Self-learning feedback calibration method and system for an image generation model

Abstract

This invention proposes a self-learning feedback calibration method and system for image generation models, belonging to the field of image generation technology. The method includes: collecting relevant data in the field of land spatial planning; performing structured processing on the relevant data, converting it into a computer-recognizable coded form, and constructing a domain knowledge coding library; selecting a basic image generation model, using LoRA fine-tuning technology, and combining it with the constructed domain knowledge coding library to fine-tune the parameters of the basic image generation model; and generating an initial land spatial planning image dataset based on the fine-tuned image generation model and given initial input conditions. Through the integration of LoRA fine-tuning technology with the ControlNet control network, this method can continuously improve the quality of the image generation model. Combined with the fusion processing of multimodal data, the generated images better meet the actual needs of land spatial planning.

Description

Technical Field

[0001] This invention proposes a self-learning feedback calibration method and system for image generation models, belonging to the field of image generation technology. Background Technology

[0002] Image generation technology has great potential in land and space planning to intuitively present planning schemes, assist in decision-making and optimization, and enhance public participation.

[0003] However, its application faces many challenges. Territorial spatial planning involves complex and specialized knowledge in many aspects such as geographic information, land use, ecological protection, and transportation layout, requiring the generated images to be highly accurate, professional, and practical.

[0004] Most of the current mainstream basic image generation models are general-purpose. Although these models can generate images in a wide range of scenarios, they lack in-depth understanding and learning of specific knowledge in the field of land and space planning. As a result, the generated images are significantly lacking in detail accuracy, compliance with professional standards, and practical operability, making it difficult to meet the actual needs of planning.

[0005] Furthermore, land spatial planning involves a large amount of multimodal data, and effectively integrating this data and extracting key features to construct control vectors is a major technical challenge. At the same time, the planning process is dynamic, requiring image generation models to have real-time adjustment and optimization capabilities to adapt to constantly changing planning conditions.

[0006] Therefore, this invention proposes a self-learning feedback calibration method and system for image generation models. By introducing domain knowledge, multimodal data fusion, ControlNet control network, and real-time parameter correction, it aims to improve the quality and adaptability of image generation models in land spatial planning. Summary of the Invention

[0007] This invention provides a self-learning feedback calibration method and system for image generation models to solve the problems mentioned in the background section above: This invention proposes a self-learning feedback calibration method for an image generation model, the method comprising: S1. Collect relevant data in the field of land and space planning, perform structured processing on the relevant data, convert it into a computer-recognizable encoding form, and construct a domain knowledge encoding library; select a basic image generation model, use LoRA fine-tuning technology, and combine it with the domain knowledge encoding library constructed above to fine-tune the parameters of the basic image generation model; based on the fine-tuned image generation model, generate an initial land and space planning image dataset according to the given initial input conditions. S2. Acquire multimodal data related to land spatial planning, fuse the multimodal data, extract key features, and construct multimodal conditional control vectors; integrate the ControlNet control network into the image generation model, and train and configure the ControlNet using the multimodal conditional control vectors; further control the initially generated land spatial planning image through ControlNet to obtain an enhanced image dataset; S3. Define feedback metrics for evaluating the quality of generated images, and use a pre-trained evaluation model to evaluate the enhanced image dataset; obtain comprehensive feedback data for each image; and construct the pose evaluation matrix of the image generation model based on the comprehensive feedback data. S4. Perform sensitivity analysis on the key parameters of the image generation model, and design a real-time parameter self-calibration algorithm based on the pose evaluation matrix and parameter sensitivity analysis results; use the self-calibration algorithm to dynamically optimize and iterate the image generation model. S5. Select a portion of land spatial planning scenarios as a test set, and generate images using a dynamically optimized image generation model; compare the generated images with actual scenarios or standard images designed by professional designers to verify the calibration results.

[0008] This invention proposes a self-learning feedback calibration system for an image generation model, comprising: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0009] The beneficial effects of this invention are as follows: By integrating LoRA fine-tuning technology with the ControlNet control network, this method can continuously improve the quality of the image generation model. Combined with the fusion processing of multimodal data, the generated images better meet the actual needs of land spatial planning.

[0010] By analyzing the feedback from the generated images in real time, and using the pose evaluation matrix and parameter sensitivity analysis, not only can problems be identified, but key parameters can also be dynamically adjusted through a self-correction algorithm, thereby ensuring that the quality of the generated images is continuously improved.

[0011] The dynamic optimization process enables image generation models to adapt to different input conditions and domain requirements. Over time, the model can continuously optimize based on new feedback, further improving its adaptability and flexibility.

[0012] By combining quantitative and qualitative evaluation of the generated images, this method can not only check the quality of images through objective standards, but also provide more human-cognitive feedback with the help of expert evaluation, thus ensuring the accuracy and effectiveness of the calibration process.

