Deep learning and ontology combination-based pepper disease identification and knowledge search system and pepper disease intelligent diagnosis method

Abstract

The invention discloses a deep learning and ontology combination-based pepper disease identification and knowledge search system and a pepper disease intelligent diagnosis method. The technical scheme is characterized in that the system comprises a pepper disease ontology construction module, a deep learning image identification module and a search integration module; the method comprises four steps. According to the invention, through deep fusion of the ontology technology and deep learning, a set of pepper disease intelligent identification and knowledge search system is constructed, and key problems in traditional agricultural disease diagnosis are significantly solved. The system realizes cross-modal bidirectional retrieval of image and semantic information, breaks through the limitation of a single-modal technology, and enables fragmented knowledge such as disease characteristics and prevention and control methods to be efficiently integrated and accurately matched. Through a structured knowledge graph and a dynamic optimization mechanism, the system not only improves the accuracy of disease identification and the integrity of information retrieval, but also greatly reduces the user operation threshold, and is especially suitable for inexperienced growers.

Description

Technical Field

[0001] The present invention belongs to the technical field of smart agriculture, and specifically relates to a pepper disease identification and knowledge search system based on the combination of deep learning and ontology, and a pepper disease intelligent diagnosis method. Background Art

[0002] In agricultural production, crop disease identification and prevention are highly dependent on the practical experience of growers. Traditional disease diagnosis methods rely on manual observation of disease phenotypic characteristics (such as color, morphology, texture changes, etc.). Although they have been proven to be reliable through long-term practice, their learning process is complex and easily limited by subjective experience, making them difficult to popularize among ordinary growers. With the development of digital technology, agricultural image acquisition and data processing capabilities have been significantly improved, but how to achieve accurate mapping of crop disease phenotypic characteristics with growers' experiential knowledge is still a technical bottleneck restricting the development of intelligent agriculture.

[0003] In the existing technology, image recognition models based on deep learning (such as YOLOv5, etc.) can efficiently complete the visual detection task of crop diseases, but their output results are mostly single labels or probability values, lacking deep association with semantic information such as disease mechanisms and prevention and control strategies. On the other hand, ontology theory provides a structured expression framework for agricultural knowledge engineering, which can systematically organize knowledge such as disease names, feature descriptions, and prevention and control methods, but its application is mostly limited to text retrieval scenarios and fails to form an effective linkage with image data. The semantic description of diseases (such as features, prevention and control methods) lacks effective association with image data, resulting in low search efficiency.

[0004] In addition, existing image recognition systems mostly focus on a single modality and are unable to simultaneously meet the needs of bidirectional retrieval of text and images. In recent years, deep learning technologies (such as YOLOv5) have performed well in agricultural image classification, but their results lack interpretability and cannot be directly associated with semantic knowledge. On the other hand, although ontology technology can build a structured knowledge system, it is difficult to process unstructured image data. Therefore, how to bridge the gap between images and semantics and realize intelligent diagnosis of "searching for text with images and searching for images with text" has become a technical problem that needs to be solved urgently in this field.

[0005] To this end, a pepper disease recognition and knowledge search system based on the combination of deep learning and ontology and an intelligent diagnosis method for pepper diseases were proposed. Summary of the invention

[0006] In view of the problems mentioned in the background art, the purpose of the present invention is to provide a pepper disease recognition and knowledge search system based on the combination of deep learning and ontology, and a pepper disease intelligent diagnosis method. The system realizes cross-modal diagnosis and prevention and control information retrieval of pepper diseases by integrating semantic knowledge graphs and image recognition technologies, and provides efficient and accurate disease management solutions for agricultural growers to solve the problems mentioned in the background art.

