support for precision plant protection. Current mainstream identification methods mostly use single-frame field images as input, determining pest and disease types through general visual feature extraction and classification networks. Some methods introduce text labels for simple multimodal fusion, which improves identification efficiency in conventional scenarios to some extent. However, these methods heavily rely on static features, fail to utilize the temporal evolution of pests and diseases, have limited ability to distinguish similar pests and diseases, and lack robustness under complex field backgrounds, lighting changes, and occlusion interference, making it difficult to meet the high-precision requirements of actual production.

[0003] Existing visual-language recognition methods generally suffer from problems such as simple modal fusion, language information being used only for auxiliary annotation, and the lack of temporal guidance mechanisms. They can only output a single pest or disease category, failing to determine evolutionary stages or verify confidence levels, resulting in a lack of interpretability and control guidance value. Furthermore, offline training modes cannot achieve sample feedback and model iterative optimization, making it difficult to adapt to recognition scenarios across different regions, crops, and emerging pests and diseases, leading to poor model generalization ability and sustained usability. Therefore, there is an urgent need in this field to break through the traditional single-frame static recognition framework and construct a pest and disease recognition method based on temporal texture association, language-guided mask matching, and evolutionary inference to address the many shortcomings of existing technologies. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a visual language-based method for identifying crop diseases and pests.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for identifying crop diseases and pests based on visual language includes the following steps:

[0007] S100, Temporal Image Acquisition and Temporal Texture Set Construction: At fixed time intervals, multiple frames of RGB images are continuously acquired from the same crop monitoring site to construct a temporal image sequence of diseases and pests containing information on the growth and changes of diseases and pests. The temporal image sequence is then subjected to inter-frame registration, illumination normalization, and enhancement processing of lesions and insect damage areas. A set of texture features that continuously change in the temporal dimension is extracted to form a temporal texture set.

[0008] S200, Bidirectional dynamic masking and mask temporal feature map generation: Bidirectional dynamic masking is performed on the temporal texture set, with the previous frame texture as the forward constraint and the next frame texture as the backward constraint, to calculate a mask temporal feature map that retains only the dynamic change features of pests and diseases and removes the static background and occlusion noise in the field.

[0009] S300, Visual Language Anchor Point Library Construction and Bimodal Feature Encoding: A visual language anchor point library containing descriptions of the temporal evolution of pests and diseases, texts of the damage stages, keywords of texture changes, and prevention and control schemes is pre-constructed. A language encoder is used to encode the text anchor point information into language-guided feature vectors, and a visual encoder is used to encode the mask temporal feature map into visual temporal feature vectors.

[0010] S400, Temporal Anchor Alignment and Coarse Identification of Pests and Diseases: Using language-guided feature vectors as supervision signals, temporal anchor alignment mapping is performed on visual temporal feature vectors. Temporal texture distance and mask matching degree are calculated respectively. Weighted comprehensive similarity is obtained by fusing according to preset weights. Coarse identification of pests and diseases is completed based on weighted comprehensive similarity.

[0011] S500, Pest and Disease Evolution Deduction and Confidence Determination: Based on the coarse recognition results and the temporal texture change rate, the probability of pest and disease evolution stages is deduced, the evolution confidence is calculated, and the evolution confidence is compared with a preset threshold. If it is higher than the threshold, it is determined as the final recognition category; if it is lower than the threshold, it returns to step S2 to perform temporal texture enhancement again.

[0012] S600, Structured Text Output and Recognition Result Presentation: Based on the final recognition category and corresponding evolutionary stage, the system matches and outputs structured text from the visual language anchor point library, including the name of the pest or disease, the level of damage, the evolutionary stage, the cause analysis, and the targeted prevention and control plan.

[0013] Preferably, the formula for calculating the mask weight in the bidirectional dynamic masking process in step S200 is:

[0014]

[0015] in, For the first Frame mask weights, For the first Frame texture feature map, For the first Frame texture feature map, For the first Frame texture feature map, Forward time series weighting coefficients, satisfying .

