Block texture generation method based on multi-objective optimization

Abstract

The invention discloses a block texture generation method based on multi-objective optimization, and the method comprises the following steps: S1, constructing a historical block multi-modal data set, and carrying out the standardized preprocessing and semantic segmentation color coding of image samples in the data set; s2, based on the preprocessed data set, generating a high-precision historical block plane initial layout by adopting a generative adversarial network; s3, performing diversified expansion under semantic guidance on the initial layout by using a low-rank self-adaptive fine-tuning diffusion model to generate a diversified candidate scheme set; s4, constructing a multi-target evaluation index system for the block texture and the road network; and S5, performing iterative optimization on candidate schemes in the diversified candidate scheme set based on a multi-objective genetic algorithm, and screening out an optimization scheme. According to the method, the problems of insufficient form precision, low generation efficiency and poor scheme comprehensive performance in the existing historical block texture reconstruction can be solved, and accurate generation and organic optimization of the historical block texture are realized.

Description

Technical Field

[0001] This invention relates to the fields of historical district preservation and urban planning, and in particular to a method for generating street texture based on multi-objective optimization. Background Technology

[0002] While Generative Adversarial Networks (GANs) have been applied to building layout generation in existing technologies, they suffer from limitations such as high dependence on training data, weak generalization ability of generated results, and poor interpretability. The emergence of diffusion models has provided a new path to solve the above problems, but in the generation of historical street fabric, they lack a sufficient understanding of the spatial structural logic and systematic modeling of the overall fabric of the street and the co-evolution of the road network, making it difficult to balance the needs of historical authenticity and modern spatial performance optimization.

[0003] Furthermore, existing methods often focus on single-dimensional generation optimization, failing to construct a multi-objective quantitative evaluation system encompassing both architectural fabric and road networks. This results in generated schemes that are difficult to meet the comprehensive needs of historical district preservation and organic renewal. Therefore, there is an urgent need for a district fabric generation method that combines morphological accuracy, generation diversity, and multi-objective optimization capabilities. Summary of the Invention

[0004] This invention provides a method for generating streetscape texture based on generative AI and multi-objective optimization. This invention addresses the problems of insufficient morphological accuracy, low generation efficiency, and poor overall performance in existing historical streetscape texture reconstruction methods, achieving accurate generation and organic optimization of historical streetscape texture.

[0005] The technical solution of this invention: a method for generating street texture based on multi-objective optimization, comprising the following steps: S1. Construct a multimodal dataset of historical blocks, and perform standardized preprocessing, semantic segmentation, and color coding on the image samples in the dataset; S2. Based on the preprocessed dataset, a generative adversarial network is used to generate a high-precision initial layout of historical street blocks. S3. The initial layout is semantically guided to be diversified and expanded using a low-rank adaptive fine-tuning diffusion model to generate a diverse set of candidate solutions. S4. Construct a multi-objective evaluation index system for street texture and road network; S5. Based on the multi-objective genetic algorithm, using the multi-objective evaluation index system as the fitness function, iteratively optimize the candidate schemes in the diverse candidate scheme set to select the optimized scheme with excellent comprehensive performance.

[0006] The aforementioned method for generating street texture based on multi-objective optimization, specifically includes the following steps in step S1: Constructing a multimodal dataset. Collect GIS images of historical districts, field survey drawings, and images from architectural case libraries; Image samples were resampled to 1024×1024 pixels at a uniform scale, and semantic segmentation color coding was used to distinguish the background, land boundary line, river, road and building textures with different orientations. Geometric symmetry enhancement and noise robustness enhancement are performed on the training set samples, and structured text labels are generated accordingly.

[0007] The aforementioned method for generating street texture based on multi-objective optimization uses a Pix2PixHD model as the generative adversarial network model; the Pix2PixHD model includes a generator and a discriminator. The generator adopts a two-stage series structure. ,in, For global generation of networks, The local enhancement network; the global generation network First, construct the overall layout and macroscopic structure of the image at a resolution of 512×512, and then enhance the local network. Subsequently, the global generator network was analyzed. The output is upsampled by 4 times and locally refined, and then fused. The provided global context information generates the final high-resolution result, ensuring rich detail and structural consistency; The discriminator adopts a multi-scale architecture. The original image and its 2x and 4x downsampled versions are judged separately. By evaluating the realism of the image at different perceptual scales, the generator is guided to produce visually consistent results at multiple levels of overall layout and local texture. The training process of the Pix2PixHD model is optimized using a composite loss function, which is expressed as follows: ; In the formula: To combat the loss, it is used to drive the generated image to approximate the real data distribution; The feature matching loss is used to stabilize training and improve structural coherence by constraining the features of the intermediate layer of the discriminator. The perceptual loss is used to constrain semantic content and texture details in the feature space using a pre-trained VGG network; and All are weighting coefficients. For the discriminator's index; Output the target.

[0008] The aforementioned multi-objective optimization-based street texture generation method, in step S3, uses a LoRA-tuned diffusion model based on the Stable Diffusion 1.5 architecture, achieved through low-rank matrix factorization. Update the original weights ,in and Both represent parameter matrices. As a transpose operation, the updated weights are represented as follows: ; The training objective is to minimize the mean squared error between the generated results and the true annotations. ; In the formula, This is the forward calculation process of the diffusion model. For conditional input, Output the target.

