Rural landscape style intelligent evaluation method and system based on model optimization and migration

Abstract

The invention provides a rural landscape style and appearance intelligent evaluation method and system based on two-stage model optimization and migration, and the method comprises the steps: employing a large-scale multi-modal model as a simulation expert to generate initial evaluation data for a rural landscape image through a two-stage method, constructing a domain data set with domain knowledge enhancement, and carrying out the optimization and migration of a large-scale multi-modal model; fine tuning is carried out on the pre-trained multi-modal large model on a domain data set, so that the model can learn and internalize professional knowledge, evaluation dimensions and detailed standards of rural landscape style and appearance evaluation, and then an intelligent evaluation system is constructed based on the fine-tuned multi-modal large model. Therefore, an artificial intelligence system with professional rural landscape style and appearance evaluation ability can be constructed at low cost and high efficiency, and intelligent and efficient evaluation of the rural landscape style and appearance is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of landscape evaluation, and specifically relates to a rural landscape intelligent evaluation method and system based on two-stage model optimization and migration. Background Art

[0002] Unprecedented demands are being placed on the planning, construction, management, and improvement of rural living environments and landscapes. Traditional rural landscape evaluation relies heavily on manual field surveys, expert experience, and subjective scoring, resulting in inefficiency, high costs, inconsistent standards, and difficulty in large-scale implementation. An objective, efficient, and intelligent evaluation tool is urgently needed to assist in this effort. While some automated or semi-automated landscape evaluation methods currently exist, these methods generally lack the depth and accuracy required for specialized evaluations, or are expensive to implement, making them difficult to apply to rural landscape evaluation tasks.

[0003] For example, one type of existing method is a landscape evaluation method based on traditional computer vision. This type of method usually extracts the underlying visual features of the image, such as color, texture, edges, vegetation index, etc., and combines them with machine learning algorithms (such as SVM and random forest) for classification or regression scoring. "Visual Landscape Quality Evaluation of Scenic Roads Based on Human Eye Field Image Recognition" discloses such a landscape evaluation method based on traditional computer vision. This method extracts different elements through image segmentation, calculates the visual feature parameters of each element, and inputs pre-trained classifiers to obtain evaluation grades or scores. However, this type of method has difficulty understanding the deep semantics and aesthetic concepts of images, such as complex dimensions such as style coordination and cultural identity; and has poor adaptability to evaluation standards. Changes in standards require redesigning features and training models.

[0004] Another existing method is a method that uses a dedicated model trained on a manually annotated dataset. "Using deeplearning to quantify the beauty of outdoor places" discloses such a method, in which a large image database containing detailed manual annotations (such as the degree of building damage, garbage exposure, green coverage, etc.) is established and a deep neural network is trained to perform end-to-end evaluation. In this method, the cost of constructing the dataset is extremely high and time-consuming and labor-intensive. Especially for rural landscapes that require detailed and multi-dimensional evaluation, the acquisition of high-quality annotated data is a huge bottleneck. At the same time, the generalization ability of the model may be limited by the coverage of the annotated data.

[0005] In recent years, Transformer-based large language models (LLMs) have made breakthrough progress, and further development has led to the development of multimodal large language models (MLLMs) capable of simultaneously understanding both images and text, such as the Qwen-VL series. These models demonstrate powerful image-text understanding and generation capabilities in general scenarios. However, when general MLLMs are directly applied to specialized fields such as rural landscape assessment, the depth, professionalism, and meticulousness of their evaluations often fail to meet specific needs, requiring targeted optimization and the infusion of domain knowledge.