[0016] Preferably, the temporal texture distance mentioned in step S400 is used to measure the temporal variation pattern of pests and diseases, and the calculation formula is as follows:

[0017]

[0018] in, For the first Frame visual temporal feature vectors For the first Frame visual temporal feature vectors This represents the total number of frames in the time-series image sequence. The smaller the value, the more regular the temporal changes in the patterns of pests and diseases.

[0019] Preferably, the mask matching degree mentioned in step S400 is used to measure the degree of matching between the language guidance features and the mask visual features, and the calculation formula is as follows:

[0020]

[0021] in, for Frame language-guided feature vectors For the first Frame mask weights.

[0022] Preferably, the weighted comprehensive similarity in step S400 is obtained by fusing the mask matching degree and the temporal texture distance, and the calculation formula is as follows:

[0023]

[0024] in, Let be the mask matching degree weight coefficient, satisfying .

[0025] Preferably, the confidence level of pest and disease evolution in step S5 is calculated by combining the weighted comprehensive similarity and the evolution rate deviation, as shown in the formula:

[0026]

[0027] in, This represents the current rate of change in the texture of pests and diseases. This represents the standard evolution rate of the pest / disease. A higher value indicates a more reliable recognition result.

[0028] Preferably, in step S500, an evolution confidence threshold is set. ,when The final recognition result is output in time. Then return to step S2 to re-execute bidirectional dynamic masking and temporal texture enhancement.

[0029] Preferably, the visual language anchor point library adopts a hierarchical structure, with one pest category corresponding to multiple different evolutionary stages, and one evolutionary stage corresponding to a set of temporal texture description text, hazard level labels, and standardized prevention and control measures.

[0030] Preferably, after the structured text is output in step S600, the temporal texture features, evolutionary confidence, recognition results and actual observation results of this recognition are stored together in the visual language anchor point library, and the language encoder and visual encoder are updated online incrementally to form a self-iterative optimization recognition closed loop.

[0031] The present invention has the following beneficial effects:

[0032] 1. This invention uses time-series image sequence acquisition and time-series texture evolution features as the core recognition basis, which completely breaks through the technical limitations of traditional single-frame static recognition. It can accurately distinguish between pests and diseases with similar morphology and similar symptoms, effectively reduce the probability of confusion and misjudgment, and greatly improve the accuracy and distinguishability of pest and disease identification.

[0033] 2. This invention uses bidirectional dynamic masking to enhance the dynamic features of pests and diseases and remove field noise. It can effectively suppress interference factors such as soil, weeds, light changes and shading, and significantly improve the purity of feature extraction and the robustness of the recognition process in complex field environments.

[0034] 3. This invention uses language anchors to guide the alignment of temporal features and the determination of confidence in the evolution of pests and diseases, so as to make the identification results interpretable, traceable, and predictable. It can simultaneously output the evolution stage and hazard level of pests and diseases, providing direct and reliable decision support for precision plant protection in the field.

[0035] 4. This invention constructs a closed loop of sample feedback and online incremental iteration of the model, which can continuously optimize the visual encoder and language encoder using measured data, continuously improve the model's generalization ability and adaptability, and can be adapted to different crops, different regions and new pest and disease identification scenarios, thus extending the model's life cycle and practical value. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the overall process of a visual language-based method for identifying crop diseases and pests proposed in this invention.

[0037] Figure 2 This is a schematic diagram illustrating the principle of bidirectional dynamic masking proposed in this invention;

[0038] Figure 3 This is a flowchart of the visual language feature alignment and similarity calculation proposed in this invention;

[0039] Figure 4 This is a flowchart illustrating the closed-loop process for the confidence extrapolation and identification of pest and disease evolution proposed in this invention. Detailed Implementation

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0041] Reference Figure 1 A method for identifying crop diseases and pests based on visual language includes the following steps:

[0042] S100, Temporal Image Acquisition and Temporal Texture Set Construction: At fixed time intervals, multiple frames of RGB images are continuously acquired from the same crop monitoring site to construct a temporal image sequence of diseases and pests containing information on the growth and changes of diseases and pests. The temporal image sequence is then subjected to inter-frame registration, illumination normalization, and enhancement processing of lesions and insect damage areas. A set of texture features that continuously change in the temporal dimension is extracted to form a temporal texture set.