[0009] The aforementioned method for generating street texture based on multi-objective optimization includes a multi-objective evaluation index system in step S4 that includes five indicators for the building texture dimension and five indicators for the road network dimension. The architectural texture dimension indicators include: texture coverage, number of small texture units, number of large texture units, average texture size, and standard deviation of texture size; The road network dimension indicators include: road network density, minimum cost path, spatiotemporal accessibility, average road width, and standard deviation of road width.

[0010] In the aforementioned multi-objective optimization-based street texture generation method, step S4, the iterative optimization based on a multi-objective genetic algorithm, includes: Each candidate solution is encoded as an individual in the potential space to construct an initial population; The fitness value of each individual is calculated based on the aforementioned multi-objective evaluation index system; Elite individuals are selected through an elite retention strategy to form an elite solution set; Convergence is determined by the relative change in the optimal fitness value between generations. If convergence is not achieved, the elite solution set is fed back to the iterative process for further optimization until the convergence condition is met.

[0011] The aforementioned multi-objective optimization-based street texture generation method uses a fitness function formalized through weighted summation as follows: ; In the formula: For the first The normalized scoring function for each indicator. For the first The function corresponding to each indicator. No. The value of the indicator; The optimization goal is .

[0012] In the aforementioned multi-objective optimization-based street texture generation method, the iterative optimization in step S5 optimizes the texture coverage of the generated scheme to 47%-55%, the average texture size to 85-175m², the road network density to 860-960m / hectare, and the spatiotemporal accessibility to 0.87-0.99.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a multimodal dataset by collecting GIS geographic information data of historical blocks, on-site survey maps, and images from a building case library. Combined with standardized processing at a unified scale and semantic segmentation color coding, it clearly distinguishes background, land boundaries, waterways, roads, and the texture of buildings facing different directions, providing a precise data foundation for model training and avoiding data errors that cause the generated results to deviate from the true form. This invention employs a two-level structure of the Pix2PixHD model: a global generation network and a local enhancement network. It first constructs a macro-layout and then refines it locally. Combined with a multi-scale discriminator and a composite loss function, it comprehensively constrains the generated results from the overall structure to the local texture, ensuring that the initial layout is highly consistent with the real historical blocks in terms of spatial logic and morphological details, thus solving the problems of low generation accuracy and poor structural coherence in traditional models.

[0014] 2. This invention is based on a diffusion model with low-rank adaptive (LoRA) fine-tuning. Using the initial layout as a semantic constraint, it generates multiple sets of differentiated candidate solutions by optimizing only a small number of low-rank parameters while retaining the capabilities of the pre-trained model. This approach reduces training costs, avoids overfitting, and can quickly adapt to historical district fabric generation tasks. It provides diverse solutions covering different building combinations and road network forms, meeting the design needs of multi-scenario comparison in historical district renewal and solving the problems of low efficiency and insufficient innovation in traditional manual design.

[0015] 3. This invention constructs a multi-objective evaluation index system encompassing architectural texture and road network. The architectural texture dimension ensures the authenticity of the historical appearance, while the road network dimension meets modern transportation and functional requirements, avoiding the problem of existing methods focusing only on a single dimension, leading to poor practicality of the solutions. This invention uses a multi-objective genetic algorithm to construct a fitness function with multi-dimensional indicators, combined with an elite retention strategy to select high-quality solutions. Convergence is determined based on intergenerational fitness changes, ultimately selecting solutions with excellent overall performance. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of a street texture generation method based on generative AI and multi-objective optimization. Figure 2 This is a schematic diagram illustrating the preprocessing of images for the training and testing sets. Figure 3 This is a diagram of the Pix2PixHD model architecture. Figure 4 Diagram of the fusion architecture for diffusion model and LoRA fine-tuning; Figure 5 A schematic diagram for generating a diverse set of candidate solutions; Figure 6 This is a schematic diagram of iterative optimization using a multi-objective genetic algorithm. Figure 7 A comparison of the spatial morphology of the original core area (OA) and the expanded area (EA) and the resulting map; Figure 8 A comparison of the spatial morphology of the original core area (OA) and the expanded area (EA) of the three models, and subsequent generated result diagrams; Figure 9 A comparison chart of architectural texture indicators between the original core area (OA) and the expanded area (EA) of the three models; Figure 10 A comparison chart of road network indicators between the original core area (OA) and the expanded area (EA) of the three models; Figure 11 A comparison chart of the total scores for each model's approach. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0018] Example: A method for generating street texture based on multi-objective optimization, such as... Figure 1 As shown, it includes the following steps: S1. Construct a multimodal dataset of historical blocks and perform standardized preprocessing, semantic segmentation, and color coding on the image samples in the dataset. In this step, to address the specific issues of morphological generation in riverside historical blocks, a structured multimodal database was constructed to provide a reliable data foundation for subsequent generative design model training and validation. The data sources integrate GIS images, field survey drawings, design materials, and a professional architectural case library to ensure the representativeness, standardization, and practical feasibility of the samples in terms of morphology.