[0006] As mentioned above, traditional computer vision methods struggle to capture high-level semantic and aesthetic information. The high cost and time required to construct high-quality, large-scale, domain-specific annotated datasets hinder the development and application of specialized models. Directly applying general-purpose MLLMs to specialized domains lacks depth and accuracy, and lacks domain-specific capabilities. Therefore, there is currently a lack of an effective method for efficiently and cost-effectively transferring the capabilities of large, general-purpose models to specialized domains, such as rural landscape assessment. Summary of the Invention

[0007] This invention aims to address the low degree of automation, poor objectivity, and low efficiency of existing rural landscape assessment methods, as well as the difficulties in injecting domain knowledge, the high cost of acquiring high-quality domain data, and the lack of professionalism in model evaluation when applying advanced multimodal large models. The purpose is to provide a method and system that can efficiently generate high-quality domain evaluation data and effectively fine-tune pre-trained multimodal large models based on this data, thereby achieving automated, multi-dimensional, and professional intelligent evaluation of rural landscapes. Specifically, the present invention adopts the following technical solutions:

[0008] The present invention provides a rural landscape intelligent evaluation method based on two-stage model optimization and migration, which has the following technical features: step S1, based on a predetermined rural landscape evaluation index system, using a large multimodal model to generate evaluation data for rural landscape images, thereby constructing an evaluation data set enhanced with domain knowledge; step S2, using the evaluation data set to fine-tune the parameters of the multimodal large model; step S3, based on the fine-tuned multimodal large model, constructing a rural landscape intelligent evaluation system, and using the system to generate intelligent evaluation and display results for the rural landscape images to be evaluated.

[0009] The intelligent rural landscape evaluation method based on two-stage model optimization and migration provided by the present invention may also have such technical features, wherein step S1 includes the following sub-steps: step S1-1, defining the rural landscape evaluation index system, which includes multiple evaluation dimensions; step S1-2, designing prompts for the large multimodal model based on the rural landscape evaluation index system; step S1-3, collecting a batch of representative rural landscape images, and pre-processing each of the rural landscape images as sample images; step S1-4, combining the sample images with the prompts and inputting them into the large multimodal model, so that the large multimodal model generates a preliminary text evaluation report for the sample images; step S1-5, extracting evaluation information of each evaluation dimension based on the text evaluation report, and performing data cleaning and formatting on the evaluation information to obtain formatted evaluation data; step S1-6, converting the formatted evaluation data into a dialogue format suitable for fine-tuning the multimodal large model, and constructing the evaluation dataset based on the sample images and the converted formatted evaluation data.

[0010] The intelligent evaluation method for rural landscape features based on two-stage model optimization and migration provided by the present invention may also have such technical features, wherein step S3 includes the following sub-steps: step S3-1, loading the fine-tuned multimodal large model and configuring an inference framework for it; step S3-2, based on the input and output of the multimodal large model, constructing a user interaction interface, at least for allowing users to upload the rural landscape images to be evaluated and displaying the evaluation results to the users; step S3-3, based on the input and output of the multimodal large model and the user interaction interface, constructing the back-end implementation logic.

[0011] The intelligent evaluation method for rural landscape features based on two-stage model optimization and migration provided by the present invention may also have such technical features, wherein, in step S1-1, the rural landscape features evaluation index system includes the following evaluation dimensions: building maintenance status, environmental cleanliness, landscape plant configuration, style coordination, facilities and functions, and historical and cultural identification.

[0012] The intelligent evaluation method for rural landscape style based on two-stage model optimization and migration provided by the present invention may also have such technical features, wherein each of the evaluation dimensions includes multiple evaluation indicators, the evaluation indicators of the building maintenance status include the cleanliness of the facade and the structural integrity, the evaluation indicators of the environmental cleanliness include road cleanliness and hardening, and the maintenance of public spaces, the evaluation indicators of the landscape plant configuration include green plant coverage and plant configuration beauty, the evaluation indicators of the style coordination include color coordination and material and style unity, the evaluation indicators of the facilities and functions include the completeness of the infrastructure and the completeness of the sign system, and the evaluation indicators of the history and cultural identity include the preservation of historical elements and the traditional pattern.

