A land space planning effect picture intelligent generation method and system

Abstract

The present application relates to a kind of land space planning effect picture intelligent generation method and system, it relates to land space planning technical field, it includes: the original image of user input and generation demand are collected;According to the original image, preprocessed image is generated;Image understanding is carried out through preprocessed image, and then image prompt word is formed;According to generation demand, clear demand prompt word;Image prompt word and demand prompt word are fused, and integrated guide word is obtained;Integrated guide word and preprocessed image are input into the scene rendering model of preestablished, and scene effect picture is generated and output.The present application has the effect of conveniently and quickly generating land space planning effect picture.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of land space planning, and in particular to a land space planning effect picture intelligent generation method and system. BACKGROUND

[0002] Land space planning is an arrangement of land space development and protection in a certain region in space and time, so as to optimize the land space development and protection pattern, realize the coordination of population, economy, resources and environment, and promote high-quality development and high-quality life.

[0003] Since the land space planning effect picture plays an irreplaceable key role in assisting decision-making, enhancing communication, guiding implementation and result display, etc., when carrying out land space planning, it is generally necessary to generate a land space planning effect picture. At present, CAD drawings are generally first made, and then the CAD two-dimensional planning drawings are cleaned and layered and imported into three-dimensional software such as SketchUp, 3dsMax, etc., to generate building, terrain, road, etc. basic white model according to the design height stretching; then realistic material mapping, landscaping vegetation, characters, vehicles, etc. are arranged, and a sun and sky system simulating natural light is set, and finally a rendering engine such as V-Ray, Lumion, etc. is used to calculate and output the effect picture.

[0004] At present, in the process of generating a land space planning effect picture, an operator needs to make CAD drawings and adjust them using three-dimensional software, and a large amount of manual operation is required at each link, so it often takes several hours or even several days to generate a high-quality effect picture, which is low in efficiency and cannot meet the needs of fast drawing and multiple scheme comparison. SUMMARY

[0005] In order to facilitate the generation of a land space planning effect picture, the present application provides a land space planning effect picture intelligent generation method and system.

[0006] In a first aspect, the present application provides a land space planning effect picture intelligent generation method, which adopts the following technical solution:

[0007] A land space planning effect picture intelligent generation method, comprising:

[0008] S1: collecting original images input by a user and generation requirements;

[0009] S2: generating a preprocessed image according to the original image;

[0010] S3: carrying out image understanding through the preprocessed image, and then forming image prompt words;

[0011] S4: calling requirement prompt words based on the generation requirements;

[0012] S5: fuse the image prompt word and the demand prompt word to obtain a comprehensive guide word;

[0013] S6: input the comprehensive guide word and the preprocessed image into a preset scene rendering model to generate and output a scene effect diagram.

[0014] By adopting the above technical solution, the original image input by the user and the generated demand can be closely matched with the actual idea of the user, thereby providing a targeted basis for subsequent operations. The image prompt word and the demand prompt word are fused to form a comprehensive guide word, and the preprocessed image is input into a scene rendering model to generate an effect diagram, thereby realizing a process from diversified input to precise output, effectively improving the degree of coincidence between the generated effect diagram and the user's expectation, reducing the workload of manual repeated adjustment, improving the efficiency and accuracy of the generation of the land space planning effect diagram, and thereby facilitating the fast generation of the land space planning effect diagram.

[0015] Optionally, the determination method of the preprocessed image comprises:

[0016] S21: based on the original image, an image type is retrieved;

[0017] S22: whether the image type belongs to a preset CAD base map type is judged;

[0018] S23: if yes, useless elements are removed from the original image, and a contour line is extracted, thereby obtaining a black-and-white draft diagram;

[0019] S24: the black-and-white draft diagram is subjected to a closed integrity check to form a CAD adjustment diagram, and the CAD adjustment diagram is taken as the preprocessed image;

[0020] S25: if no, the original image is subjected to picture preprocessing to form a picture adjustment diagram, and the picture adjustment diagram is taken as the preprocessed image.

[0021] By adopting the above technical solution, the original image type is distinguished, the CAD base map type is subjected to specific processing, useless elements are removed, a contour line is extracted, and closed integrity is checked, thereby effectively simplifying the image content and highlighting the key information, thereby providing a clear and accurate basic graph for subsequent analysis and processing. The picture of the non-CAD base map type is subjected to picture preprocessing, thereby ensuring that the original images of different sources can be converted into preprocessed images conducive to the generation of the effect diagram through appropriate operations, and the adaptability of the method to various image data is enhanced.