[0019] The sample selection followed the following principles: (1) the layout of the blocks and the waterways were closely related; (2) the spatial texture characteristics were clearly distinguishable; (3) it had certain historical and cultural value; and (4) the image data was complete, including a clear scale and geographical reference. Finally, a total of 60 representative historical blocks were selected, divided into 50 cases as the training set and 10 cases as the test set, and the test set did not participate in any training process.

[0020] To adapt to the input requirements of deep learning models, all image samples were standardized using Adobe Photoshop. First, the images were resampled to 1024×1024 pixels at a uniform scale (1:2000) and saved as PNGs to preserve clear outlines and color information. For color encoding, the RGB color filling strategy commonly used in semantic segmentation was adopted, assigning specific color identifiers to different spatial elements: background (0, 0, 0), land boundary line (255, 0, 0), river (100, 150, 200), road (100, 100, 100), buildings with rooftops parallel to the river (255, 255, 0), and buildings with rooftops perpendicular to the river (255, 255, 255).

[0021] Two images are generated for each case: training set A corresponds to the state of "removing existing buildings while retaining the land boundary line," while training set B retains the original building texture. The test set samples use the same preprocessing workflow as training set A. This image design balances visual appeal and machine readability, meeting the pix2pixHD model's requirement for paired image input, such as... Figure 2 As shown.

[0022] To enable cross-modal training between text and images, each image is accompanied by a structured text description (Prompt). The label content is concise and clear, highlighting keywords related to spatial structure and morphological features, and embedding a multi-dimensional semantic control framework. This framework introduces adjustable weight coefficients x ∈ [0, 2] to adjust the semantic intensity, enhancing the LoRA model's sensitivity to morphological features and the controllability of generation.

[0023] To improve the model's generalization ability and address the limited number of training samples, this embodiment implements two types of data augmentation strategies: Geometric symmetry enhancement: By performing horizontal and vertical flipping operations, the training samples are expanded to four times their original size. Horizontal flipping strengthens the model's recognition of shape orientation invariance, suitable for learning the mirrored layout of riverside buildings; vertical flipping, by introducing vertical symmetry, enhances the model's ability to extract local textural features. Noise robustness enhancement: Zero-mean Gaussian noise is applied to simulate image degradation or sensor interference that may be encountered in real-world applications, forcing the model to learn more robust feature representations.

[0024] After the above enhancement processing, the original training samples finally constitute a moderately sized and structurally rigorous multimodal historical street texture database, providing a data foundation for the efficient training and reliability evaluation of subsequent neural network models.

[0025] S2. Based on the preprocessed dataset, a generative adversarial network (GAN) is used to generate a high-precision initial planar layout of historical blocks. This step first generates a high-fidelity initial planar layout of historical blocks using the Pix2PixHD model. Then, using the aforementioned output as semantic conditions, a diffusion model fine-tuned with low-rank adaptive (LoRA) technology is driven to diversify the scheme set, ultimately generating a planar layout scheme that strictly adheres to the input semantic constraints while possessing rich morphological variability. Specifically: This embodiment uses the Pix2PixHD model as the core generation architecture. For example... Figure 3 As shown, this model is an advanced conditional generative adversarial network (cGAN) designed specifically for high-resolution image synthesis tasks. Its core innovation lies in the introduction of a coarse-to-fine generator and multi-scale discriminator structure to effectively address the issues of detail preservation and training stability in high-resolution image generation.

[0026] The generator adopts a two-stage series structure. ,in, For global generation of networks, The local enhancement network; the global generation network First, construct the overall layout and macroscopic structure of the image at a resolution of 512×512, and then enhance the local network. Subsequently, the global generator network was analyzed. The output is upsampled by 4 times and locally refined, and then fused. The provided global context information generates the final high-resolution result (e.g., 1024×1024), ensuring rich detail and structural consistency; The discriminator adopts a multi-scale architecture. The original image and its 2x and 4x downsampled versions are judged separately. By evaluating the realism of the image at different perceptual scales, the generator is guided to produce visually consistent results at multiple levels of overall layout and local texture. The training process of the Pix2PixHD model is optimized using a composite loss function, which is expressed as follows: ; In the formula: To combat the loss, it is used to drive the generated image to approximate the real data distribution; The feature matching loss is used to stabilize training and improve structural coherence by constraining the features of the intermediate layer of the discriminator. The perceptual loss is used to constrain semantic content and texture details in the feature space using a pre-trained VGG network; and All are weighting coefficients. For the discriminator's index; Output the target.

[0027] The Pix2PixHD model generated a high-fidelity initial layout of the historical district plan, laying the morphological foundation for subsequent diversified expansions. This initial scheme will serve as a semantic condition, driving the next stage of the diffusion model to diversify further.

[0028] S3. A low-rank adaptive fine-tuning diffusion model is used to semantically guide the diversification of the initial layout, generating a diverse set of candidate schemes. In this step, to achieve diversified generation of historical street layouts, this study adopts a diffusion model as the core generation architecture and introduces low-rank adaptive (LoRA) technology for efficient domain adaptation. This method achieves rapid adaptation to the historical street texture generation task under limited computing resources by embedding lightweight trainable parameters into the pre-trained diffusion model.