[0013] The intelligent rural landscape assessment method based on two-stage model optimization and migration provided by the present invention may also have the following technical features: in step S1-2, the prompt includes a system prompt and a user prompt. The system prompt is used to assign the role of rural landscape review expert to the large-scale multimodal model and inform the evaluation criteria and output format requirements. The output format requirements are to output a text evaluation report including the following content: the score of each evaluation indicator, the evaluation reason, the overall score, the overall rating, and a comprehensive review. The user prompt includes image information of the rural landscape image to be evaluated and evaluation instructions. In step S1-4, after combining the sample image with the prompt, the application programming interface of the large-scale multimodal model is batch called programmatically.

[0014] The intelligent rural landscape assessment method based on two-stage model optimization and migration provided by the present invention may also have the following technical features: In step S1-5, structured information is extracted from the text evaluation report, including the scores of each evaluation indicator, the evaluation reasons, the overall score, the overall rating, and the comprehensive review. The structured information is then cleaned and formatted to obtain the formatted evaluation data. In step S1-6, the converted formatted evaluation data includes system role information, user role information, assistant role information, and image fields.

[0015] The intelligent evaluation method for rural landscape features based on two-stage model optimization and migration provided by the present invention may also have such technical features, wherein, in step S3-3, the back-end implementation logic includes: constructing an input message structure suitable for the multimodal large model based on the input rural landscape image, and inputting it into the multimodal large multimodal model for reasoning; obtaining the text evaluation report generated by the multimodal large multimodal model, and displaying it to the user on the user interaction interface.

[0016] The intelligent evaluation method for rural landscape features based on two-stage model optimization and migration provided by the present invention may also have such technical features, wherein, in step S2, the evaluation data set is used to fine-tune the multimodal large model through a low-rank adaptive method.

[0017] The present invention provides an intelligent evaluation system for rural landscape features, which has the following technical features. It is constructed through the above-mentioned intelligent evaluation method for rural landscape features based on two-stage model optimization and migration. The system at least includes: an input display module for allowing users to input rural landscape images to be evaluated and displaying the generated evaluation results to the user; and an intelligent evaluation generation module, including the fine-tuned multimodal large model, for generating the evaluation results for the rural landscape images to be evaluated.

[0018] Functions and effects of the invention

[0019] The method and system for intelligent rural landscape assessment based on two-stage model optimization and migration provided by the present invention can construct an artificial intelligence system with professional rural landscape assessment capabilities at a low cost and high efficiency through a two-stage approach. The constructed system has the following advantages:

[0020] Reduce the cost and difficulty of data acquisition: By using large multimodal models to assist in generating high-quality initial evaluation data, the reliance on expensive and time-consuming manual expert annotation is significantly reduced, making it more cost-effective to build fine-tuned datasets for the rural landscape field.

[0021] Improve the professionalism and accuracy of evaluation: By fine-tuning the pre-trained multimodal model on a specific domain dataset, the model can learn and internalize the professional knowledge, evaluation dimensions and detailed standards of rural landscape evaluation. Its evaluation results are more in-depth, accurate and in line with actual needs than general models.

[0022] Achieve efficient and automated landscape evaluation: The entire evaluation process is triggered by users uploading pictures, and the model automatically completes analysis and report generation, greatly improving the efficiency of rural landscape evaluation and supporting large-scale, rapid surveys and assessments.

[0023] Enhance the objectivity and consistency of evaluation: Evaluation based on a unified model and preset standards reduces the deviation caused by human subjective factors and ensures the relative objectivity and consistency of the evaluation results.

[0024] Good scalability and adaptability: The evaluation criteria and dimensions can be adjusted and expanded by modifying the system prompts and fine-tuning the evaluation data set, so that the system can adapt to the rural landscape evaluation needs of different regions and types.