[0043] S200, Bidirectional Dynamic Masking Processing and Mask Timing Feature Map Generation: Refer to Figure 2 The temporal texture set is subjected to bidirectional dynamic masking, with the previous frame texture as the forward constraint and the next frame texture as the backward constraint, to calculate a masked temporal feature map that retains only the dynamic changes of pests and diseases and removes static background and occlusion noise in the field; the mask weight calculation formula for bidirectional dynamic masking is:

[0044]

[0045] in, For the first Frame mask weights, For the first Frame texture feature map, For the first Frame texture feature map, For the first Frame texture feature map, Forward time series weighting coefficients, satisfying .

[0046] S300, Visual Language Anchor Library Construction and Bimodal Feature Encoding: A visual language anchor library containing descriptions of the temporal evolution of pests and diseases, text of the damage stage, keywords of texture changes, and control measures is pre-constructed. A language encoder is used to encode the text anchor information into language-guided feature vectors, and a visual encoder is used to encode the masked temporal feature map into visual temporal feature vectors. The visual language anchor library adopts a hierarchical structure, with one pest and disease category corresponding to multiple different evolutionary stages, and one evolutionary stage corresponding to a set of temporal texture description text, damage level labels, and standardized control measures.

[0047] S400, Temporal Anchor Alignment and Coarse Pest and Disease Identification: Using language-guided feature vectors as supervision signals, temporal anchor alignment mapping is performed on visual temporal feature vectors. Temporal texture distance and mask matching degree are calculated separately. A weighted comprehensive similarity is obtained by fusing these values ​​according to preset weights. Coarse pest and disease identification is then completed based on this weighted comprehensive similarity. The temporal texture distance is used to measure the temporal variation pattern of pests and diseases; the calculation formula is as follows:

[0048]

[0049] in, For the first Frame visual temporal feature vectors For the first Frame visual temporal feature vectors This represents the total number of frames in the time-series image sequence. The smaller the value, the more regular the temporal changes in the patterns of pests and diseases.

[0050] Mask matching degree is used to measure the degree of matching between language guidance features and mask visual features. The calculation formula is as follows:

[0051]

[0052] in, for Frame language-guided feature vectors For the first Frame mask weights.

[0053] For the specific procedures of steps S300 and S400 above, please refer to [the provided text]. Figure 3 .

[0054] S500. Pest and Disease Evolution Deduction and Confidence Determination: Based on the coarse identification results and the temporal texture change rate, the probability of pest and disease evolution stages is deduced, and the evolution confidence is calculated. The evolution confidence is compared with a preset threshold. If it is higher than the threshold, it is determined as the final identification category; if it is lower than the threshold, the process returns to step S2 to perform temporal texture enhancement again. The pest and disease evolution confidence is calculated by weighted comprehensive similarity and evolution rate deviation, using the following formula:

[0055]

[0056] in, This represents the current rate of change in the texture of pests and diseases. This represents the standard evolution rate of the pest / disease. A higher value indicates a more reliable recognition result.

[0057] Specifically, set the evolution confidence threshold. ,when The final recognition result is output in time. Then return to step S2 to re-execute bidirectional dynamic masking and temporal texture enhancement.

[0058] S600, Structured Text Output and Recognition Result Presentation: Based on the final recognition category and corresponding evolutionary stage, structured text containing the pest / disease name, damage level, evolutionary stage, causal analysis, and targeted prevention and control measures is matched and output from the visual language anchor point library. After outputting the structured text, the temporal texture features, evolutionary confidence, recognition results, and actual observation results of this recognition are stored in the visual language anchor point library. The language encoder and visual encoder are then updated incrementally online, forming a self-iterative and optimized recognition closed loop.