[0029] The generation mechanism of the diffusion model is based on inverse learning of the forward noise addition process. The forward process is controlled by variance scheduling parameters. Original data asymptotically converted to Gaussian noise Its transformation process follows the Markov chain hypothesis: ; The reverse process involves training the U-Net network. Learn to gradually denoise to reconstruct a structured planar graph from random noise.

[0030] To efficiently adapt the pre-trained diffusion model (based on Stable Diffusion 1.5), this study employs the LoRA fine-tuning technique. This method utilizes low-rank matrix factorization... Update the original weights ,in , , and rank The updated weights are represented as follows: ; This method reduces the number of training parameters by two orders of magnitude, significantly improving fine-tuning efficiency while maintaining model generation capabilities.

[0031] The training objective is to minimize the mean squared error between the generated results and the true annotations. ; In the formula, This is the forward calculation process of the diffusion model. For conditional input, Output the target.

[0032] This joint framework fully leverages the advantages of diffusion models in generating complex structures, while achieving a balance between computational efficiency and generation quality through LoRA technology. It provides an effective technical path for the diversified generation of historical district textures, such as... Figure 4 As shown.

[0033] S4. Construct a multi-objective evaluation index system for street texture and road network; the multi-objective evaluation index system includes five indicators for building texture dimension and five indicators for road network dimension; The architectural texture dimension indicators include: texture coverage, number of small texture units, number of large texture units, average texture size, and standard deviation of texture size; The road network dimension indicators include: road network density, minimum cost path, spatiotemporal accessibility, average road width, and standard deviation of road width.

[0034] Specifically, the texture coverage index is defined as the ratio (percentage) of the building footprint area to the total land area, reflecting the overall building density and spatial compactness: ; In the formula: The total area of ​​the textured area. This represents the total area of ​​the red plots.

[0035] Its principle originates from the classic measure of building density in urban morphology, and it is usually applied at the scale of urban areas or regions (such as 1-5 km²) to assess land use intensity. In the study of historic districts, texture coverage can be used to judge the preservation integrity of traditional building complexes and the degree of intrusion of modern development. High coverage often reflects a continuous traditional texture, but attention should also be paid to the potential impact of over-densification on the original spatial structure.

[0036] The number of small texture units is a statistical indicator representing the total number of independent building units with an area of ​​less than 30 m², used to characterize the degree of spatial fragmentation. ; In the formula: Area of ​​a single textured block (m²) 2 ).

[0037] Its core function lies in identifying the distribution of fine-grained architectural elements, making it suitable for micro-analysis at the block or plot level. In the context of historic districts, this indicator helps to capture the remaining condition of traditional alleyways, outbuildings, or courtyard structures, thereby determining whether the original "capillary" spatial network has been preserved.

[0038] The number of large-scale building units is a statistical indicator of the total number of independent building units with an area greater than 240 m², used to identify the distribution characteristics of large buildings. ; As a representative indicator of "coarse-grained" space, it is mainly used in the study of historical districts to identify whether large-scale new buildings have been inserted, and then to assess their harmony with the traditional texture. It is an important basis for judging "style conflict".

[0039] The average size index of the texture is calculated as the arithmetic mean of the base area of ​​all building units (m²). 2 ), reflecting the benchmark scale of regional architectural space: ; Its principle is based on a statistical description of the concentration trend of building unit area, and it is applicable to multi-scale analysis from street blocks to urban areas. In historical districts, this value can be used to compare the scale characteristics of buildings in different historical periods, identify whether the scale of the original buildings has been distorted due to reconstruction or addition, and provide data support for "scale restoration".

[0040] Road network density, as a fundamental indicator, quantifies the coverage and connectivity potential of the road network by measuring the total road length per unit area (m / hectare). ; In the formula: The total length of the road (m) The area is in hectares.

[0041] This indicator originates from classic urban morphology studies and is applicable to systematic analysis at the city or district scale (1-10 km²). In the context of historic districts, high road network density often corresponds to the "dense road network" characteristics of traditional streets and alleys, which helps improve pedestrian accessibility; however, it is also necessary to combine the actual width and traffic capacity of the streets and alleys to comprehensively evaluate whether they meet modern use needs without destroying the traditional spatial texture.

[0042] The minimum cost path metric, based on graph theory, objectively assesses traffic efficiency by calculating the shortest path length (m) between the two farthest points in the network. ; In the formula: Represents a node and The shortest path length between them.

[0043] This indicator is applicable to micro-analysis at the street level and can effectively identify connectivity barriers caused by dead-end roads and winding alleys, providing an important basis for optimizing the pedestrian network of historic districts.

[0044] The spatiotemporal accessibility index, based on the theory of time geography, assesses the walkability and human-scale nature of spatial services by calculating the proportion of road areas accessible within a 2-hour walking radius. ; In the formula: This is a set of nodes that can be reached within 2 hours. This is the total set of nodes.