[0025] Provide decision-making support: It can provide quantitative, structured evaluation results and detailed analysis for rural planners, managers, researchers, etc., and provide a scientific decision-making basis for rural landscape planning, design optimization, governance improvement and other work. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 Flowchart of a rural landscape intelligent evaluation method based on two-stage model optimization and migration in an embodiment of the present invention;

[0027] Figure 2 is a flow chart of step S1 in an embodiment of the present invention;

[0028] Figure 3 is a flow chart of step S3 in an embodiment of the present invention;

[0029] Figure 4 This is a structural block diagram of the rural landscape intelligent evaluation system according to an embodiment of the present invention;

[0030] Figure 5 This is an example of a user interaction interface in an embodiment of the present invention. Figure 1 ;

[0031] Figure 6 This is an example of a user interaction interface in an embodiment of the present invention. Figure 2 . DETAILED DESCRIPTION

[0032] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following is a detailed explanation of the rural landscape intelligent evaluation method and system based on two-stage model optimization and migration of the present invention in combination with embodiments and drawings.

[0033] <Example>

[0034] Figure 1 This is a flow chart of the rural landscape intelligent evaluation method based on two-stage model optimization and migration in this embodiment.

[0035] like Figure 1 As shown, the rural landscape intelligent evaluation method based on two-stage model optimization and migration in this embodiment mainly includes the following steps:

[0036] Step S1, constructing a domain evaluation dataset: Based on a predefined rural landscape evaluation index system, a large multimodal model is used to generate domain evaluation data for rural landscape images, thereby constructing an evaluation dataset enhanced with domain knowledge.

[0037] Step S2, efficient parameter fine-tuning: perform efficient parameter fine-tuning on the multimodal large model based on the evaluation dataset.

[0038] Step S3, intelligent evaluation and display: construct an intelligent evaluation system for rural landscape features based on the fine-tuned multimodal large model, and use the system to generate intelligent evaluation and display results for the rural landscape images to be evaluated.

[0039] The above steps will be described in detail below.

[0040] Step S1 (first stage), constructing a domain evaluation dataset: Based on a predefined rural landscape evaluation index system, a large multimodal model is used to generate domain knowledge-rich and structured evaluation data for a batch of rural landscape images, thereby constructing a domain knowledge-enhanced evaluation dataset.

[0041] Figure 2 This is a flow chart of step S1 in this embodiment.

[0042] like Figure 2 As shown, step S1 specifically includes the following sub-steps:

[0043] Step S1-1, defining a detailed rural landscape evaluation index system.

[0044] For example, the rural landscape evaluation index system can include six evaluation dimensions, each of which contains multiple evaluation indicators: 1) Building maintenance status, such as facade cleanliness and structural integrity; 2) Environmental cleanliness, such as road cleanliness and hardening, and public space maintenance; 3) Landscape plant configuration, such as green plant coverage and plant configuration beauty; 4) Style coordination, such as color coordination, material and style unity; 5) Facilities and functions, such as infrastructure completeness and sign system completeness; 6) Historical and cultural identity, such as the preservation of historical elements and traditional patterns.

[0045] Step S1-2: Based on the rural landscape evaluation index system, prompts are designed for the selected large multimodal model, including system prompts and user prompts.

[0046] Among them, large multimodal models are existing powerful models, such as OpenAI GPT-4o, which can be used as a "simulated expert" to give domain-knowledge-rich evaluations of rural landscape images.

[0047] In step S1-2, system prompts and user prompts are designed for the OpenAI GPT-4o API. The system prompt assigns the large multimodal model the role of "rural landscape review expert" and explicitly informs it of the evaluation criteria and output format requirements based on the rural landscape evaluation index system. For example, the format requirements may include scoring each evaluation indicator in the rural landscape evaluation index system, and providing the evaluation reason, overall score, overall rating, and summary (comprehensive review). The user prompt contains the image information of the rural landscape image to be evaluated and specific evaluation instructions.

[0048] In step S1-3, a batch of representative rural landscape images are collected and each image is preprocessed as a sample image. For example, each image is converted into a format suitable for API input through preprocessing. Preprocessing can include operations such as cropping, compression, and transcoding.