[0013] By setting iteration termination conditions and a continuous optimization mechanism, this method ensures the long-term optimization capability of the model, and can continuously improve the accuracy and practicality of image generation. Attached Figure Description

[0014] Figure 1 This is a diagram illustrating the steps of the method described in this invention; Figure 2 As described in this invention Figure 1 A diagram illustrating the steps of the S2 method. Detailed Implementation

[0015] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0016] One embodiment of the present invention, such as Figure 1 As shown, a self-learning feedback calibration method for an image generation model includes: S1. Collect relevant data in the field of land and space planning, including architectural style standards, topographic features, and planning layout principles; perform structured processing on the relevant data, convert it into a computer-recognizable coded form, and construct a domain knowledge coding library; select a basic image generation model, use LoRA (Low-Rank Adaptation) fine-tuning technology, and combine it with the constructed domain knowledge coding library to fine-tune the parameters of the basic image generation model, so that it initially has the ability to generate images that conform to the characteristics of the land and space planning field; based on the fine-tuned image generation model, according to the given initial input conditions, including the basic information of the planning area and the planning objectives, generate an initial land and space planning image dataset; S2. In addition to the initial input conditions, acquire multimodal data related to land spatial planning, including satellite remote sensing images, geographic information system (GIS) data, and field survey photos; fuse the multimodal data, extract key features, and construct multimodal conditional control vectors; integrate the ControlNet control network into the image generation model, and train and configure the ControlNet using the multimodal conditional control vectors; further control the initially generated land spatial planning images using ControlNet to obtain an enhanced image dataset; S3. Define feedback metrics for evaluating the quality of generated images; evaluate the enhanced image dataset using a pre-trained evaluation model; obtain comprehensive feedback data for each image; construct the pose evaluation matrix of the image generation model based on the comprehensive feedback data. S4. Perform sensitivity analysis on the key parameters of the image generation model to determine which parameters have a significant impact on the quality of the generated images. These key parameters include the learning rate, batch size, and network layer parameters. Based on the pose evaluation matrix and the parameter sensitivity analysis results, design a real-time parameter self-calibration algorithm. This algorithm automatically adjusts the values ​​of key parameters based on feedback data from the currently generated images to optimize model performance. For example, if the style consistency index is low, the algorithm can appropriately adjust style-related network layer parameters; if the detail accuracy index is unsatisfactory, the algorithm can adjust parameters such as the learning rate to improve the model's ability to capture details. The self-calibration algorithm is used to dynamically optimize and iterate the image generation model. After each image generation, parameters are adjusted promptly based on feedback data, and then the image is generated again and evaluated until the quality of the generated image reaches the preset standard or the maximum number of iterations is reached. Through continuous iteration, the model gradually adapts to the needs of the land spatial planning field, achieving dynamic optimization. S5. Select a representative set of land spatial planning scenarios as a test set, and generate images using a dynamically optimized image generation model. Compare the generated images with actual scenarios or standard images designed by professional designers to verify the validity and accuracy of the calibration results. Verification is conducted using a combination of objective indicators (such as structural similarity index, peak signal-to-noise ratio, etc.) and subjective evaluation (scores by experts). After verification, if the calibration results meet the requirements, the final image generation model is packaged and output. This model can be applied to actual land spatial planning projects, providing planners with high-quality image generation services that meet the needs of the field, and assisting in planning decisions and design work.

[0017] The working principle and effects of the above technical solution are as follows: By introducing a domain knowledge coding library and fusing multimodal data, the model can generate high-quality images that conform to the characteristics of land and space planning, improving the accuracy and detail of the images; through dynamic optimization iteration based on self-correction algorithms, the model can automatically adjust key parameters according to feedback data, reducing the frequency and complexity of manual adjustments, improving the automation and intelligence level of the model, and reducing the need for manual intervention. By combining the ControlNet control network with multimodal conditional control vectors, the model is enhanced in terms of style consistency and detail accuracy, making the generated images more in line with expectations in terms of style, detail, and planning rationality; thus enhancing the control over image style and detail representation. Through real-time parameter self-calibration, the image generation model can be continuously optimized based on feedback, improving the model's adaptability and flexibility in different planning scenarios and meeting the needs of different land and space planning projects. By comprehensively using multiple evaluation indicators such as style consistency, detail accuracy, and planning rationality, and combining expert scores and objective indicators for verification, the quality of the generated images is ensured to better meet actual needs, thus improving the scientific nature and accuracy of model evaluation. The self-calibrated image generation model can generate images that more closely resemble real-world planning scenarios, thus providing planners with more accurate image generation services and supporting the implementation of planning decisions. By providing high-quality image generation results that meet the needs of land and space planning, it can effectively assist planners in working more efficiently during the design and decision-making process, improving the overall quality and decision-making efficiency of land and space planning projects.

[0018] By using automated feedback calibration and dynamic optimization, the model can achieve optimization iteration in a shorter time, reducing the time cost of traditional manual training and adjustment, and improving the training efficiency of the model.