[0022] Optionally, after the CAD adjustment diagram is taken as the preprocessed image, the method further comprises:

[0023] S241: input the CAD adjustment map into a preset ControlNet neural network architecture to obtain conditional output results and unconditional output results;

[0024] S242: calculate the conditional output results and the unconditional output results by using a preset inference formula to form a line drawing control condition map, and replace the preprocessed image with the line drawing control condition map.

[0025] By using the above technical solutions, the CAD adjustment map is input into the ControlNet neural network architecture to obtain the conditional output results and the unconditional output results, and then the line drawing control condition map is formed by using the inference formula and the preprocessed image is replaced, so as to further process the CAD image in depth, so that the image contains more control information that can be used to guide subsequent rendering, thereby improving the accuracy and controllability of the final generated land space planning effect picture in structure and details.

[0026] Optionally, the demand prompt word generation method comprises:

[0027] S41: extracting an application scenario and a picture size from the generated demand;

[0028] S42: determining a scene prompt word according to the application scenario;

[0029] S43: determining a size prompt word in combination with the picture size;

[0030] S44: checking whether the generated demand contains a demand style picture input by a user;

[0031] S45: if yes, carrying out image understanding on the demand style picture to form a demand style prompt word;

[0032] S46: combining the scene prompt word, the size prompt word and the demand style prompt word as the demand prompt word;

[0033] S47: if no, combining the scene prompt word and the size prompt word as the demand prompt word.

[0034] By using the above technical solutions, the application scenario, picture size and other key information are extracted from the generated demand, and the scene prompt word and the size prompt word are determined accordingly. If there is a demand style picture, a demand style prompt word can also be generated and combined, so that the user's demand for the land space planning effect picture in terms of application scenario, size specification and specific style is fully considered, and more rich and accurate text guidance can be provided for the scene rendering model, helping to generate an effect picture that better meets the user's use purpose and aesthetic preference.

[0035] Optionally, the scene rendering model construction method comprises:

[0036] S61: Collect a space planning effect drawing data set;

[0037] S62: Classify the space planning effect drawing data set to obtain a single style data set;

[0038] S63: Generate a single style description text for the single style data set;

[0039] S64: Label the single style data set with the single style description text to form a labeled data set;

[0040] S65: Based on the SDXL model as a base model, carry out low-rank adaptive training on the labeled data set to generate a single style rendering model;

[0041] S66: Integrate the single style rendering model to obtain a comprehensive rendering model, and use the comprehensive rendering model as the scene rendering model.

[0042] By adopting the above technical solution, by collecting a space planning effect drawing data set and classifying it to generate a single style data set, after labeling, a single style rendering model is generated by low-rank adaptive training based on the SDXL model. Finally, a comprehensive rendering model is obtained by integration, so that the comprehensive rendering model constructed by learning multiple style features can generate diversified style land space planning effect drawings, meeting the differentiated needs of different projects and different users for the style of effect drawings.

[0043] Optionally, the method for forming the single style data set comprises:

[0044] S621: Retrieve a single planning effect drawing from the space planning effect drawing data set;

[0045] S622: Identify the single thing coverage area value, single thing category and single thing color value in the single planning effect drawing;

[0046] S623: Calculate the area proportion value according to the single thing coverage area value;

[0047] S624: Determine the single thing reference value in combination with the single thing color value, the area proportion value and the single thing category;

[0048] S625: Screen the single planning effect drawing according to each single thing reference value to form a single style data set.

[0049] By adopting the technical scheme, the single style dataset is formed by screening and filtering by identifying the single thing coverage area value, type and color value in the single planning effect picture, calculating the area proportion value and determining the single thing reference value, so as to accurately extract data with similar style characteristics, provide data samples with strong pertinence and consistent style for subsequent model training, and help to train a rendering model with more accurate grasp of a specific style, thereby improving the quality of the model in generating a specific style land space planning effect picture.