[0007] The above technical objectives of the present invention are achieved through the following technical solutions:

[0008] A pepper disease recognition and knowledge search system based on the combination of deep learning and ontology, comprising:

[0009] A pepper disease ontology construction module for structurally organizing the semantic data of pepper diseases based on ontology theory to form a knowledge system including disease names, characteristics, and prevention and control methods;

[0010] A deep learning image recognition module that uses the YOLOv5 model to detect and classify pepper disease images and extract disease phenotype characteristics;

[0011] A search integration module that associates the semantic data of the ontology construction module with the feature data of the image recognition module, and uses the HMS-Rank hybrid modal search sorting function to dynamically balance the text and image modal weights, supporting users to perform cross-modal searches through text or image input and returning the complete information of the diseases;

[0012] A dynamic update module that automatically optimizes the classification accuracy of the ontology knowledge graph and the deep learning model based on the DKG-IO dynamic knowledge graph incremental optimization algorithm according to user feedback and newly collected disease data.

[0013] In the above pepper disease recognition and knowledge search system based on the combination of deep learning and ontology, among them: the pepper disease ontology construction module includes:

[0014] A semantic annotation unit that defines the semantic attributes of diseases by expert annotating the disease data collected from pepper planting bases;

[0015] An ontology relationship modeling unit that establishes logical associations between diseases based on color, shape, and epidermal change characteristics to form an extensible pepper disease ontology knowledge graph, and realizes dynamic correction of node weights and association relationships through the DKG-IO dynamic knowledge graph incremental optimization algorithm.

[0016] In the above pepper disease recognition and knowledge search system based on the combination of deep learning and ontology, among them: the deep learning image recognition module includes:

[0017] A data set construction unit, which constructs a training data set based on clustered pepper disease images, including disease images and their annotated labels;

[0018] The model optimization unit adjusts the YOLOv5 network parameters through transfer learning to adapt to the small sample recognition scenario of pepper diseases.

[0019] The above-mentioned pepper disease identification and knowledge search system based on deep learning and ontology, wherein: the search integration module uses the following method to achieve cross-modal search:

[0020] When querying text, semantic keywords are matched through the ontology knowledge graph to return relevant disease information and associated images;

[0021] When querying images, feature vectors are extracted through the deep learning model, and similarity matching is performed with disease features in the knowledge base. The weighted scores are calculated based on the HMS-Rank hybrid modal search ranking function, and the corresponding semantic information is returned.

[0022] The HMS-Rank formula is:

[0023] Score(q,d)=λ·Stext(q,d)+(1-λ)·Simage(q,d);

[0024]

[0025] in:

[0026] λ is a dynamic weight ranging from 0 to 1. The dynamic weight λ is adaptively adjusted according to the input quality confidence;

[0027] Stext is the similarity score between the text query and the knowledge base semantics;

[0028] Simage is the similarity score between image features and knowledge base features;

[0029] Conftext / image is the confidence of the text or image input, generated by the input quality assessment model;

[0030] T is the temperature coefficient, which controls the smoothness of weight adjustment;

[0031] q is the search content entered by the user;

[0032] d is the disease information entry or document in the knowledge base;

[0033] Stext(q,d) is the similarity score between query q and document d in text mode;

[0034] Simage(q,d) is the similarity score between query q and document d in image modality.

[0035] 5. The chili disease identification and knowledge search system based on the combination of deep learning and ontology according to claim 1, characterized in that: the dynamic update module optimizes the knowledge graph through the following steps:

[0036] Quantify the confidence of user feedback

[0037] Update the node weight wnew = wold + γ·Cf·Δw, and trigger the modification of the knowledge graph structure, where:

[0038] wold / new is the old / new weight of the knowledge graph node;

[0039] γ is the learning rate;

[0040] Cf is the feedback confidence;

[0041] Δw is the weight increment.

[0042] The above-mentioned chili disease identification and knowledge search system based on the combination of deep learning and ontology, wherein: the chili disease ontology knowledge graph includes the following levels:

[0043] Disease category layer, defining the common disease types of clustered upright chili peppers;

[0044] Feature description layer, refining the color, shape, and disease spot distribution characteristics of each disease;

[0045] Prevention and control method layer, associating diseases with their corresponding pesticide use and environmental regulation measures.

[0046] The above-mentioned chili disease identification and knowledge search system based on the combination of deep learning and ontology, wherein: the training data set contains at least 5,000 disease images annotated by plant protection experts, covering more than 10 common disease types of clustered upright chili peppers.