[0049] In step S1-4, the sample image is combined with the prompt designed in step S1-2 and input into the large multimodal model, so that the large multimodal model generates a preliminary text evaluation report for each sample image based on the sample image and the prompt.

[0050] Among them, after combining the sample images with the prompts, the API interface of the general large-scale multimodal model is called in batches through programming to generate evaluation reports in batches.

[0051] Step S1-5: extracting various evaluation information based on the preliminary text evaluation report, and performing data cleaning and formatting on the extracted evaluation information to obtain formatted evaluation data.

[0052] The design parsing logic automatically extracts structured information from the text evaluation reports returned by the large multimodal model, including the scores of each evaluation indicator, the reasons for the evaluation, the overall score, the overall rating, and the comprehensive comments. The ratings can include, for example, four levels: excellent, good, fair, and poor.

[0053] The extracted structured information is then cleaned to remove irrelevant characters and standardized. A manual review process is introduced to proofread and correct low-quality or parsed data generated by large multimodal models to improve the quality of the dataset. Finally, the cleaned, formatted, and proofread structured evaluation data is saved as JSON or other standard formats.

[0054] Step S1-6: Convert the formatted evaluation data into a specific dialogue format suitable for fine-tuning of a downstream multimodal large model, and construct an evaluation dataset for model fine-tuning based on the sample images and the converted formatted evaluation data.

[0055] Each piece of data suitable for fine-tuning a large multimodal model typically contains:

[0056] System role message: defines the role played by the multimodal large model and the evaluation criteria followed.

[0057] User role message: contains image placeholders (such as ) and user evaluation requests (such as "What is the rural landscape style reflected in this picture").

[0058] Assistant role message: Contains detailed evaluation report text generated and structured by a large multimodal model.

[0059] Image field: points to the actual sample image file path.

[0060] Through the above steps, a JSONL format evaluation dataset for model fine-tuning is finally generated.

[0061] Step S2 (second stage), efficient parameter fine-tuning: perform efficient parameter fine-tuning on the multimodal large model based on the evaluation dataset.

[0062] That is, the domain knowledge-enhanced evaluation dataset generated in the first stage is used to fine-tune a pre-trained multimodal large model to enable it to acquire professional evaluation capabilities for rural landscape features.

[0063] Specifically, a pre-trained multimodal large model with superior performance is selected as the base model, such as Qwen2.5-VL-7B. This base model is then fine-tuned using efficient parameter tuning techniques, such as Low-Rank Adaptation (LoRA). This adapts the knowledge and capabilities of the large multimodal model to the specific task of rural landscape assessment, using minimal computational resources and time.

[0064] Before starting the fine-tuning process, configure the fine-tuning framework and related parameters. The fine-tuning framework can be, for example, Llama-Factory. Related parameters include: base model path, custom dataset path and name (i.e., pointing to the domain evaluation dataset generated in step S1), fine-tuning type (i.e., using the LoRA method), LoRA method-specific parameters, and training parameters. Training parameters include learning rate, batch size, training rounds, optimizer, etc.

[0065] The fine-tuning process can then be initiated, where the base model learns how to generate a detailed assessment report that meets predetermined criteria based on the input rural landscape image and assessment request. The fine-tuned multimodal large model is stored, thus preserving the LoRA adapter weights.

[0066] Step S3, intelligent evaluation and display: construct an intelligent evaluation system for rural landscape features based on the fine-tuned multimodal large model, and use the system to generate intelligent evaluation and display results for the rural landscape images to be evaluated.

[0067] Figure 3 This is a flow chart of step S3 in this embodiment.

[0068] like Figure 3 As shown, step S3 specifically includes the following sub-steps:

[0069] Step S3-1: load the fine-tuned multimodal large model and configure an inference framework for it to improve inference efficiency.

[0070] For example, an efficient reasoning framework such as vLLM can be adopted, and the sampling parameters of the multimodal large model can be configured to control the quality of the generated text.