[0050] Optionally, the method for determining the single thing reference value comprises:

[0051] S6241: retrieving a thing type reference color interval corresponding to the single thing type;

[0052] S6242: comparing the single thing color value with the thing type reference color interval to determine a color change value;

[0053] S6243: determining an area proportion influence value and an area color adjustment value according to the area proportion value;

[0054] S6244: determining a color influence value based on the area color adjustment value and the color change value;

[0055] S6245: determining a thing type reference value according to the single thing type;

[0056] S6246: combining the color influence value, the thing type reference value and the area proportion influence value to determine a comprehensive reference value, and taking the comprehensive reference value as the single thing reference value.

[0057] By adopting the technical scheme, the thing type reference color interval is retrieved, the color change value is determined by comparison, the related reference values are determined according to the area proportion value, and the comprehensive reference value is finally obtained as the single thing reference value, so that the factors such as thing type, color and area proportion are comprehensively considered, the screening and construction of the single style dataset are more accurate and reasonable, and the fineness of the rendering model trained based on this is improved.

[0058] Optionally, the method for generating the single style rendering model comprises:

[0059] S651: retrieving a data quantity value from the labeled dataset;

[0060] S652: determining a learning rate adjustment value and a training frequency adjustment value according to the data quantity value;

[0061] S653: retrieving a model reference learning rate and a reference training frequency from the SDXL model;

[0062] S654: calculate the sum of the model reference learning rate and the learning rate adjustment value as a learning rate correction value;

[0063] S655: calculate the sum of the reference training times and the training times adjustment value as a training times correction value;

[0064] S656: based on the learning rate correction value, the training times correction value and the labeled data set, the SDXL model is adaptively trained by low rank to form a single training model, and the single training model is used as the single style rendering model.

[0065] By adopting the above technical solution, the learning rate adjustment value and the training times adjustment value are determined according to the data number value of the labeled data set, the correction value is obtained by combining the reference learning rate and the reference training times of the SDXL model, the single style rendering model is generated by low rank adaptive training, and the way of flexibly adjusting the training parameters according to the data characteristics can make the model better adapt to the characteristics of the labeled data set in the training process, improve the model training efficiency and effect, and generate a rendering model that can more accurately reflect the style characteristics in the data set, thereby improving the quality and style fit of the land space planning effect map generation.

[0066] Optionally, after generating and outputting the scene effect map, it further includes:

[0067] S71: collect the image selection area and modification prompt word input by the user;

[0068] S72: select from the scene effect map according to the image selection area to form a scene selection effect map and a selection residual effect map;

[0069] S73: input the scene selection effect map and the modification prompt word into a preset scene rendering model to generate a selection adjustment effect map;

[0070] S74: fuse the selection adjustment effect map and the selection residual effect map to form a region adjustment effect map and output.

[0071] By adopting the above technical solution, the image selection area and modification prompt word input by the user are collected, the scene effect map is partially selected and re-input into the scene rendering model to generate a selection adjustment effect map, and then the selection adjustment effect map is fused with the selection residual effect map to output a region adjustment effect map, thereby providing the user with the ability to locally modify and optimize the generated effect map, meeting the user's demand for adjusting specific regions of the effect map according to new requirements or new ideas in the land space planning process, and enhancing the flexibility and interactivity of the generation method.

[0072] In a second aspect, the present application provides a land space planning effect map intelligent generation system, which adopts the following technical solution:

[0073] An intelligent generation system of land space planning effect picture, comprising:

[0074] The acquisition module is configured to acquire an original image, generate a demand, a spatial planning effect picture dataset, an image selection area, and a modification prompt word.

[0075] The memory stores a program for implementing the method of any one of the first aspect.

[0076] The processor loads and executes the program stored in the memory.

[0077] In summary, the present application has at least one of the following beneficial technical effects:

[0078] 1. By acquiring the original image and the generated demand input by the user, the actual idea of the user can be closely matched, and a targeted basis is provided for subsequent operation. The image prompt word and the demand prompt word are fused to form a comprehensive guide word, and the preprocessed image is input into a scene rendering model to generate an effect picture, realizing the process from diversified input to accurate output, effectively improving the consistency of the generated effect picture with the user's expectation, reducing the workload of manual repeated adjustment, improving the efficiency and accuracy of the generation of the land space planning effect picture, and thus facilitating the fast generation of the land space planning effect picture;

[0079] 2. By distinguishing the type of the original image, the CAD base map type is processed specifically, useless elements are removed, contour lines are extracted, and the closure integrity is checked, which can effectively simplify the image content and highlight the key information, providing a clear and accurate basic graph for subsequent analysis and processing. For non-CAD base map type, the picture is preprocessed to ensure that the original images from different sources can be converted into preprocessed images that are beneficial to the generation of effect pictures through appropriate operations, thereby enhancing the adaptability of the method to various image data;