[0047] The above-mentioned chili disease identification and knowledge search system based on the combination of deep learning and ontology, wherein: the interactive interface of the search integration module supports the following functions:

[0048] Multi-modal input: allowing users to upload images or input text descriptions for query;

[0049] Result display: displaying the disease name, feature description, prevention and control methods, and similar case images in a combination of text and pictures, and dynamically sorting the search results through the HMS-Rank hybrid modal search sorting function.

[0050] The test and verification of the chili disease identification and knowledge search system based on the combination of deep learning and ontology include:

[0051] Evaluate the performance of the image recognition module using accuracy and recall rate, and the accuracy of the test set is not less than 95%;

[0052] Verify the usability and ease of use of the search results through user satisfaction surveys.

[0053] The present invention also provides a method for intelligent diagnosis of pepper diseases in a pepper disease recognition and knowledge search system based on the combination of deep learning and ontology, including the following steps:

[0054] S1. Collect pepper disease images and semantic data, and construct an ontology knowledge base and a training data set;

[0055] S2. Perform real-time detection and classification of disease images through a deep learning model;

[0056] S3. According to the text or image input by the user, cross-modal match the knowledge base data and return the diagnosis result through the HMS-Rank hybrid modal search sorting function;

[0057] S4. Recommend prevention and control measures based on the diagnosis results, and update the knowledge base through the DKG-IO dynamic knowledge graph incremental optimization algorithm to optimize the subsequent search accuracy.

[0058] In summary, the present invention mainly has the following beneficial effects:

[0059] Through the deep integration of ontology technology and deep learning, the present invention constructs a set of intelligent pepper disease recognition and knowledge search systems, which significantly solves the key problems in traditional agricultural disease diagnosis. The system realizes cross-modal bidirectional retrieval of image and semantic information, breaks through the limitations of single-modal technology, and enables efficient integration and precise matching of fragmented knowledge such as disease characteristics and prevention methods. Through the structured knowledge graph and dynamic optimization mechanism, the system not only improves the accuracy of disease recognition and the integrity of information retrieval, but also greatly reduces the user operation threshold, especially suitable for inexperienced growers. In addition, the modular design endows the system with strong expansion ability, which can adapt to different crop disease scenarios, provides a replicable and sustainable technical solution for agricultural intelligent management, and has broad practical value and promotion potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is an overall schematic diagram of the pepper disease recognition and knowledge search system based on the combination of deep learning and ontology of the present invention;

[0061] Figure 2 is a workflow diagram of the pepper disease recognition and knowledge search system based on the combination of deep learning and ontology of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0062] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0063] Embodiment 1

[0064] Reference Figure 1 , this embodiment provides a pepper disease recognition and knowledge search system based on the combination of deep learning and ontology, including: a pepper disease ontology construction module, a deep learning image recognition module, a search integration module, and a dynamic update module, where:

[0065] The pepper disease ontology construction module is used to structurally organize the semantic data of pepper diseases based on ontology theory to form a knowledge system including disease names, characteristics, and prevention and control methods;

[0066] The deep learning image recognition module uses the YOLOv5 model to detect and classify pepper disease images and extract disease phenotype characteristics;

[0067] The search integration module associates the semantic data of the ontology construction module with the feature data of the image recognition module, supports users to perform cross-modal searches through text or image input, and returns the complete information of the disease.

[0068] The search integration module associates the semantic data of the ontology construction module with the feature data of the image recognition module, adopts a hybrid modal search ranking function (HMS-Rank) to dynamically balance the weights of text and image modalities, supports users to perform cross-modal searches through text or image input, and returns the complete information of the disease;

[0069] The dynamic update module automatically optimizes the classification accuracy of the ontology knowledge graph and the deep learning model based on the dynamic knowledge graph incremental optimization algorithm (DKG-IO) according to user feedback and newly collected disease data. The dynamic update module automatically optimizes the knowledge base and the model through user feedback. The accuracy of the system is significantly improved after iteration, and the technological advancement is maintained in the long term.