[0059] The identification and comparison logic structure of steps S500 and S600 above is as follows: Figure 4 As shown.

[0060] Experimental Design

[0061] Experimental subjects: five common crops: tomatoes, peanuts, corn, rice and wheat, including samples from the early, middle and late evolution stages of pests and diseases, covering morphologically similar and easily confused pest and disease types and conventional field disturbance scenarios.

[0062] Sample setup: 200 test samples were set for each crop, with each sample consisting of time-series images of the same locus, for a total of 1000 samples.

[0063] Test environment: Natural field environment, including normal light variations and a small amount of weeds and leaves blocking the light.

[0064] Evaluation indicators: identification accuracy, similar pest and disease differentiation rate, and average identification confidence level.

[0065] Group settings

[0066] Example 1: The entire process of the crop disease and pest identification method based on visual language of the present invention is fully adopted, including time-series image acquisition, bidirectional dynamic masking, visual language anchor point library, time-series anchor point alignment, evolutionary confidence determination, and model self-iterative update.

[0067] Example 2: Using some of the technologies of this invention, the temporal image acquisition and visual feature extraction are retained, while bidirectional dynamic masking, visual language anchor guidance, evolutionary confidence determination and model self-iterative optimization are eliminated.

[0068] Comparative example: Using existing conventional pest and disease identification methods, there is single-frame image input, static visual feature extraction, no temporal information, no multimodal fusion, no confidence verification and model iteration.

[0069] Example 1 (using the method of this invention in its entirety)

[0070] Implementation steps

[0071] 1. Collect multiple frames of RGB images of the same monitoring site at fixed time intervals to construct a time-series image sequence of pests and diseases. After inter-frame registration, illumination normalization, and enhancement processing of lesions and insect damage areas, extract the time-series texture feature set.

[0072] 2. Perform bidirectional dynamic masking on the temporal texture set to generate a masked temporal feature map that retains only the dynamic changes of pests and diseases, and removes the interference of static background in the field.

[0073] 3. Call the visual language anchor library to generate language-guided feature vectors and visual temporal feature vectors through the language encoder and visual encoder, respectively.

[0074] 4. Using language-guided features as supervision signals, we complete temporal anchor point alignment, calculate temporal texture distance and mask matching degree, fuse them to obtain weighted comprehensive similarity, and complete coarse identification of pests and diseases.

[0075] 5. Combine the coarse recognition results with the texture change rate to perform evolutionary deduction, calculate the evolutionary confidence. If it is higher than the threshold, determine the final category; if it is lower than the threshold, return to the mask enhancement step.

[0076] 6. Output structured text such as pest and disease name, damage level, evolution stage, and control plan, and feed the identified data back to update the anchor point library and encoder to achieve model self-iterative optimization.

[0077] The identification data is shown in the table below:

[0078] Table 1: Statistical Table of Identification Results of Five Crop Diseases and Pests in Example 1

[0079] crop Number of test samples Correctly identify numbers Recognition accuracy Differentiation rate of similar diseases and pests Average recognition confidence tomato 200 189 94.5% 91.2% 0.88 peanut 200 186 93.0% 89.7% 0.86 corn 200 191 95.5% 92.8% 0.89 rice 200 188 94.0% 90.9% 0.87 wheat 200 185 92.5% 89.1% 0.85 average 1000 939 93.9% 90.74% 0.87

[0080] It should be noted that in the table above, the distinction between similar diseases and pests refers to the two diseases and pests with the most similar morphology in each crop (tomato: early blight vs. late blight; rice: rice blast vs. sesame leaf spot; wheat: stripe rust vs. leaf rust; corn: large leaf spot vs. small leaf spot; rapeseed: sclerotinia rot vs. downy mildew).