[0080] 3. By acquiring the image selection area and the modification prompt word input by the user, the scene effect picture is partially selected and input into the scene rendering model to generate a selected and adjusted effect picture, which is then fused with the remaining effect picture to output a regional adjustment effect picture, thereby providing the user with the ability to locally modify and optimize the generated effect picture, meeting the user's demand for adjusting the specific area of the effect picture according to new demands or new ideas in the land space planning process, and enhancing the flexibility and interactivity of the generation method. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 is a method flowchart of intelligent generation of land space planning effect picture;

[0082] Figure 2 is a determination method flowchart of preprocessed image;

[0083] Figure 3 is a generation method flowchart of demand prompt words;

[0084] Figure 4 is a construction method flowchart of a scene rendering model. DETAILED DESCRIPTION

[0085] The application will be described in further detail below with reference to the accompanying drawings and embodiments.

[0086] An intelligent generation method of land space planning effect pictures, by collecting the original image input by the user and the generation demand, distinguishing the CAD base map and the non-CAD base map type for the original image respectively for pretreatment, then carrying out image understanding through the pretreated image to form an image prompt word, extracting the application scene and the picture size from the generation demand to combine into a demand prompt word, and fusing the two into a comprehensive guide word; meanwhile, taking the SDXL model as the basis, generating a single style rendering model through the space planning effect picture data set classification, labeling and low rank adaptive training, and obtaining a scene rendering model after integration, inputting the comprehensive guide word and the pretreated image into the model to generate a scene effect picture, and supporting subsequent local area adjustment and optimization of the effect picture according to the user demand, so as to facilitate the fast generation of land space planning effect pictures.

[0087] Reference Figure 1 , the embodiment of the application discloses an intelligent generation method of land space planning effect pictures, which comprises:

[0088] S1: collecting the original image input by the user and the generation demand.

[0089] Among them, the original image refers to the initial image provided by the user in the land space planning effect picture generation scene for serving as the generation basis. The type of the original image includes a CAD base map or a real picture.

[0090] The generation demand refers to various expectations and requirements proposed by the user for the land space planning effect picture. The generation demand includes an application scene, a picture size, a style preference and an image angle, etc.

[0091] The original image is obtained after being actively uploaded to the system by the user, and the generation demand is obtained after being actively uploaded or selecting the options in the preset webpage interface by the user.

[0092] S2: generating a pretreated image according to the original image.

[0093] Among them, the pretreated image refers to the image obtained after the original image input by the user is processed.

[0094] The original image is processed to generate a pretreated image, which is convenient for subsequent generation of effect pictures.

[0095] To further ensure the rationality of the preprocessed image, further separate analysis and calculation of the preprocessed image are required, which are described in detail by the following steps.

[0096] Referring to Figure 2 The determination method of the preprocessed image includes the following steps:

[0097] S21: Retrieving the image type based on the original image.

[0098] The image type refers to the format of the original image and the type corresponding to the content characteristics. The image type includes CAD base map type or real picture type.

[0099] The format of the original image is retrieved as the image type.

[0100] S22: Determine whether the image type belongs to the preset CAD base map type. If yes, execute S23; if no, execute S25.

[0101] The CAD base map type refers to a pre-defined image category that meets the CAD design specification and spatial planning purpose. The CAD base map type is set by the operator in advance.

[0102] By determining whether the image type belongs to the preset CAD base map type, subsequent processing of the image type is facilitated.

[0103] S23: Removing useless elements from the original image and extracting contour lines to obtain a black and white line drawing.

[0104] The useless elements refer to information in the original image that is irrelevant to the land space planning structure and will interfere with subsequent contour extraction and rendering. Common useless elements include annotated text (such as building area value, plot name annotation), dimension marking line (such as dimension line and arrow indicating road width), layer auxiliary line (such as grid line, reference datum line), redundant graphics (such as design draft traces, temporary annotation symbols), etc.

[0105] The contour line refers to a line that can reflect the core structure of the land space planning, such as plot boundary line, road center line / edge line, building outer wall contour line, water system shoreline, etc., which is a key visual element that embodies the planning space layout.