[0045] This indicator has particular value in neighborhood-scale (500 m-1 km radius) applications, effectively measuring the ease with which residents and tourists in historic districts can access service facilities and cultural attractions. It is a core assessment tool for building pedestrian-friendly historic districts.

[0046] The average road width dimension index reflects the overall scale characteristics of the street space by calculating the arithmetic mean (m) of the widths of all roads:

[0047] In the analysis of street cross-section scale, this indicator plays an important role in identifying the inappropriate widening of historical streets and alleys due to the adaptation to modern traffic needs, and helps to maintain the original spatial proportions of historical blocks.

[0048] The standard deviation index for road width dimensions reflects the homogeneity of street space by measuring the dispersion (m) of road width values.

[0049] This indicator is applicable to comparative analysis of multiple streets within a district. A lower standard deviation usually indicates a well-planned and well-preserved traditional neighborhood, while a higher dispersion suggests scale disorder caused by renovations over multiple periods, providing a quantitative basis for conservation planning.

[0050] This embodiment constructs a comprehensive evaluation system encompassing two dimensions: architectural fabric and road network, based on the spatial characteristics and protection requirements of traditional settlements. The scoring criteria for each indicator are rigorously validated through architectural analysis, ensuring that the optimization process effectively balances the preservation of historical features with modern functional needs.

[0051] In terms of architectural texture, the scoring criteria fully consider the protection requirements of traditional spatial forms. A texture coverage rate of 50% is the ideal threshold, a figure derived from systematic surveys of historical districts and consistent with the 45%-55% building density characteristic of traditional settlements. The control standards for the number of small texture units (<30m²) and large texture units (>240m²) are based on the analysis of the relationships between traditional building groups, aiming to prevent spatial fragmentation and avoid building volumes that are incompatible with the historical environment. The average texture size is ideally set at 120m², and the standard deviation of texture size is also set at 120m². These two values ​​are established with reference to the common scale of traditional dwellings, satisfying modern usage needs while maintaining the organic integrity and spatial rhythm of the building group.

[0052] In terms of road network, the scoring criteria reflect the inheritance and development of traditional street and alley spatial characteristics. The ideal road network density is 1000 meters per hectare, consistent with the high-density road network characteristics of water towns. The minimum cost path scoring mechanism balances traffic efficiency with traditional spatial features, while spatiotemporal accessibility directly uses proportional values ​​to emphasize the pedestrian-oriented nature of historical districts. The average road width, set at 6 meters for maximum score and 10 meters for minimum score, precisely corresponds to the typical scale of traditional streets and alleys, while the control of the standard deviation of road width ensures a balance between the uniformity and diversity of street space.

[0053] This system adopts an equal-weighting strategy, with each indicator having a weight of 0.1, reflecting the emphasis on balancing multi-dimensional values ​​in the organic renewal of historical districts. This evaluation system transforms professional protection requirements into quantifiable optimization objectives, providing a scientific basis for evaluating the fitness of genetic algorithms and ensuring that the optimization results both respect historical environmental characteristics and meet modern development needs.

[0054] The fitness function, which combines the above ten indicators, can be formalized by weighted summation as follows: ; In the formula: For the first The normalized scoring function for each indicator. For the first The function corresponding to each indicator. No. The value of the indicator; The optimization goal is .

[0055] Table 1 shows the weight distribution and normalized values: Table 1

[0056] S5. Based on a multi-objective genetic algorithm, using the multi-objective evaluation index system as the fitness function, iteratively optimize the candidate schemes in the diverse candidate scheme set to select the optimized scheme with excellent comprehensive performance. In this step, the iterative optimization is based on the diverse planar map schemes generated in the early stage, and improves the comprehensive performance of the candidate schemes in terms of texture structure and road network performance by simulating the natural evolution process.

[0057] The optimization process encodes each planar solution as an individual in the latent space. , construct containing The initial population of each scheme A custom evaluation system is used to calculate the fitness score for each individual based on ten performance indicators. Establish quantitative performance-oriented optimization goals.

[0058] The algorithm maintains population quality through an elite retention strategy, prior to each generation of selection. Elite individuals constitute the elite solution set. Convergence is determined by the relative change in optimal fitness values ​​across successive generations: ; in, This is a preset convergence threshold. If convergence fails, the algorithm feeds back the elite solution set to the iterative process for continued optimization until the convergence condition is met.

[0059] The optimization mechanism organically combines the creative output of generative AI with multi-objective optimization capabilities. Through performance-driven directed evolution, it achieves an effective balance between complex morphological constraints and functional requirements, providing scientific decision support for the organic renewal of historical districts. Figure 5 A schematic diagram illustrating the generation of diverse candidate solution sets is shown; Figure 6 A schematic diagram of iterative optimization is shown.

[0060] The method steps of the present invention have been described in detail above, and the present invention will be further described with reference to specific examples.

[0061] 1. Dataset Through a systematic data processing and enhancement process, this embodiment successfully constructed a multimodal texture database of the riverside historical district, as shown in Table 2: Table 2

[0062] The database ultimately contains 510 high-quality image samples. The training set was expanded from the original 50 sets of samples to 250 sets through data augmentation, while the test set retained 10 independent sets of samples. Simultaneously, 500 structured text labels were generated, constructing a complete text-image pairing dataset. Data augmentation effect analysis shows that geometric symmetry augmentation effectively improved the model's ability to recognize spatial orientation, while noise robustness augmentation significantly improved the model's stability under interference environments.