[0070] By integrating ontology construction and deep learning technologies, the system realizes cross-modal searches (mutual text and image searches) for pepper diseases, solves the problem of the separation of semantic and image information in traditional methods, and significantly improves the diagnostic efficiency and the comprehensiveness of information obtained by users.

[0071] Specifically, in this embodiment, the pepper disease ontology construction module includes a semantic annotation unit and an ontology relationship modeling unit, where:

[0072] The semantic annotation unit defines the semantic attributes of diseases by expertly annotating the disease data collected from the chili pepper planting base.

[0073] The ontology relationship modeling unit establishes the logical associations among diseases based on color, shape, and epidermal change characteristics, forms an extensible ontology knowledge graph of chili pepper diseases, and dynamically corrects the node weights and association relationships through the DKG-IO dynamic knowledge graph incremental optimization algorithm.

[0074] Semantic annotation and ontology relationship modeling structure the fragmented disease knowledge into an extensible knowledge graph, significantly improving the retrieval speed of disease characteristics and supporting the dynamic addition of new disease types.

[0075] Specifically, in this embodiment, the deep learning image recognition module includes:

[0076] The dataset construction unit constructs a training dataset based on the images of clustered and upward-pointing chili pepper diseases, including disease images and their annotation labels.

[0077] The model optimization unit adjusts the YOLOv5 network parameters through transfer learning to adapt to the small-sample recognition scenario of chili pepper diseases.

[0078] The optimized model (transfer learning) based on small-sample training solves the problem of scarce agricultural image data. The model achieves a recognition accuracy of 96.2% with only 5000 images for training and covers more than 10 common diseases.

[0079] Specifically, in this embodiment, the search integration module implements cross-modal search using the following method:

[0080] When querying text, semantic keywords are matched through the ontology knowledge graph, and relevant disease information and associated images are returned.

[0081] When querying images, feature vectors are extracted through the deep learning model, similarity matching is performed with the disease characteristics in the knowledge base, and the weighted score is calculated based on the HMS-Rank hybrid modal search ranking function, and the corresponding semantic information is returned.

[0082] The HMS-Rank formula is as follows:

[0083] Score(q,d)=λ·Stext(q,d)+(1-λ)·Simage(q,d);

[0084]

[0085] Where λ is the dynamic weight, ranging from 0 to 1, representing the contribution ratio of the text and image modalities.

[0086] Stext is the similarity score between the text query and the knowledge base semantics (based on cosine similarity);

[0087] Simage is the similarity score between the image features and the knowledge base features (based on Euclidean distance);

[0088] Conftext / image is the confidence of the text or image input, generated by the input quality assessment model (e.g., image clarity score, text keyword coverage);

[0089] T is the temperature coefficient, which controls the smoothness of weight adjustment (default value 0.5). The larger the value, the smoother the weight change;

[0090] q is the search content input by the user, which can be a text description (such as "leaves have black spots") or an uploaded disease image, representing the query request to be matched and is the starting point of the search;

[0091] d is the disease information entry (document) in the knowledge base, which contains structured data such as disease name, feature description, prevention and control methods, and associated images, representing the target object to be matched and is the end point of the search;

[0092] Stext(q,d) is the similarity score between the query q and the document d in the text modality;

[0093] Simage(q,d) is the similarity score between the query q and the document d in the image modality;

[0094] The dynamic weight λ is adaptively adjusted according to the input quality confidence.

[0095] For example: The user uploads a blurred diseased leaf image (Confimage = 0.3) and enters the text "leaves have black spots" (Conftext = 0.8);

[0096] Calculated: The system preferentially relies on text search (weight 73%);

[0097] Scenario 1: The user uploads a blurred diseased leaf image (Confimage = 0.9), but there is no text input. At this time, λ = 0, and it completely relies on image matching.

[0098] Technical effects:

[0099] Adaptive search: Dynamically allocate modality weights according to the input quality, and the relevance of the search results is significantly improved;

[0100] Enhanced fault tolerance: When there is a blurred image or blurred text input, the system automatically reduces the impact of the low-quality modality;

[0101] Response efficiency: The calculation time of the weighted score is short, meeting the requirements of real-time interaction.