[0019] In one embodiment of the present invention, S1 includes: S11. Obtain relevant data in the field of land and space planning, including architectural style norms (such as classical, modern, and regional architectural styles), topographic features (such as natural geographical elements such as mountains, plains, and water bodies), and planning layout principles (such as functional zoning, traffic flow, and ecological protection); and perform structured processing on the collected relevant data to transform knowledge into computer-recognizable coding forms, such as nodes and edges in a graph database, to construct a domain knowledge coding library; S12. Select an existing image generation model as the base model and conduct a comprehensive evaluation based on model performance and resource consumption; use LoRA (Low-Rank Adaptation) fine-tuning technology, combined with a domain knowledge encoding library, to fine-tune the parameters of the base model; S13. Based on the actual needs of land and space planning, set initial input conditions, including basic information of the planning area (geographical location, area, surrounding environment, etc.) and planning objectives (residential, commercial, industrial, ecological, etc.); based on the fine-tuned image generation model, generate an initial land and space planning image dataset according to the given initial input conditions, and annotate the generated images, including information such as building type, terrain features, and planning layout.

[0020] The working principle and effect of the above technical solution are as follows: by structuring relevant data in the field of land and space planning and converting it into a computer-recognizable encoding form, such as nodes and edges in a graph database, the efficiency of data management and retrieval is significantly improved, ensuring the accurate transmission and application of knowledge.

[0021] By constructing a domain knowledge coding library, we can efficiently integrate and manage various complex data such as architectural styles, terrain features, and planning layouts, reducing the complexity of manual processing and integration. By using LoRA fine-tuning technology in conjunction with the domain knowledge coding library to optimize the basic image generation model, the model can better adapt to the characteristics of the land spatial planning field, improving the accuracy and relevance of image generation.

[0022] By setting initial input conditions based on the actual needs of land and space planning, image datasets that accurately meet planning requirements can be generated, improving the practicality and operability of planning and design. By annotating the generated images with information such as building type, terrain features, and planning layout, the interpretability and usability of the image dataset are increased, ensuring that the images not only have visual representation but also contain detailed planning information, facilitating subsequent analysis and application.

[0023] The fine-tuned image generation model can generate customized images according to specific planning needs, providing planners with more personalized and high-precision planning image data, which helps improve the quality and efficiency of planning and design. By comprehensively evaluating the performance and resource consumption of existing image generation models, the most suitable base model can be selected for fine-tuning. This reduces resource consumption and optimization time costs while ensuring image quality, thus improving the overall system's operating efficiency.

[0024] The fine-tuned model can generate images that meet the requirements according to different planning needs, improving the model's adaptability and flexibility in different planning scenarios.

[0025] One embodiment of the present invention, such as Figure 2 As shown, S2 includes: S21. Acquire multimodal data related to land and space planning, including satellite remote sensing images, geographic information system (GIS) data, field survey photos, and drone aerial videos; and preprocess the collected multimodal data. S22. Based on the multimodal fusion network in deep learning, data from different modalities are fused to extract key features; the fused features are represented as vectors to construct a multimodal conditional control vector, which contains information about the image's style, details, layout, and other aspects. S23. Integrate the ControlNet control network into the image generation model to control the image generation process according to the given condition vector; and use multimodal conditional control vectors to train and configure ControlNet so that it can accurately capture the features of the land spatial planning field. S24. Further control the initially generated land spatial planning image using ControlNet to enhance the image's performance in terms of style and detail, resulting in an enhanced image dataset.

[0026] The working principle and effects of the above technical solution are as follows: By combining multimodal data, it is possible to capture more comprehensive multi-dimensional information of land spatial planning, improving the efficiency and accuracy of data processing. By using a multimodal fusion network in deep learning, different types of data are fused, simplifying the data integration process, effectively extracting key information, and reducing the complexity of manual data processing and analysis.

[0027] Based on multimodal conditional control vectors, the image generation process can be precisely controlled, ensuring that the generated images meet the specific details and style required by the planning, thus improving the customization and accuracy of the planned images. By integrating the ControlNet control network, the image generation process can be controlled more finely according to the given conditional vectors, enhancing the flexibility and operability of the image generation process.

[0028] By leveraging ControlNet to further control and optimize the initially generated images, the images' performance in terms of style, detail, and layout is enhanced, resulting in more realistic and visually impactful images. Applying multimodal conditional control vectors to ControlNet training enables the model to quickly adapt and generate images that meet requirements, thereby reducing the computational resources and time costs required during training and optimization to some extent.

[0029] With the help of the control network, images with different styles and layouts can be generated according to different national spatial planning needs, which significantly improves the diversity and flexibility of image generation.