[0063] The quality assessment of the dataset shows that: (1) the spatial distribution covers 10 typical cities, ensuring geographical representativeness; (2) the architectural texture types cover three main patterns: parallel layout, vertical layout and mixed layout; (3) the sample complexity distribution is reasonable, including a multi-level structure from simple texture units to complex texture combinations.

[0064] The application process includes three main stages: First, a generative model is trained based on the constructed multimodal dataset; second, the land use boundary of the historical block is input into the trained model to generate various texture layout schemes; finally, the texture evaluation index system proposed in this paper is used to quantitatively evaluate the generated results.

[0065] 2. Generation of historical district layouts based on generative AI 2.1, Pix2pixHD Based on the Pix2PixHD model architecture constructed above, experiments were conducted to generate the floor plan layout of historical districts. The experiments used a multimodal historical district image dataset, including 250 training samples and 10 test samples. All input images were uniformly preprocessed to a resolution of 1024×1024 pixels and a semantic segmentation color coding system was adopted to clearly distinguish spatial elements such as buildings, roads, and waterways.

[0066] For model training, the Adam optimizer is used for parameter updates, with an initial learning rate set to 2×10⁻⁶. -4 The training process employs a phased learning rate strategy: a constant learning rate is maintained for the first 300 epochs to promote stable model convergence, followed by linear decay for the next 700 epochs to refine parameter tuning. The total training duration is 1000 epochs, with a cumulative training time of approximately 18 hours.

[0067] To verify the scientific validity and effectiveness of the evaluation index system constructed in this study, and to identify the core problems faced by historical blocks at the levels of texture and road network, this paper first conducts a systematic quantitative comparison of the original core area (OA) and the expanded area (EA). The spatial morphology comparison and subsequent generation results are as follows: Figure 7 As shown in Table 3, the evaluation index scores of the original core area (OA), the expanded area (EA), and the AI-generated area (GA) are compared. Table 3

[0068] Note: Pix2PixHD (GA1) is the single output of the deterministic model; LoRA (GA2) and Pix2PixHD-LoRA (GA3) are the mean ± standard deviation of the results of multiple runs of the probabilistic model (n=40 for GA2, n=100 for GA3).

[0069] The two regions showed significant differences in several key indicators, objectively revealing the damage that existing expansion activities have caused to the integrity of the historical fabric and the road network at a human scale.

[0070] Quantitative results indicate that the core problem faced by EA is not insufficient texture coverage, but rather a severe disorder in the scale order of texture. Although EA has a slight advantage in the number of small textures (0.98), its scores on key dimensions such as the number of large texture units (0.58), the average size of texture (0.43), and the standard deviation of texture size (0.00) are significantly lower than those of OA. This reveals that existing development practices have led to a trend towards homogenization and uniformity in texture composition, severely damaging the multi-layered and organically composite texture layers unique to OA.

[0071] In terms of road network structure, EA lags behind OA in all road network indicators, especially in the "average road width" (0.64) and "standard deviation of road width" (0.38), where EA (0.98, 0.72) show significant differences. This suggests that EA's road reconstruction may have adopted a simple widening strategy to accommodate motor vehicle traffic, while ignoring the natural changes in road width. This weakens the "narrow street and dense network" characteristics of historical streets and alleys as containers for social interaction, resulting in a rigid and monotonous spatial experience.

[0072] It is evident that quantitative analysis accurately identifies the root cause of the disharmonious appearance within the "Build Control Zone" (EA) of historic districts: a severe decline in the diversity of textural scales and a systematic degradation of the human-centered scale of the road network. This conclusion points to a core optimization direction for subsequent generative design: effective textural generation models should not merely pursue morphological imitation, but should focus on reconstructing an organic textural scale spectrum and restoring the refined structure and pedestrian-friendliness of the road network. Figure 6 As shown, subsequent experiments, based on this diagnostic conclusion, are dedicated to generating an AI generation scheme that can repair existing defects in EA and effectively continue the spatial genes of OA.

[0073] Furthermore, to evaluate the effectiveness of different generation paradigms, a systematic comparison was conducted between Pix2PixHD, LoRA, and the proposed Pix2PixHD-LoRA hybrid model, with generated samples as follows: Figure 8 As shown, this study not only focuses on the performance of the generated results relative to the benchmarks (original core area OA and expanded area EA) in terms of quantitative metrics, but also emphasizes the analysis of the inherent trade-offs between different models in terms of morphological accuracy, functional rationality, and generation diversity. The comparison results are as follows: Figure 9-11 As shown, where Figure 9 and Figure 10 In the figure, the bar chart on the horizontal axis represents the generation results of the original core area, the expanded area, the pix2pixHD model, the LoRA model, and the Pix2PixHD-LoRA hybrid model of this invention.