[0106] The black and white line drawing refers to an image that only retains black contour lines and white background after removing useless elements.

[0107] When the image type belongs to the preset CAD base type, it is indicated that this is a CAD base type, so the original image is scanned and matched with the preset useless element feature library. When the element in the feature library is matched, it is directly marked and deleted, and the lineart edge extraction algorithm is used to extract the contour line, so as to obtain a black and white line drawing, which is convenient for subsequent use.

[0108] The useless element feature library pre-stores features such as the font type of the text label, the "line + arrow" combination form of the size line, the light color / dotted line attribute of the auxiliary line, etc. The useless element feature library is pre-set by the operator.

[0109] S24: Perform a closed integrity check on the black and white line drawing to form a CAD adjustment drawing, and use the CAD adjustment drawing as a preprocessing image.

[0110] The CAD adjustment drawing refers to an image formed by correcting the contour line that is not closed in the black and white line drawing.

[0111] The breakpoints and notches in the black and white line drawing are detected, and the positions with breakpoints and notches are filled to form a CAD adjustment drawing. The CAD adjustment drawing is used as a preprocessing image, thereby improving the accuracy of the obtained preprocessing image.

[0112] In order to further ensure the rationality of the preprocessing image, it is necessary to make further separate analysis and calculation on the preprocessing image. The specific steps are as follows.

[0113] After using the CAD adjustment drawing as a preprocessing image, the following steps are further included.

[0114] S241: Input the CAD adjustment drawing into the preset ControlNet neural network architecture to obtain conditional output results and unconditional output results.

[0115] The ControlNet neural network architecture refers to a neural network structure pre-deployed in the land space planning effect drawing generation system and used for adding spatial condition control. The ControlNet neural network architecture is used to receive control images such as the CAD adjustment drawing, to provide accurate structural constraints for subsequent image generation, and to avoid structural deviation in the generated effect drawing. The ControlNet neural network architecture is pre-set by the operator.

[0116] The conditional output result refers to a predicted noise result generated by the model in combination with the structure characteristics of the CAD adjustment drawing as a spatial structure constraint condition input into the ControlNet. The unconditional output result refers to a predicted noise result output by the ControlNet based on the default generation logic of the base model (such as SDXL) without inputting additional constraint conditions such as the CAD adjustment drawing.

[0117] The conditional output result and the unconditional output result are obtained by inputting the CAD adjustment drawing into the preset ControlNet neural network architecture, which facilitates subsequent use.

[0118] S242: The conditional output result and the unconditional output result are calculated using a preset inference formula to form a line drawing control condition drawing, and the line drawing control condition drawing is used to replace the preprocessed image.

[0119] The inference formula refers to a preset calculation formula for fusing the conditional output result and the unconditional output result.

[0120] The line drawing control condition drawing refers to an image that strengthens the structure constraint of the CAD adjustment drawing obtained by calculating the preset inference formula.

[0121] The inference formula is specifically: the final output predicted noise = unconditional output result + hyperparameter * (conditional output result - unconditional output result).

[0122] The hyperparameter refers to a parameter for controlling the strength of the structure constraint, which is preset by the operator according to actual needs. The hyperparameter usually takes a value range of 0.5 to 2.0, and the higher the structure requirement, the larger the value of the hyperparameter.

[0123] By calculating the conditional output result and the unconditional output result using the inference formula, the final output predicted noise is obtained, and then the decoding function of the VAE (Variational Autoencoder) is used to map the final output predicted noise to image features in the pixel space and perform image reconstruction to form a line drawing control condition drawing with a white background, a black outline, and a clearer edge and more accurate structure. The line drawing control condition drawing is used to replace the preprocessed image, thereby improving the accuracy of the obtained preprocessed image.

[0124] S25: The original image is subjected to picture preprocessing to form a picture adjustment drawing, and the picture adjustment drawing is used as the preprocessed image.

[0125] The picture adjustment drawing refers to an image obtained by subjecting a non-CAD base image to picture preprocessing.

[0126] When the image type does not belong to the preset CAD base map type, it is indicated that the image type is a real picture type at this time, so the format (such as JPEG, PNG, BMP) of the original image is read and converted into a preset standard format, and the picture is subjected to denoising processing, color correction, contrast enhancement and other quality optimization, so as to form a picture adjustment graph, and the picture adjustment graph is used as a pretreatment image, so as to improve the accuracy of the obtained pretreatment image.