[0081] Example 2 (using part of the method of this invention)

[0082] Implementation steps

[0083] 1. Collect crop images in chronological order and extract texture features without performing bidirectional dynamic masking.

[0084] 2. Only visual encoders are used to extract features, without visual language anchor point library guidance and temporal feature alignment operations.

[0085] 3. The recognition results are directly output based on visual feature similarity, without evolutionary confidence determination or online model iteration updates.

[0086] The identification data is shown in the table below:

[0087] Table 2: Statistical Table of Identification Results of Five Crop Diseases and Pests in Example 2

[0088] crop Number of test samples Correctly identify numbers Recognition accuracy Differentiation rate of similar diseases and pests Average recognition confidence tomato 200 176 88.0% 82.3% 0.79 peanut 200 173 86.5% 80.8% 0.77 corn 200 179 89.5% 84.1% 0.80 rice 200 175 87.5% 82.0% 0.78 wheat 200 172 86.0% 79.9% 0.76 average 1000 875 87.5% 81.82% 0.78

[0089] It should be noted that the test samples in the table above are completely identical to those in Example 1, and the similar pests and diseases are distinguished from those in Example 1 (tomato: early blight vs. late blight; rice: rice blast vs. sesame leaf spot; wheat: stripe rust vs. leaf rust; corn: large leaf spot vs. small leaf spot; rapeseed: sclerotinia rot vs. downy mildew).

[0090] Comparative examples (existing conventional identification methods)

[0091] Implementation steps

[0092] 1. Using a single frame of static RGB image as the recognition input, extract traditional static visual features.

[0093] 2. The disease and pest type determination is completed using a conventional classification network, without temporal features or visual-linguistic multimodal fusion.

[0094] 3. It only outputs the category of pests and diseases, without judging the evolutionary stage, confidence verification, or model iteration capability.

[0095] The identification data is shown in the table below:

[0096] Table 3: Statistical Table of Identification Results of Diseases and Pests of Five Crops in the Comparative Example

[0097] crop Number of test samples Correctly identify numbers Recognition accuracy Differentiation rate of similar diseases and pests Average recognition confidence tomato 200 165 82.5% 73.6% 0.71 peanut 200 161 80.5% 71.8% 0.69 corn 200 168 84.0% 75.2% 0.72 rice 200 163 81.5% 72.9% 0.70 wheat 200 159 79.5% 70.6% 0.68 average 1000 816 81.6% 72.82% 0.70

[0098] It should be noted that the test samples in the table above are completely identical to those in Example 1, and the similar pests and diseases are distinguished from those in Example 1 (tomato: early blight vs. late blight; rice: rice blast vs. sesame leaf spot; wheat: stripe rust vs. leaf rust; corn: large leaf spot vs. small leaf spot; rapeseed: sclerotinia rot vs. downy mildew).

[0099] Experimental conclusions

[0100] I. This invention significantly improves the overall accuracy of pest and disease identification.

[0101] Example 1 achieved an average recognition accuracy of 93.90%, maintaining a high accuracy rate of over 94% on crops such as tomatoes, corn, and rice. This represents a 6.4 percentage point improvement over Example 2's average accuracy of 87.50%, and a 12.3 percentage point improvement over the comparative example's average accuracy of 81.60%. The data demonstrates that this invention, through temporal texture features, bidirectional dynamic masking, and visual-language dual-modal fusion, effectively overcomes the shortcomings of static recognition, significantly improving recognition accuracy. It exhibits stable and reliable recognition performance on five common crops, meeting the precision requirements of practical field applications.

[0102] II. This invention significantly enhances the ability to distinguish between similar pests and diseases.