[0030] In one embodiment of the present invention, step S22 includes: Preliminary intra-modal feature extraction is performed on the preprocessed multimodal data, including: semantic features are extracted using the BERT model for text modalities, visual features are extracted using the ResNet model for image modalities (e.g., remote sensing images), and topological features are extracted using graph neural networks for vector modalities (e.g., GIS spatial data), to obtain basic feature data for each modality; The basic feature data of each modality are input into the cross-modal interaction layer of the multimodal fusion network. The association weights between different modal features (such as the spatial correspondence weights between text semantic features and remote sensing image visual features) are calculated through the attention mechanism to construct the cross-modal association feature matrix. Based on the cross-modal association feature matrix, deep interactive operations are performed through the feature fusion layer of the multimodal fusion network to extract common key features across modalities that reflect the inherent correlation of multimodal data, and obtain a fused feature set. The importance of the fused feature set is evaluated (e.g., feature contribution is calculated using the GBDT model), low-importance noisy features are filtered out, and principal component analysis (PCA) is used for dimensionality normalization to map the high-dimensional fused features to a low-dimensional feature space, thereby obtaining refined fused feature data. The refined and fused feature data is input into the feature vector transformation layer and mapped into a dense vector of fixed dimensions through a fully connected network. The vector dimension is dynamically adjusted according to the complexity of the land and space planning scenario (for example, a 256-dimensional vector is used for planning in ecologically sensitive areas, and a 512-dimensional vector is used for planning in urban construction areas), thus obtaining a fused feature vector set. By combining the core constraints of territorial spatial planning (such as the proportion of land use types, the threshold of spatial development intensity, and the boundary of ecological protection red lines), the fusion feature vector set is fine-tuned under constraints (for example, by optimizing the dimensional weights of the vectors corresponding to the constraints through gradient descent) to construct a multimodal condition control vector that combines multimodal information and planning constraints.

[0031] The working principle and effect of the above technical solution are as follows: by using different deep learning models to extract features of text, image and vector modalities respectively, the extraction of features of each modality is more efficient and accurate, thereby improving the overall effect of multimodal data processing; through the attention mechanism in the cross-modal interaction layer, the correlation weight between different modalities can be automatically calculated, making the feature fusion between modalities more concise and efficient, and reducing the complexity of manually selecting and matching different modal features.

[0032] By constructing a cross-modal correlation feature matrix, the intrinsic correlation between different modalities can be accurately captured, and key common features across modalities can be extracted, enhancing the model's ability to understand complex relationships between data, thereby improving the representativeness and accuracy of the fusion results. By evaluating the feature importance of the fusion feature set and combining it with principal component analysis for dimensionality normalization, low-importance noise features can be effectively filtered out, thereby improving the efficiency and accuracy of feature selection.

[0033] Principal component analysis (PCA) maps high-dimensional fused features to a low-dimensional feature space, effectively reducing computational burden, improving feature vector processing efficiency, and optimizing resource consumption in subsequent calculations. By dynamically adjusting the vector dimension according to the complexity of the land spatial planning scenario, the generated feature vectors can be adapted to the complexity of different planning scenarios, enhancing the model's flexibility and scalability.

[0034] By combining the core constraints of territorial spatial planning and fine-tuning the integrated feature vector set, the feature vectors generated by the model can better meet the actual planning needs, accurately capture features related to planning constraints, and improve the control accuracy of the planning scheme.

[0035] By fine-tuning the multimodal conditional control vector, multimodal information can be better integrated with planning constraints, ensuring that the generated land spatial planning image accurately reflects the actual needs and constraints of the plan, thus improving the reliability and practicality of the generated image.

[0036] In one embodiment of the present invention, step S23 includes: The parameters of the backbone feature extraction layer of the image generation model are frozen (preserving the basic feature mapping capability of the pre-trained weights), and only the parameter fine-tuning permission of the upper generation module of the model is opened to obtain the basic generation model architecture adapted to ControlNet integration. We construct a dynamic connection channel between the feature interaction layer of the ControlNet control network and the intermediate feature layer of the image generation model. We achieve gradient isolation between the control network and the generation model through zero convolution initialization (initializing the convolution kernel weights to a zero matrix) to obtain the initial ensemble framework. A modality adapter is embedded in the initial integration framework to convert multimodal conditional control vectors into control feature maps that match the feature dimensions of the image generation model (e.g., converting 256-dimensional control vectors into 64×64×8 spatial feature tensors), thus establishing a spatial mapping relationship between conditional vectors and the generation process. A ternary training dataset is constructed, consisting of multimodal conditional control vectors, corresponding planning target images, and spatial constraint masks (such as ecological protection red line masks and urban development boundary masks). The spatial constraint masks convert vector constraints into pixel-level control signals through morphological operations. The design includes a hierarchical loss function, including control feature matching loss (calculating the cosine distance between the control feature map and the intermediate features of the generated model), planning element semantic loss (extracting the matching degree between the element semantics of the generated image and the control vector through the U-Net structure), and spatial topology loss (constraining the adjacency relationship of various land uses in the generated image to conform to the planning logic). A two-stage training strategy is adopted. In the first stage, the parameters of the image generation model are frozen, and only the control module of ControlNet is trained to learn the basic control logic. In the second stage, 10% of the parameters of the upper layer of the generation model are unfrozen, and the control network and the generation module are trained together to achieve the coordinated optimization of control signals and generation style. Based on the trained model, a control intensity adjustment mechanism is constructed. By dynamically scaling the weight coefficients of the multimodal conditional control vector (range 0.1-1.0), the differentiated control intensity configuration for different elements (such as traffic networks and green space systems) in the planning image generation process is realized, thus completing the scenario-based deployment of ControlNet.