[0071] Step S3-2: Based on the input and output of the multimodal large model, a user interaction interface is constructed for users to upload rural landscape images to be evaluated and the evaluation results are displayed to the users.

[0072] For example, a user-friendly interactive interface can be built using a web UI framework such as Gradio. The user interface allows users to perform the following operations: upload the rural landscape image to be evaluated, trigger the evaluation, view the evaluation results output by the multimodal large model, and clear the input and output.

[0073] Step S3-3: Build the backend implementation logic based on the input and output of the multimodal large model and the user interaction interface.

[0074] Among them, the back-end implementation logic includes: constructing an input message structure suitable for the multimodal large model based on the uploaded rural landscape images, that is, constructing an input message structure consistent with that during fine-tuning, inputting it into the multimodal large model for reasoning, and obtaining the text evaluation report generated by the multimodal large model, and displaying it to the user in the front-end user interaction interface.

[0075] Figure 4 This is a structural block diagram of the rural landscape intelligent evaluation system in this embodiment.

[0076] like Figure 4 As shown, this embodiment also provides a rural landscape intelligent evaluation system 10 corresponding to the above method, which includes an input display module 11, an input processing module 12, an intelligent evaluation generation module 13, an output processing module 14 and a general control module 15 for controlling the operation of the above modules.

[0077] Among them, the input display module 11 is used to provide the above-mentioned user interaction interface for users to upload rural landscape images to be evaluated, start the intelligent evaluation generation process, and display the rural landscape intelligent evaluation results generated for the uploaded images to users.

[0078] The input processing module 12 is used to process the images uploaded by the user and the evaluation requests input, including converting the images into a format suitable for the multimodal large model, generating corresponding prompts, thereby constructing an input message structure consistent with that during fine-tuning, and passing it to the intelligent evaluation generation module 13.

[0079] The intelligent evaluation generation module 13 includes the above-mentioned fine-tuned multimodal large model and reasoning framework, and is used to receive input messages from the input processing module 12, and generate a text evaluation report based on the received input message, and pass it to the output processing module 14.

[0080] The output processing module 14 is used to receive the output of the model from the intelligent evaluation generation module 13, convert the output into a standardized scoring format, and pass it to the input display module 11 for display. The input display module 11 displays the corresponding evaluation results to the user.

[0081] Figure 5 This is an example of the user interaction interface in this embodiment. Figure 1 , showing the user interaction interface in the initial state; Figure 6 This is an example of the user interaction interface in this embodiment. Figure 2 , shows the user interaction interface after the user uploads the image and starts the evaluation.

[0082] like Figure 5 and Figure 6 As shown, illustratively, the user interaction interface 20 may include prompt text information, an image upload area 21 , an evaluation result display area 22 , a start evaluation button 23 , and a reset button 24 .

[0083] The image upload area 21 is used for users to upload rural landscape images to be evaluated. The area may contain multiple function buttons. Users can click the corresponding function button in the area according to the prompt information and select a local image to upload, or they can drag and drop the local image to the area for upload according to the prompt information. Figure 6 As shown, this area is also used to display a preview of the uploaded image after the image is successfully uploaded for user verification.

[0084] The start evaluation button 23 is for the user to click to start the intelligent evaluation process.

[0085] The evaluation result display area 22 is used to display the intelligent evaluation results to the user, such as Figure 6As shown, the evaluation results are organized according to the six evaluation dimensions mentioned above. For each evaluation dimension, the scores and evaluation reasons of each item under that dimension are displayed respectively, and the overall score and comprehensive comments are displayed at the end.

[0086] The reset button 24 is for the user to click to clear the image upload area 21 and the evaluation result display area 22 so as to re-upload the image and start the next smart evaluation.