[0103] In Example 1, the average differentiation rate for similar pests and diseases was 90.74%, significantly higher than the 81.82% in Example 2 and the 72.82% in the comparative example, representing increases of 8.92 and 17.92 percentage points, respectively. In easily confused pest and disease identification scenarios in the field, this invention, relying on temporal evolution features and linguistic anchor point alignment, can accurately capture the textural differences of pests and diseases over time, effectively distinguishing pest and disease types with similar symptoms and morphologies, significantly reducing the probability of confusion and misjudgment, and solving the problem of insufficient differentiation ability for similar pests and diseases in existing technologies.

[0104] Third, this invention effectively improves the reliability and credibility of the recognition results.

[0105] The average recognition confidence level in Example 1 was 0.87, in Example 2 it was 0.78, and in the comparative example it was only 0.70. This invention uses an evolutionary confidence level determination mechanism to quantitatively verify the recognition results. If the confidence level does not reach the threshold, the result is returned for re-enhancement, avoiding the output of low-confidence results. Simultaneously, it compares the temporal texture change patterns with the standard evolution rate, making the recognition results interpretable and traceable. The overall reliability is far higher than some technical solutions without verification mechanisms and existing conventional methods, providing a stable and reliable basis for field plant protection decisions.

[0106] IV. Core Technology Aspects in Improving Recognition Performance

[0107] Example 2, lacking bidirectional dynamic masking, visual-language anchor guidance, and evolutionary confidence determination, showed significantly lower performance across all metrics compared to Example 1. This demonstrates that dynamic masking for noise reduction, visual-language bimodal alignment, and evolutionary confidence verification are crucial for improving accuracy and reliability. The comparative example, employing single-frame static recognition, non-temporal, and multimodal fusion, exhibited the worst performance, proving that temporal image sequences and pest evolution characteristics are key to overcoming the limitations of traditional single-frame recognition. The synergistic effect of the complete technical solution in this invention achieves full-process optimization of feature extraction, modal fusion, result verification, and model iteration, making performance improvement inevitable.

[0108] Fifth, this invention has stronger practicality and scenario adaptability.

[0109] Example 1 can simultaneously output the name of pests and diseases, their damage level, evolutionary stage, causal analysis, and targeted control solutions. Simultaneously, through data feedback, the model can be iteratively updated online, continuously adapting to different regions, planting conditions, crops, and new pest and disease identification scenarios. Its generalization ability and long-term practicality are significantly superior to existing technologies that only output categories and lack iterative capabilities. Experimental results fully demonstrate that the method of this invention is more in line with the actual needs of precision plant protection in smart agriculture and has high value for widespread application.

[0110] In summary, the visual language-based crop pest and disease identification method of this invention is significantly superior to existing conventional identification methods and some technical solutions in terms of identification accuracy, ability to distinguish similar pests and diseases, reliability of results, practicality and generalization. It effectively solves the technical defects of traditional methods such as reliance on static features, simple modality fusion, easy confusion of similar pests and diseases, unreliable results, and inability to iteratively optimize, fully achieving the purpose of the invention and fully verifying all the beneficial effects of this invention.