[0102] The cross-modal search mechanism allows users to query either by text or image, significantly shortening the system response time and remarkably improving the matching rate of search results, which can meet the needs of different user scenarios.

[0103] Among them, the dynamic update module optimizes the knowledge graph through the following steps:

[0104] Quantify the confidence of user feedback

[0105] Update the node weight wnew = wold + γ·Cf·Δw and trigger the modification of the knowledge graph structure, where:

[0106] wold / new is the old / new weight of the knowledge graph node, reflecting the importance of the node;

[0107] γ is the learning rate (default 0.05), controlling the weight update rate;

[0108] Cf is the feedback confidence (0 - 1), quantifying the credibility of user feedback (for example: expert feedback is regarded as "valid", and the same feedback from ordinary users needs to be counted 3 times as 1 valid time);

[0109] Δw is the weight increment, determined by the feedback type (such as the feedback of "the control method is ineffective" corresponding to Δw = -0.1, and adding a new solution corresponding to Δw = +0.2);

[0110] For example:

[0111] Scenario 1: The user repeatedly feedbacks that "the recommended pesticide for anthrax is ineffective" (10 total feedbacks, 8 valid feedbacks), then Cf = 0.8.

[0112] Node weight update: w new = 0.6 + 0.05 × 0.8 × (-0.1) = 0.596, triggering the associated alternative solution (such as "azoxystrobin").

[0113] Scenario 2: An expert adds a control method for "disease-resistant varieties" (Δw = +0.2), and the node weight is increased from 0.4 to 0.42, and this solution is preferentially displayed.

[0114] Technical effects:

[0115] Self-optimization of the knowledge base: The update efficiency of node weights is significantly improved, effectively reducing the manual maintenance cost;

[0116] Error correction: The weights of invalid nodes are gradually reduced, and the proportion of error information is decreased;

[0117] Dynamic expansion: Newly added nodes (such as new disease types) automatically inherit the association relationship, adapting to the iterative needs of agricultural knowledge;

[0118] Among them, HMS-Rank and DKG-IO work together:

[0119] Improve the accuracy of results through dynamic weighting during the search phase;

[0120] In the feedback phase, the knowledge base is optimized through weight correction, forming a closed loop of "search-feedback-optimization";

[0121] HMS-Rank introduces the temperature coefficient (TT) balance weight to adjust the sensitivity;

[0122] DKG-IO quantifies the feedback type (Δw) to implement a differentiated update strategy.

[0123] For example, in the deployment of a pepper planting base, after a user uploaded a blurred image, HMS-Rank set the text weight to 0.8 and returned the result of "anthracnose". When the user reported "diagnosis error", the DKG-IO dynamic knowledge graph incremental optimization algorithm reduced the weight of the node and associated it with the "disease" node. After one month of operation, the system's misdiagnosis rate dropped from 8% to 3%.

[0124] Specifically, in this embodiment, the pepper disease ontology knowledge graph includes the following levels:

[0125] The disease category layer defines the common disease types of cluster pepper;

[0126] The feature description layer refines the color, shape, and spot distribution characteristics of each disease;

[0127] The control method layer links diseases with their corresponding pesticide use and environmental control measures.

[0128] The layered ontology knowledge graph (disease category-characteristics-prevention and control methods) makes the knowledge organization logic clear, improves the information completeness of user query results by 40%, and reduces the risk of misdiagnosis.

[0129] Specifically, in this embodiment, the training data set contains at least 5,000 disease images annotated by plant protection experts, covering more than 10 common diseases of clustered peppers. Large-scale expert-annotated data sets (≥5,000 images, covering 10 diseases) ensure the generalization ability of the model, and the recognition accuracy fluctuation in cross-regional testing is less than 3%.

[0130] Specifically, in this embodiment, the interactive interface of the search integration module supports the following functions:

[0131] Multimodal input: Allow users to upload images or enter text descriptions for queries;

[0132] Result display: The disease name, feature description, prevention and control methods, and images of similar cases are displayed in a combination of text and graphics, and the search results are dynamically sorted through the HMS-Rank hybrid modality search and sorting function.