[0037] The working principle and effect of the above technical solution are as follows: by freezing the backbone feature extraction layer of the image generation model, only the upper generation module is fine-tuned, and the connection between the feature layers is dynamically adjusted through ControlNet integration, so that the generated image more accurately meets the requirements of multimodal conditional control vector and planning; through a phased training strategy, the first stage only trains the ControlNet control module without updating most of the parameters of the generation model, which reduces the amount of computation and training time and optimizes the training efficiency.

[0038] By using a modality adapter, the multimodal conditional control vector is matched with the feature dimensions of the image generation model, effectively establishing a spatial mapping relationship between the control vector and the generation process, and enhancing the influence of the control signal on image generation. By designing a hierarchical loss function, including control feature matching loss, planning element semantic loss and spatial topology loss, it is ensured that the generated image meets the planning requirements while preserving the correctness of the spatial topology and planning elements.

[0039] By employing zero-convolution initialization, gradient isolation between the control network and the generative model is ensured, avoiding interference from invalid control signals in the generation process, thereby optimizing the image generation effect. Through the control intensity adjustment mechanism, the weight coefficients of the control vector can be dynamically adjusted according to the differentiated control requirements of different elements in the planned image generation process, achieving flexible planned image generation.

[0040] By employing a two-stage training strategy, the control network and generation module are progressively optimized, resulting in more stable model training, accelerated convergence, and improved quality of the final generated images. By combining multimodal control vectors with spatial constraint masks, the generated images accurately reflect the different requirements of the planning objectives, satisfying complex planning scenario changes and enhancing the diversity and adaptability of planning image generation.

[0041] In one embodiment of the present invention, S3 includes: S31. Define the feedback indicators; the feedback indicators include: Style consistency index is used to measure the similarity between generated images and typical styles in the field of land spatial planning; for example, by calculating the style distance between generated images and standard style images. The detail accuracy metric is used to evaluate the accuracy of details in an image, including architecture and terrain; for example, by comparing the generated image with detail elements in a real scene or professional design drawing. Planning rationality indicators are used to check whether the images conform to the layout principles and regulatory requirements of the national land spatial planning. For example, whether the functional zoning is reasonable and whether the traffic flow is smooth.

[0042] Other metrics: Depending on actual needs, other feedback metrics can also be defined, such as image sharpness and color matching.

[0043] S32. Using the pre-collected labeled data, train the evaluation model, and calculate the feedback index defined above based on the trained evaluation model. S33. Use the trained evaluation model to evaluate the enhanced image dataset and obtain comprehensive feedback data for each image. S34. Based on the comprehensive feedback data, construct the pose evaluation matrix of the image generation model; the rows of the pose evaluation matrix represent different input conditions or parameter settings, the columns represent different feedback indicators, and the elements in the matrix represent the quality score of the generated image under specific input conditions and parameter settings. S35. Analyze the constructed attitude evaluation matrix to identify the model's performance under different conditions.

[0044] The working principle and effect of the above technical solution are as follows: by defining multi-dimensional feedback indicators (such as style consistency, detail accuracy, planning rationality, etc.), the quality of the generated image can be comprehensively evaluated, enabling the image generation model to generate images that meet planning requirements more accurately; by using the trained evaluation model to automatically calculate feedback indicators, subjective bias in the manual evaluation process is avoided, and the consistency and objectivity of the evaluation results are improved.

[0045] By constructing a pose evaluation matrix for an image generation model, we can systematically analyze the model's performance under different input conditions and parameter settings, gain a deeper understanding of the model's strengths and weaknesses in different scenarios, and thus provide a basis for subsequent optimization. By analyzing the pose evaluation matrix, we can identify the model's performance bottlenecks and advantages under specific conditions, guide the direction of subsequent optimization, and enable the model to adapt to different needs more efficiently in practical applications.

[0046] By evaluating the comprehensive feedback data of each image and combining it with the model's pose evaluation matrix, the problems in image generation can be clearly identified, allowing for targeted improvements and reducing uncertainty in the optimization process. By analyzing the image quality scores under different input conditions or parameter settings, fine-grained adjustments to the image generation model can be achieved, making the model more adaptable and flexible in different application scenarios.

[0047] By calculating feedback metrics such as style consistency, detail accuracy, and planning rationality, a quantitative evaluation basis can be provided for the quality of the generated image, improving the transparency and interpretability of the image generation process and enhancing users' trust in the generated results. By defining multiple feedback metrics that cover multiple dimensions of evaluation such as style, detail, and planning, the performance of the generated image can be comprehensively evaluated, making the image generation model more comprehensive and reliable in practical applications.