[0074] As can be seen, the single Pix2PixHD model (GA1), as a deterministic model, reveals the fundamental limitations of this type of model through its single output. Although it partially corrects the shortcomings of EA in texture generation—raising the standard deviation of texture size from 0 to 0.71 and maintaining a high texture coverage (0.88) and a small texture quantity (0.96)—indicating its ability to learn and generate uniform and dense texture patterns, the number of large textures it generates is abnormally high (0.10), while the average texture size score (0.55) remains unsatisfactory, revealing its inherent inadequacy in controlling the distribution of complex texture scales. More importantly, in road network generation, its performance is comprehensively inferior to OA, especially in spatiotemporal accessibility (0.66) and road width control (0.48), indicating its difficulty in coordinating complex functional and structural relationships. Furthermore, its single generation result, which fails to provide alternative solutions, greatly limits its application value in planning and design practice.

[0075] The independent LoRA model (GA2) exhibits strong generative diversity, generating a large number of diverse schemes. However, its quantitative results (total score 5.38±1.64) reveal serious problems: extremely high output instability and systematically low overall quality. It lags far behind OA in key indicators, even falling short of EA (e.g., texture coverage 0.46±0.15). Large standard deviations (e.g., road network indicators generally exceeding ±0.24) indicate that its output fluctuates wildly between "complete failure" and "occasional feasibility." While this model can serve as a "creative engine," its inherent geometric inaccuracies (e.g., blurred building boundaries and poor road network connectivity) make it difficult to independently generate reliable planning schemes.

[0076] The proposed Pix2PixHD-LoRA hybrid model (GA3) successfully addresses the aforementioned challenges, achieving a balance between quality, diversity, and stability. In terms of morphological fidelity, GA3 significantly surpasses GA1 across all key metrics and approaches the level of OA. Its average texture size score (0.85) is 54.5% higher than GA1 (0.55). More importantly, it surpasses OA (0.74) in texture size standard deviation (0.95), indicating its ability to generate more uniform and harmonious texture patterns. Regarding road network functionality, GA3 achieves a milestone, with spatiotemporal accessibility (0.96) and road network density (0.94) approaching the optimal values ​​of OA, and the average road width score (0.88) being 83.3% higher than GA1 (0.48). Crucially, GA3 achieves a high score of 8.06 and an extremely low standard deviation (±0.19) while generating 100 schemes. This result has profound practical implications: it demonstrates that the hybrid framework does not simply trade off the quantity of generated solutions against the quality of outputs. Instead, through the high-precision structural priors provided by Pix2PixHD, it effectively constrains and guides the generation process of LoRA branches, focusing its sampling process on high-quality solution space regions. Therefore, designers do not obtain a large number of candidate solutions of varying quality, but rather multiple (e.g., hundreds) high-quality alternative designs that excel in both morphological rationality and functional integrity, and exhibit high consistency.

[0077] Comprehensive comparison results show that both Pix2PixHD and LoRA have significant limitations when applied independently: the former generates only a single, unremarkable result, while the latter has low geometric accuracy and poor stability, making it difficult to directly apply in practical engineering. The hybrid model proposed in this invention achieves a synergistic effect beyond simple combinations of single models through a serial architecture, achieving a superior overall balance in the three key dimensions of morphological accuracy, functional rationality, and generation diversity. This framework provides a digital generation approach that combines algorithmic innovation with engineering practicality to address the dual challenges of "continuing traditional fabric" and "adapting to modern functions" in the protection and renewal of historic districts.

[0078] 2.3 Multi-objective iterative optimization To verify the comprehensive improvement effect of multi-objective optimization iteration on historical block generation schemes, this study, based on the candidate scheme set generated by the Pix2pixHD-LoRA fusion model, aimed at the synergistic optimization of urban fabric structure and road network performance, and conducted multiple rounds of iterative search using an equal-weight comprehensive evaluation system. After three generations of iterations, the optimization process reached the convergence condition (final intergenerational improvement rate 0.7% < threshold τ=0.01), and the indicators of each dimension showed a systematic improvement, as shown in Table 4: Table 4

[0079] As shown in Table 4, in terms of architectural texture, the texture coverage increased from 0.94±0.05 to 0.99±0.02, the average texture size score rose from 0.81±0.16 to 0.86±0.11, and the size standard deviation remained within the reasonable range of 0.65–0.69, reflecting a trend towards more rational texture composition and enhanced order. The number of large texture units also increased from 0.79±0.06 to 0.82±0.06, indicating further strengthening of functional integrity.

[0080] In terms of road network, the road network density score improved from 0.91±0.05 to 0.95±0.01, the spatiotemporal accessibility score significantly improved from 0.90±0.14 to 0.98±0.03, the average road width score reached the full score of 1.00 from 0.91±0.15, and the standard deviation of road width also improved to 0.75±0.02, reflecting the overall optimization of the road network structure in terms of connectivity efficiency and scale coordination.

[0081] Through this process, the overall score of the solution improved from 8.08±0.95 to 8.58±0.32, with a significant increase and a marked narrowing of the standard deviation, indicating that the optimization framework not only improved the overall performance but also enhanced the stability and reliability of the output solution.