[0087] Functions and Effects of the Embodiments

[0088] The method and system for intelligent rural landscape assessment based on two-stage model optimization and migration provided in this embodiment can construct an artificial intelligence system with professional rural landscape assessment capabilities at a low cost and high efficiency through the two-stage approach. The constructed system has the following advantages:

[0089] Reduce the cost and difficulty of data acquisition: By using large multimodal models to assist in generating high-quality initial evaluation data, the reliance on expensive and time-consuming manual expert annotation is significantly reduced, making it more cost-effective to build fine-tuned datasets for the rural landscape field.

[0090] Improve the professionalism and accuracy of evaluation: By fine-tuning the pre-trained multimodal large model on a specific domain dataset, the model can learn and internalize the professional knowledge, evaluation dimensions and detailed standards of rural landscape evaluation. Its evaluation results are more in-depth, accurate and in line with actual needs than general models.

[0091] Achieve efficient and automated landscape evaluation: The entire evaluation process is triggered by users uploading pictures, and the model automatically completes analysis and report generation, greatly improving the efficiency of rural landscape evaluation and supporting large-scale, rapid surveys and assessments.

[0092] Enhance the objectivity and consistency of evaluation: Evaluation based on a unified model and preset standards reduces the deviation caused by human subjective factors and ensures the relative objectivity and consistency of the evaluation results.

[0093] Good scalability and adaptability: The evaluation criteria and dimensions can be adjusted and expanded by modifying the system prompts and fine-tuning the evaluation data set, so that the system can adapt to the rural landscape evaluation needs of different regions and types.

[0094] Provide decision-making support: It can provide quantitative, structured evaluation results and detailed analysis for rural planners, managers, researchers, etc., and provide a scientific decision-making basis for rural landscape planning, design optimization, governance improvement and other work.

[0095] In the embodiment, the system development structure is clear, the degree of modularity is high, and the subsequent maintenance and function expansion are convenient. At the same time, the libraries and frameworks that the system relies on are all mainstream open source tools with good community support and portability.

[0096] Furthermore, in the embodiment, the defined rural landscape evaluation index system includes six evaluation dimensions, each dimension contains multiple clear evaluation indicators. Through the designed system prompts, the large multimodal model scores each evaluation indicator and gives specific reasons for the scoring. Therefore, the evaluation of rural landscape is comprehensive, the evaluation standards are clear, and it has good interpretability.

[0097] Furthermore, the system provides a user-friendly interactive interface. Users can trigger the intelligent analysis and evaluation process by clicking to select a rural landscape image and clicking a start button. The operation is convenient and intuitive, and basically requires no additional learning cost, which has great practical application value.

[0098] The above embodiments are merely illustrative of specific implementations of the present invention, and the present invention is not limited to the scope of the description of the above embodiments. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention as claimed. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A rural landscape intelligent evaluation method based on two-stage model optimization and migration, characterized by: The following steps are involved: Step S1, based on a predetermined rural landscape evaluation index system, using a large multimodal model to generate evaluation data for rural landscape images, thereby constructing an evaluation dataset enhanced with domain knowledge; Step S2, fine-tuning the parameters of the multimodal large model using the evaluation dataset; Step S3: constructing a rural landscape intelligent evaluation system based on the fine-tuned multimodal large model, and using the system to generate intelligent evaluation and display results for the rural landscape image to be evaluated.

2. The rural landscape intelligent evaluation method based on two-stage model optimization and migration according to claim 1, Its characteristics are: Wherein, step S1 includes the following sub-steps: Step S1-1, defining the rural landscape evaluation index system, which includes multiple evaluation dimensions; Step S1-2, designing prompts for the large multimodal model based on the rural landscape evaluation index system; Step S1-3, collecting a batch of representative rural landscape images, and pre-processing each of the rural landscape images as a sample image; Step S1-4, combining the sample image and the prompt and inputting the combined result into the large multimodal model, so that the large multimodal model generates a preliminary text evaluation report for the sample image; Step S1-5, extracting evaluation information of each evaluation dimension based on the text evaluation report, and performing data cleaning and formatting on the evaluation information to obtain formatted evaluation data; Step S1-6: convert the formatted evaluation data into a conversation format suitable for fine-tuning the multimodal large model, and construct the evaluation dataset based on the sample image and the converted formatted evaluation data.