[0048] In one embodiment of the present invention, step S4 includes: S41. Perform sensitivity analysis on the key parameters of the image generation model to obtain sensitivity analysis results. The sensitivity analysis results include determining which parameters have a greater impact on the quality of the generated image. The key parameters include learning rate, batch size, network layer parameters, etc. S42. Based on the pose evaluation matrix and parameter sensitivity analysis results, and based on the real-time parameter self-correction algorithm, the values ​​of key parameters are automatically adjusted according to the feedback data of the currently generated image. S43. Utilize a self-calibration algorithm to dynamically optimize and iterate the image generation model; after each image generation, adjust the parameters promptly based on feedback data, and generate the image again for evaluation. S44. Set iteration termination conditions, such as the quality of the generated image reaching a preset standard or the maximum number of iterations being reached; through continuous iteration, the model gradually adapts to the needs of the land and space planning field, achieving dynamic optimization.

[0049] The working principle and effect of the above technical solution are as follows: By performing sensitivity analysis on key parameters, it is possible to identify which parameters have a greater impact on the quality of the generated image, thereby optimizing the adaptability of the model and enabling the model to be adjusted according to actual needs, thus improving the accuracy of image generation; through the real-time parameter self-correction algorithm, the model can automatically adjust the values ​​of key parameters, reducing the complexity of manual parameter tuning, lowering the cost of manual intervention, and improving the overall optimization efficiency.

[0050] By performing real-time dynamic optimization iterations based on feedback data, the model can adapt to new data and requirements in a timely manner, continuously optimize during the generation process, and automatically adjust according to the evaluation results, thereby improving the quality and stability of image generation. By setting iteration termination conditions and adjusting parameters after each image generation, the model can gradually stabilize after multiple iterations, ensuring that the quality of the generated images meets the preset standards and avoiding overtraining or parameter instability.

[0051] By automatically adjusting parameters and stopping iterations at appropriate times, overtraining and unnecessary computation are avoided, thereby saving computational resources and improving training efficiency. Through dynamic optimization iteration, the image generation model can continuously adapt to the ever-changing needs of the field of land spatial planning, ensuring that the generated images meet the latest planning standards and requirements.

[0052] By combining the posture evaluation matrix and sensitivity analysis results, the model can adjust parameters in a timely manner for specific conditions, ensuring a more accurate optimization process. It can also adjust during real-time generation, improving the real-time feedback capability of optimization. By automatically adjusting the values ​​of key parameters, the model can flexibly adjust according to different input conditions and planning requirements, ensuring the generation of high-quality images in diverse planning scenarios.

[0053] In one embodiment of the present invention, S42 includes: The feedback data of the currently generated image is analyzed to extract key quality assessment indicators, including style consistency, detail accuracy, and planning rationality. The extracted key quality assessment indicators are then preprocessed. Based on the parameter sensitivity analysis results, a parameter adjustment strategy is formulated. For parameters that have a significant impact on image quality (such as learning rate and network layer parameters), a larger adjustment range and frequency are set. For parameters that have a smaller impact (such as batch size), a smaller adjustment range is set or the parameters are kept unchanged. At the same time, the interaction between parameters is considered to avoid conflicts and instability in the parameter adjustment process. After each image is generated, the key parameter values ​​that need to be adjusted are calculated using a real-time parameter self-correction algorithm based on feedback data and parameter adjustment strategies. The calculated new parameter values ​​are updated into the image generation model, and retraining or parameter fine-tuning is performed. After parameter updates and model retraining, new images are generated and their quality is evaluated. The effect of parameter adjustment is verified by comparing the differences between the new and old images in various quality indicators. If the adjustment effect is not ideal, the parameter adjustment strategy or parameter value is further adjusted until a satisfactory adjustment effect is achieved.

[0054] The working principle and effect of the above technical solution are as follows: By formulating specific parameter adjustment strategies based on the parameter sensitivity analysis results, the key parameters affecting image quality can be finely adjusted, which can more accurately optimize the performance of the image generation model and improve the quality of the generated images; by considering the interaction between parameters when adjusting parameters, conflicts and instabilities that may occur during parameter adjustment can be avoided, thereby ensuring the stability and reliability of the model during the optimization process.

[0055] By analyzing the feedback data of the generated images and extracting key quality assessment indicators (such as style consistency, detail accuracy, and planning rationality), the feedback data can be effectively used to guide parameter adjustment, making the adjustment process more targeted and efficient. Through the real-time parameter self-correction algorithm, the model can adjust parameters in a timely manner based on the feedback data after each image generation, making the image generation process more adaptive and able to maintain high-quality output in the face of constantly changing needs and data.