[0082] As can be seen, the generative-optimization integrated framework constructed in this invention effectively guides the initial scheme set to converge towards a high-quality solution set that combines spatial performance and landscape characteristics through a multi-objective iterative mechanism, providing a technically feasible and result-controllable digital path for the organic renewal of historical blocks. The results show that, while maintaining the accuracy of inherited forms, the proposed method significantly expands the comprehensive performance of the generated schemes in multiple dimensions, including textural rationality, road network efficiency, and output stability.

[0083] In summary, this invention successfully constructs a generative design framework that combines morphological accuracy, generation diversity, and multi-objective performance optimization capabilities.

Claims

1. A method for generating street texture based on multi-objective optimization, characterized in that, Includes the following steps: S1. Construct a multimodal dataset of historical blocks, and perform standardized preprocessing, semantic segmentation, and color coding on the image samples in the dataset; S2. Based on the preprocessed dataset, a generative adversarial network is used to generate a high-precision initial layout of historical street blocks. S3. The initial layout is semantically guided to be diversified and extended using a low-rank adaptive fine-tuning diffusion model to generate a diverse set of candidate solutions. S4. Construct a multi-objective evaluation index system for street texture and road network; S5. Based on the multi-objective genetic algorithm, using the multi-objective evaluation index system as the fitness function, iteratively optimize the candidate schemes in the diverse candidate scheme set to select the optimized scheme with excellent comprehensive performance.

2. The method for generating street texture based on multi-objective optimization according to claim 1, characterized in that, The construction of the multimodal dataset in step S1 specifically includes: Collect GIS images of historical districts, field survey drawings, and images from architectural case libraries; Image samples were resampled to 1024×1024 pixels at a uniform scale, and semantic segmentation color coding was used to distinguish the background, land boundary line, river, road and building textures with different orientations. Geometric symmetry enhancement and noise robustness enhancement are performed on the training set samples, and structured text labels are generated accordingly.

3. The method for generating street texture based on multi-objective optimization according to claim 1, characterized in that, The generative adversarial network model is a Pix2PixHD model; the Pix2PixHD model includes a generator and a discriminator. The generator adopts a two-stage series structure. ,in, For global generation of networks, The local enhancement network; the global generation network First, construct the overall layout and macroscopic structure of the image at a resolution of 512×512, and then enhance the local network. Subsequently, the global generator network was analyzed. The output is upsampled by 4 times and locally refined, and then fused. The provided global context information generates the final high-resolution result, ensuring rich detail and structural consistency; The discriminator adopts a multi-scale architecture. The original image and its 2x and 4x downsampled versions are judged separately. By evaluating the realism of the image at different perceptual scales, the generator is guided to produce visually consistent results at multiple levels of overall layout and local texture. The training process of the Pix2PixHD model is optimized using a composite loss function, which is expressed as follows: ； In the formula: To combat the loss, it is used to drive the generated image to approximate the real data distribution; The feature matching loss is used to stabilize training and improve structural coherence by constraining the features of the intermediate layer of the discriminator. The perceptual loss is used to constrain semantic content and texture details in the feature space using a pre-trained VGG network; and All are weighting coefficients. For the discriminator's index; Output the target.

4. The method for generating street texture based on multi-objective optimization according to claim 1, characterized in that, The LoRA fine-tuning diffusion model in step S3 is based on the Stable Diffusion 1.5 architecture and is achieved through low-rank matrix factorization. Update the original weights ,in and Both represent parameter matrices. As a transpose operation, the updated weights are represented as follows: ； The training objective is to minimize the mean squared error between the generated results and the true annotations. ； In the formula, This is the forward calculation process of the diffusion model. For conditional input, Output the target.

5. The method for generating street texture based on multi-objective optimization according to claim 1, characterized in that, The multi-objective evaluation index system in step S4 includes five indicators for the building texture dimension and five indicators for the road network dimension. The architectural texture dimension indicators include: texture coverage, number of small texture units, number of large texture units, average texture size, and standard deviation of texture size; The road network dimension indicators include: road network density, minimum cost path, spatiotemporal accessibility, average road width, and standard deviation of road width.

6. The method for generating street texture based on multi-objective optimization according to claim 5, characterized in that, In step S4, the iterative optimization based on the multi-objective genetic algorithm includes: Each candidate solution is encoded as an individual in the potential space to construct an initial population; The fitness value of each individual is calculated based on the aforementioned multi-objective evaluation index system; Elite individuals are selected through an elite retention strategy to form an elite solution set; Convergence is determined by the relative change in the optimal fitness value between generations. If convergence is not achieved, the elite solution set is fed back to the iterative process for further optimization until the convergence condition is met.

7. The method for generating street texture based on multi-objective optimization according to claim 6, characterized in that, The fitness function is formalized by weighted summation as follows: ； In the formula: For the first The normalized scoring function for each indicator. For the first The function corresponding to each indicator. No. The value of the indicator; The optimization goal is .

8. The method for generating street texture based on multi-objective optimization according to claim 1, characterized in that, In step S5, iterative optimization improves the texture coverage of the generated scheme to 47%-55%, the average texture size to 85-175m², the road network density to 860-960m / hectare, and the spatiotemporal accessibility to 0.87-0.99.

Images