3. The rural landscape intelligent evaluation method based on two-stage model optimization and migration according to claim 2, Its characteristics are: in, Step S3 includes the following sub-steps: Step S3-1, loading the fine-tuned multimodal large model and configuring an inference framework for it; Step S3-2: constructing a user interaction interface based on the input and output of the multimodal large model, at least for allowing the user to upload the rural landscape image to be evaluated and displaying the evaluation results to the user; Step S3-3: construct backend implementation logic based on the input and output of the multimodal large model and the user interaction interface.

4. The rural landscape intelligent evaluation method based on two-stage model optimization and migration according to claim 3 is characterized by: in, In step S1-1, the rural landscape evaluation index system includes the following evaluation dimensions: building maintenance status, environmental cleanliness, landscape plant configuration, style coordination, facilities and functions, and historical and cultural identification.

5. The rural landscape intelligent evaluation method based on two-stage model optimization and migration according to claim 4 is characterized by: in, Each evaluation dimension contains multiple evaluation indicators. The evaluation indicators of building maintenance status include facade cleanliness, structural integrity, The evaluation indicators of environmental cleanliness include road cleanliness and hardening, and public space maintenance. The evaluation indicators of landscape plant configuration include green plant coverage rate, plant configuration aesthetics, The evaluation indexes of style coordination include color coordination, material and style unity, The evaluation indicators of the facilities and functions include the completeness of infrastructure and the completeness of signage systems. The evaluation indicators of historical and cultural identity include the preservation of historical elements and traditional patterns.

6. The rural landscape intelligent evaluation method based on two-stage model optimization and migration according to claim 5 is characterized by: in, In step S1-2, the prompts include system prompts and user prompts. The system prompts the user to assign the role of the rural landscape review expert to the large multimodal model and informs the evaluation criteria and output format requirements. The output format is required to output a text evaluation report including the following contents: the score of each evaluation indicator, the evaluation reason, the overall score, the overall rating, and a comprehensive review. The user prompt includes image information of the rural landscape image to be evaluated and an evaluation instruction, In step S1-4, after combining the sample image with the prompt, the application programming interface of the large multimodal model is called in batches through programming.

7. The rural landscape intelligent evaluation method based on two-stage model optimization and migration according to claim 6 is characterized by: in, In step S1-5, structured information is extracted from the text evaluation report, including the scores of each evaluation indicator, the evaluation reasons, the overall score, the overall rating, and the comprehensive review, and the structured information is cleaned and formatted to obtain the formatted evaluation data. In step S1-6, the converted formatted evaluation data includes system role information, user role information, assistant role information, and image fields.

8. The rural landscape intelligent evaluation method based on two-stage model optimization and migration according to claim 5, Its characteristics are: in, In step S3-3, the backend implementation logic includes: Constructing an input message structure suitable for the multimodal large model based on the input rural landscape image, and inputting the structure into the multimodal large multimodal model for reasoning; The text evaluation report generated by the multimodal large multimodal model is obtained and displayed to the user on the user interaction interface.

9. The rural landscape intelligent evaluation method based on two-stage model optimization and migration according to claim 1 is characterized by: in, In step S2, the evaluation dataset is used to fine-tune the multimodal large model through a low-rank adaptive method.

10. An intelligent rural landscape evaluation system, characterized by: The system is constructed by the rural landscape intelligent evaluation method based on two-stage model optimization and migration according to any one of claims 1 to 9, and comprises at least: an input display module for allowing a user to input a rural landscape image to be evaluated, and displaying a generated evaluation result to the user; and An intelligent evaluation generation module includes the fine-tuned multimodal large model, and is used to generate the evaluation result for the rural landscape image to be evaluated.

Images