[0111] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying crop diseases and pests based on visual language, characterized in that, Includes the following steps: S100, Temporal Image Acquisition and Temporal Texture Set Construction: At fixed time intervals, multiple frames of RGB images are continuously acquired from the same crop monitoring site to construct a temporal image sequence of diseases and pests containing information on the growth and changes of diseases and pests. The temporal image sequence is then subjected to inter-frame registration, illumination normalization, and enhancement processing of lesions and insect damage areas. A set of texture features that continuously change in the temporal dimension is extracted to form a temporal texture set. S200, Bidirectional dynamic masking and mask temporal feature map generation: Bidirectional dynamic masking is performed on the temporal texture set, with the previous frame texture as the forward constraint and the next frame texture as the backward constraint, to calculate a mask temporal feature map that retains only the dynamic change features of pests and diseases and removes the static background and occlusion noise in the field. S300, Visual Language Anchor Point Library Construction and Bimodal Feature Encoding: A visual language anchor point library containing descriptions of the temporal evolution of pests and diseases, texts of the damage stages, keywords of texture changes, and prevention and control schemes is pre-constructed. A language encoder is used to encode the text anchor point information into language-guided feature vectors, and a visual encoder is used to encode the mask temporal feature map into visual temporal feature vectors. S400, Temporal Anchor Alignment and Coarse Identification of Pests and Diseases: Using language-guided feature vectors as supervision signals, temporal anchor alignment mapping is performed on visual temporal feature vectors. Temporal texture distance and mask matching degree are calculated respectively. Weighted comprehensive similarity is obtained by fusing according to preset weights. Coarse identification of pests and diseases is completed based on weighted comprehensive similarity. S500, Pest and Disease Evolution Deduction and Confidence Determination: Based on the coarse recognition results and the temporal texture change rate, the probability of pest and disease evolution stages is deduced, the evolution confidence is calculated, and the evolution confidence is compared with a preset threshold. If it is higher than the threshold, it is determined as the final recognition category; if it is lower than the threshold, it returns to step S2 to perform temporal texture enhancement again. S600, Structured Text Output and Recognition Result Presentation: Based on the final recognition category and corresponding evolutionary stage, the system matches and outputs structured text from the visual language anchor point library, including the name of the pest or disease, the level of damage, the evolutionary stage, the cause analysis, and the targeted prevention and control plan.

2. The method for identifying crop diseases and pests based on visual language according to claim 1, characterized in that, The formula for calculating the mask weight in the bidirectional dynamic masking process described in step S200 is as follows: in, For the first Frame mask weights, For the first Frame texture feature map, For the first Frame texture feature map, For the first Frame texture feature map, Forward time series weighting coefficients, satisfying .

3. The method for identifying crop diseases and pests based on visual language according to claim 1, characterized in that, The temporal texture distance mentioned in step S400 is used to measure the temporal variation pattern of pests and diseases, and the calculation formula is as follows: in, For the first Frame visual temporal feature vectors For the first Frame visual temporal feature vectors This represents the total number of frames in the time-series image sequence. The smaller the value, the more regular the temporal changes in the patterns of pests and diseases.

4. The method for identifying crop diseases and pests based on visual language according to claim 1, characterized in that, The mask matching degree mentioned in step S400 is used to measure the degree of matching between the language guidance features and the mask visual features, and the calculation formula is as follows: in, for Frame language-guided feature vectors For the first Frame mask weights.

5. The method for identifying crop diseases and pests based on visual language according to claim 1, characterized in that, The weighted comprehensive similarity mentioned in step S400 is obtained by fusing the mask matching degree and the temporal texture distance, and the calculation formula is as follows: in, Let be the mask matching degree weight coefficient, satisfying .

6. The method for identifying crop diseases and pests based on visual language according to claim 1, characterized in that, The confidence level of pest and disease evolution mentioned in step S5 is calculated by combining the weighted comprehensive similarity and the evolution rate deviation, as shown in the formula: in, This represents the current rate of change in the texture of pests and diseases. This represents the standard evolution rate of the pest / disease. A higher value indicates a more reliable recognition result.

7. The method for identifying crop diseases and pests based on visual language according to claim 6, characterized in that, In step S500, the evolution confidence threshold is set. ,when The final recognition result is output in time. Then return to step S2 to re-execute bidirectional dynamic masking and temporal texture enhancement.

8. The method for identifying crop diseases and pests based on visual language according to claim 1, characterized in that, The visual language anchor library adopts a hierarchical structure, with one pest category corresponding to multiple different evolutionary stages, and one evolutionary stage corresponding to a set of temporal texture description text, hazard level labels, and standardized prevention and control measures.

9. The method for identifying crop diseases and pests based on visual language according to claim 1, characterized in that, After the structured text is output in step S600, the temporal texture features, evolutionary confidence, recognition results and actual observation results of this recognition are stored in the visual language anchor point library. The language encoder and visual encoder are updated online incrementally to form a self-iterative optimization recognition closed loop.

Images