[0133] The multi-modal interaction interface reduces the operation threshold for users. The display in a combination of text and graphics greatly improves the information understanding efficiency, especially suitable for growers with limited educational levels.

[0134] The test and verification of the pepper disease identification and knowledge search system based on the combination of deep learning and ontology include:

[0135] The performance of the image recognition module is evaluated using accuracy and recall rates, and the accuracy rate of the test set is not less than 95%.

[0136] The practicality and usability of the search results are verified through user satisfaction surveys.

[0137] Verified through two indicators of accuracy rate (≥95%) and user satisfaction, the system has both technical reliability and practical application value, and the user reuse rate has been greatly improved.

[0138] Example 2

[0139] Reference Figure 2 , this embodiment provides an intelligent diagnosis method for pepper diseases of a pepper disease identification and knowledge search system based on the combination of deep learning and ontology, including the following steps:

[0140] S1. Collect pepper disease images and semantic data, and construct an ontology knowledge base and a training data set;

[0141] S2. Real-time detect and classify the disease images through a deep learning model;

[0142] S3. According to the text or image input by the user, cross-modal match the knowledge base data and return the diagnosis result through the HMS-Rank hybrid modality search and sorting function;

[0143] S4. Recommend prevention and control measures based on the diagnosis result, and update the knowledge base through the DKG-IO dynamic knowledge graph incremental optimization algorithm to optimize the subsequent search accuracy.

[0144] The intelligent diagnosis method standardizes the data collection, model training, and user interaction processes, realizes the closed-loop management from disease identification to prevention and control recommendations, and significantly improves the overall prevention and control efficiency.

[0145] Advantages compared with the existing technology

[0146] Innovation: For the first time, ontology and deep learning are combined to achieve cross-modal search for agricultural diseases;

[0147] Practicality: Adapt to complex agricultural scenarios through dynamic optimization and multi-modal interaction;

[0148] Scalability: The modular design supports extension to the diagnosis of other crop diseases, with great application potential.

[0149] In summary, the present invention constructs a smart pepper disease identification and knowledge search system by deeply integrating ontology technology and deep learning, significantly solving the key problems in traditional agricultural disease diagnosis. The system realizes cross-modal bidirectional retrieval of image and semantic information, breaks through the limitations of single-modal technology, and enables efficient integration and precise matching of fragmented knowledge such as disease characteristics and prevention methods. Through the structured knowledge graph and dynamic optimization mechanism, the system not only improves the accuracy of disease identification and the integrity of information retrieval, but also greatly reduces the user operation threshold, especially suitable for inexperienced growers. In addition, the modular design endows the system with strong expansion ability, which can adapt to different crop disease scenarios, providing a replicable and sustainable technical solution for agricultural intelligent management, and having extensive practical value and promotion potential.

[0150] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made in these embodiments without departing from the principle and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A chili disease recognition and knowledge search system based on the combination of deep learning and ontology, characterized in that: Including: A chili disease ontology construction module, which is used to structurally organize the semantic data of chili diseases based on ontology theory to form a knowledge system including disease names, characteristics, and prevention and control methods; A deep learning image recognition module, which uses the YOLOv5 model to detect and classify chili disease images and extract disease phenotype characteristics; A search and integration module, which associates the semantic data of the ontology construction module with the feature data of the image recognition module, dynamically balances the text and image modality weights using the HMS-Rank hybrid modality search and ranking function, supports users to perform cross-modal searches through text or image input, and returns the complete information of the disease; A dynamic update module, which automatically optimizes the classification accuracy of the ontology knowledge graph and the deep learning model based on the DKG-IO dynamic knowledge graph incremental optimization algorithm according to user feedback and newly collected disease data.

2. The chili disease recognition and knowledge search system based on the combination of deep learning and ontology according to claim 1, wherein: The chili disease ontology construction module includes: A semantic annotation unit, which defines the semantic attributes of diseases by expert annotating the disease data collected from chili planting bases; An ontology relationship modeling unit, which establishes logical associations between diseases based on color, shape, and epidermal change characteristics to form an extensible chili disease ontology knowledge graph, and dynamically corrects the node weights and association relationships through the DKG-IO dynamic knowledge graph incremental optimization algorithm.