[0056] By setting different adjustment magnitudes and frequencies in the adjustment strategy (setting larger magnitudes for parameters with greater impact and smaller magnitudes or keeping them unchanged for parameters with less impact), over-adjustment can be reduced, frequent ineffective optimizations can be avoided, and the efficiency of model training can be improved. By comparing the differences between new and old images on various quality indicators, it is possible to verify whether the adjustment effect has achieved the expected goal, thereby ensuring that each adjustment has actual improvement and improving the traceability and credibility of the optimization process.

[0057] By introducing multiple quality assessment indicators such as style consistency, detail accuracy, and planning rationality, the image generation results can be comprehensively evaluated from multiple dimensions, improving the adaptability of the image generation model to diverse quality requirements. By evaluating each generated image and continuously optimizing and adjusting parameters based on feedback data, the model performance can be gradually improved in each iteration, enhancing the efficiency and accuracy of the overall optimization iteration.

[0058] In one embodiment of the present invention, step S5 includes: S51. Select a representative set of land spatial planning scenarios as the test set; generate images for the test set using a dynamically optimized image generation model; S52. Use objective indicators (such as structural similarity index, peak signal-to-noise ratio, etc.) to compare the generated image with the standard image of the actual scene to evaluate the similarity and quality of the image. S53. Subjectively evaluate the generated images using an expert knowledge base, such as by scoring or descriptive feedback, to assess aspects such as style consistency, detail accuracy, and planning rationality; combine the results of objective indicators and subjective evaluations to comprehensively verify the effectiveness and accuracy of the calibration results. S54. After verification, if the calibration results meet the requirements, the final image generation model will be packaged, including model parameters, structure, dependency libraries and other information; the packaged model will be output and applied to actual land and space planning projects to provide planners with high-quality image generation services that meet the needs of the field, and to assist in planning decision-making and design work.

[0059] S55. Establish a continuous model optimization mechanism, regularly collect user feedback and new domain knowledge, and iteratively update the model.

[0060] The working principle and effect of the above technical solution are as follows: by selecting representative land and space planning scenarios as test sets and verifying the performance of the model in practical applications, the generated images can be ensured to more accurately meet the needs of the planning field, thereby improving the applicability and practical value of the model in actual projects; by combining objective indicators and subjective evaluations from expert knowledge bases, the quality of the generated images can be comprehensively evaluated, thereby reducing the bias that may be caused by a single evaluation method and improving the comprehensiveness and reliability of the evaluation results.

[0061] By comparing the generated images with standard images of the actual scene and using objective indicators for similarity and quality assessment, the details of image quality can be accurately grasped, improving the fit between the generated results and the expected standards. By establishing an expert knowledge base and collecting user feedback, combined with a regular iterative update mechanism, the image generation model can be ensured to continuously adapt to the latest domain requirements, constantly optimize the generation effect, and improve the dynamic adaptability of the model.

[0062] Validated calibration results ensure that the generated model meets actual needs, reducing the risk of errors or non-compliance with actual requirements during model application and improving the stability and reliability of the model in actual projects. By applying the optimized and validated image generation model to land and space planning projects, it can assist planners in providing high-quality image support, help decision-makers better understand planning schemes, and improve the scientific nature and efficiency of planning decisions.

[0063] By encapsulating the final image generation model, including its parameters, structure, and dependencies, the model can be easily ported and applied, improving its convenience and scalability. By establishing a continuous optimization mechanism and regularly updating the model, adjustments can be made based on the latest user feedback and domain knowledge, ensuring the model maintains high efficiency and accuracy in long-term applications and enhancing its long-term adaptability and innovation capabilities.

[0064] One embodiment of the present invention provides a self-learning feedback calibration system for an image generation model, comprising: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0065] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A self-learning feedback calibration method for an image generation model, characterized in that, The method includes: S1. Collect relevant data in the field of land and space planning, perform structured processing on the relevant data, convert it into a computer-recognizable encoding form, and construct a domain knowledge encoding library; select a basic image generation model, use LoRA fine-tuning technology, and combine it with the domain knowledge encoding library constructed above to fine-tune the parameters of the basic image generation model; based on the fine-tuned image generation model, generate an initial land and space planning image dataset according to the given initial input conditions. S2. Acquire multimodal data related to land spatial planning, fuse the multimodal data, extract key features, and construct multimodal conditional control vectors; integrate the ControlNet control network into the image generation model, and train and configure the ControlNet using the multimodal conditional control vectors; further control the initially generated land spatial planning image through ControlNet to obtain an enhanced image dataset; S3. Define feedback metrics for evaluating the quality of generated images, and use a pre-trained evaluation model to evaluate the enhanced image dataset; obtain comprehensive feedback data for each image; and construct the pose evaluation matrix of the image generation model based on the comprehensive feedback data. S4. Perform sensitivity analysis on the key parameters of the image generation model, and design a real-time parameter self-calibration algorithm based on the pose evaluation matrix and parameter sensitivity analysis results; use the self-calibration algorithm to dynamically optimize and iterate the image generation model. S5. Select a portion of land spatial planning scenarios as a test set, and generate images using a dynamically optimized image generation model; compare the generated images with actual scenarios or standard images designed by professional designers to verify the calibration results.