3. The chili disease identification and knowledge search system based on the combination of deep learning and ontology according to claim 1, characterized in that: The deep learning image recognition module includes: A dataset construction unit, which constructs a training dataset based on the images of clustered and upward-pointing chili diseases, including disease images and their annotation labels; A model optimization unit, which adjusts the YOLOv5 network parameters through transfer learning to adapt to the small-sample recognition scenario of chili diseases.

4. The chili disease recognition and knowledge search system based on the combination of deep learning and ontology according to claim 1, characterized in that: The search and integration module implements cross-modal search using the following method: When querying by text, semantic keywords are matched through the ontology knowledge graph, and relevant disease information and associated images are returned; When querying by image, feature vectors are extracted through the deep learning model, similarity matching is performed with the disease characteristics in the knowledge base, and weighted scores are calculated based on the HMS-Rank hybrid modality search and ranking function, and the corresponding semantic information is returned; The HMS-Rank formula is: ； ； Where: λ is the dynamic weight, with a range of 0 to 1, and the dynamic weight λ is adaptively adjusted through the input quality confidence; is the similarity score between the text query and the semantics of the knowledge base; is the similarity score between the image features and the knowledge base features; is the confidence of the text or image input, generated by the input quality assessment model; T is the temperature coefficient, which controls the smoothness of weight adjustment; q is the search content input by the user; d is the disease information entry or document in the knowledge base; is the similarity score of query q and document d under the text modality; is the similarity score between the query q and the document d in the image modality.

5. The pepper disease identification and knowledge search system based on the combination of deep learning and ontology according to claim 1, characterized in that: The dynamic update module optimizes the knowledge graph through the following steps: ； Update node weights , and trigger the knowledge graph structure correction, where: is the old / new weight of the knowledge graph node; γ is the learning rate; C f is the feedback confidence; Δw is the weight increment.

6. The chili disease recognition and knowledge search system based on the combination of deep learning and ontology according to claim 2, characterized in that: The chili disease ontology knowledge graph includes the following levels: Disease category layer, which defines the common disease types of clustered and upward-pointing chili; Feature description layer, which refines the color, shape, and disease spot distribution characteristics of each disease; Prevention and control method layer, which associates diseases with their corresponding pesticide use and environmental regulation measures.

7. The chili disease recognition and knowledge search system based on the combination of deep learning and ontology according to claim 3, characterized in that: The training dataset contains at least 5,000 disease images annotated by plant protection experts, covering more than 10 common disease types of clustered and upward-pointing chili.

8. The pepper disease identification and knowledge search system based on the combination of deep learning and ontology according to claim 1, characterized in that: The interaction interface of the search and integration module supports the following functions: Multi-modal input: Allows users to upload images or input text descriptions for querying; Result display: Display the disease name, feature description, control methods, and images of similar cases in the form of combination of text and pictures, and dynamically sort the search results through the HMS-Rank hybrid modality search and sorting function.

9. The pepper disease identification and knowledge search system based on the combination of deep learning and ontology according to claim 1, characterized in that: The test and verification of the pepper disease recognition and knowledge search system based on the combination of deep learning and ontology include: Use accuracy and recall to evaluate the performance of the image recognition module, and the accuracy of the test set is not less than 95%; Verify the practicability and usability of the search results through user satisfaction surveys.

10. The intelligent diagnosis method for pepper diseases of the pepper disease recognition and knowledge search system based on the combination of deep learning and ontology according to any one of claims 1-9, characterized in that, It includes the following steps: S1. Collect pepper disease images and semantic data, and construct an ontology knowledge base and a training data set; S2. Real-time detect and classify disease images through a deep learning model; S3. According to the text or image input by the user, cross-modally match the knowledge base data and return the diagnostic results through the HMS-Rank hybrid modality search and sorting function; S4. Recommend control measures based on the diagnostic results, and update the knowledge base through the DKG-IO dynamic knowledge graph incremental optimization algorithm to optimize the subsequent search accuracy.

Images