2. The self-learning feedback calibration method for an image generation model according to claim 1, characterized in that, S1 includes: S11. Obtain relevant data in the field of land and space planning, and perform structured processing on the collected data to transform knowledge into a computer-recognizable coded form and build a domain knowledge coding library. S12. Select an existing image generation model as the base model and conduct a comprehensive evaluation based on the model's performance and resource consumption; use LoRA fine-tuning technology, combined with a domain knowledge encoding library, to fine-tune the parameters of the base model. S13. Based on the actual needs of land spatial planning, set initial input conditions; based on the fine-tuned image generation model, generate an initial land spatial planning image dataset according to the given initial input conditions, and annotate the generated images.

3. The self-learning feedback calibration method for an image generation model according to claim 1, characterized in that, The S2 includes: S21. Acquire multimodal data related to land and space planning, and preprocess the collected multimodal data; S22. Based on the multimodal fusion network in deep learning, data from different modalities are fused to extract key features; the fused features are represented as vectors to construct multimodal conditional control vectors. S23. Integrate the ControlNet control network into the image generation model to control the image generation process according to the given condition vector; and use multimodal conditional control vectors to train and configure ControlNet. S24. Further control the initially generated land spatial planning image using ControlNet to obtain an enhanced image dataset.

4. The self-learning feedback calibration method for an image generation model according to claim 1, characterized in that, The S3 includes: S31. Define the feedback indicators; S32. Using the pre-collected labeled data, train the evaluation model, and calculate the feedback index defined above based on the trained evaluation model. S33. Use the trained evaluation model to evaluate the enhanced image dataset and obtain comprehensive feedback data for each image. S34. Based on the comprehensive feedback data, construct the pose evaluation matrix of the image generation model; S35. Analyze the constructed attitude evaluation matrix to identify the model's performance under different conditions.

5. The self-learning feedback calibration method for an image generation model according to claim 4, characterized in that, The feedback metrics include: Style consistency index is used to measure the similarity between generated images and typical styles in the field of land spatial planning; The detail accuracy metric is used to evaluate the accuracy of details in an image, including buildings and terrain. The planning rationality index is used to check whether the image conforms to the layout principles and normative requirements of the national land spatial planning.

6. The self-learning feedback calibration method for an image generation model according to claim 4, characterized in that, The rows of the pose evaluation matrix represent different input conditions or parameter settings, the columns represent different feedback indicators, and the elements in the matrix represent the quality score of the generated image under specific input conditions and parameter settings.

7. The self-learning feedback calibration method for an image generation model according to claim 1, characterized in that, The S4 includes: S41. Perform sensitivity analysis on the key parameters of the image generation model and obtain the sensitivity analysis results; S42. Based on the pose evaluation matrix and parameter sensitivity analysis results, and based on the real-time parameter self-correction algorithm, the values ​​of key parameters are automatically adjusted according to the feedback data of the currently generated image. S43. Utilize a self-calibration algorithm to dynamically optimize and iterate the image generation model; after each image generation, adjust the parameters promptly based on feedback data, and generate the image again for evaluation. S44. Set iteration termination conditions, and through continuous iteration, make the model gradually adapt to the needs of the field of land and space planning, and achieve dynamic optimization.

8. The self-learning feedback calibration method for an image generation model according to claim 7, characterized in that, S42 includes: The feedback data of the currently generated image is analyzed to extract key quality assessment index data, and the extracted key quality assessment index data is preprocessed. Based on the results of the parameter sensitivity analysis, a parameter adjustment strategy is formulated. After each image is generated, the key parameter values ​​that need to be adjusted are calculated using a real-time parameter self-correction algorithm based on feedback data and parameter adjustment strategies. The calculated new parameter values ​​are updated into the image generation model, and retraining or parameter fine-tuning is performed. After parameter updates and model retraining, new images are generated and their quality is evaluated. The effect of parameter adjustment is verified by comparing the differences between the new and old images in various quality indicators.

9. The self-learning feedback calibration method for an image generation model according to claim 1, characterized in that, The S5 includes: S51. Select a portion of land spatial planning scenarios as the test set; generate images for the test set using a dynamically optimized image generation model; S52. Use objective indicators to compare the generated images with standard images of the actual scene to evaluate the similarity and quality of the images; S53. Subjectively evaluate the generated images using an expert knowledge base, and comprehensively verify the effectiveness and accuracy of the calibration results by combining objective indicators and the results of the subjective evaluation. S54. After verification, if the calibration results meet the requirements, the final image generation model will be encapsulated and the encapsulated model will be output. S55. Establish a continuous model optimization mechanism, regularly collect user feedback and new domain knowledge, and iteratively update the model.

10. A self-learning feedback calibration system for an image generation model, comprising: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 9.

Images