[0127] In the embodiment, the standard format is set as a GeoTIFF format, and geographical coordinate information (such as latitude and longitude, projection coordinate system) of the image can be retained, so as to facilitate subsequent matching with geographical data (such as a plot boundary vector graph) of land space planning.

[0128] S3: image understanding is carried out through the pretreatment image, and then an image prompt word is formed.

[0129] The image prompt word refers to a text sentence describing the core content of the pretreatment image. The image prompt word includes information such as a ground object type (such as a building, a road, and a green land), a spatial layout (such as buildings distributed along both sides of a road), and a quantity feature (such as 3 multi-storey houses), which can assist a scene rendering model in accurately capturing structure and content features of the pretreatment image, and is an important part of generating a comprehensive guide word.

[0130] The pretreatment image is input into a preset image conversion language model, so as to identify each feature in the pretreatment image, and then the identified visual features are converted into a structured text description, and the semantic analysis result is integrated into a natural language prompt word according to the logic of the ground object type, the spatial layout, and the key feature, so as to obtain the image prompt word, which is convenient for subsequent use.

[0131] When the pretreatment image is a CAD base map type, the contour line corresponding ground object type (by matching line shape and layout, such as rectangular contour + dense distribution for residential buildings, and long strip continuous line for roads), spatial layout relationship (such as a road penetrating through the center of a region, and a building located on the east side of a road), and special identification (such as a closed curve + internal texture for an artificial lake) are identified.

[0132] When the pretreatment image is a real picture type, real scene ground objects (by matching texture and shape, such as a green continuous region for a green land, and a gray regular region for a building roof), ground object states (such as a 2-storey slope roof structure for a building, and an asphalt pavement for a road), and geographical correlations (such as a river extending along the north side boundary of a region) are identified.

[0133] For example, when identifying the CAD adjustment map of village planning, "5 rectangular residential buildings (distributed along the south side of the east-west road), 1 8-meter-wide main road (through the village), and 1 circular central green land (located on the north side of the middle section of the main road)" are identified. At this time, the visual feature of "5 rectangular residential buildings distributed along the south side of the east-west road" is analyzed as the text semantics of "5 multi-story residential buildings linearly distributed on the south side of the east-west main road with a building spacing of about 10 meters". After optimization, the image prompt word is formed as "national space planning scene, including 5 multi-story residential buildings (rectangular contour, linearly distributed on the south side of the east-west main road with a building spacing of about 10 meters), 1 8-meter-wide east-west main road (through the whole village), and 1 circular central green land (diameter about 20 meters, located on the north side of the middle section of the main road)".

[0134] S4: Retrieving demand prompt words based on generated demand.

[0135] The demand prompt word refers to a text description used to clearly express the user's dimensional demand for application scenarios, size specifications, and style preferences in the national space planning effect map.

[0136] By retrieving application scenarios, size specifications, and style preferences from the generated demand, and then converting each demand into a prompt word, the demand prompt word is formed, which is convenient for subsequent use.

[0137] In order to further ensure the rationality of the demand prompt word, it is necessary to make a further separate analysis and calculation of the demand prompt word. The specific steps are as follows.

[0138] Referring to Figure 3 , the demand prompt word generation method includes the following steps:

[0139] S41: Extracting application scenarios and picture sizes from generated demand.

[0140] The generated demand includes application scenarios and picture sizes. Application scenarios refer to the specific use scenarios of the user generating the national space planning effect map, reflecting the purpose and use environment of the effect map. Application scenarios include residential planning general plan, residential planning bird's eye view, commercial planning bird's eye view, industrial planning bird's eye view, educational planning bird's eye view, residential renovation map, river regulation map, road regulation map, etc.

[0141] Picture size refers to the pixel resolution or corresponding physical paper size of the national space planning effect map expected by the user. Picture size is, for example, 1024x1024, 2048x2048, 2480x3508, 2100x2970, etc.

[0142] S42: Determining scene prompt words according to application scenarios.

[0143] The scene prompt word refers to a text description generated based on the extracted application scene, which is used to clearly adapt the scene demand of the land space planning effect picture.

[0144] By inputting the application scene into the preset scene prompt database, the scene prompt word is matched and obtained, which is convenient for subsequent use.

[0145] The scene prompt database pre-stores a comparison table of different application scenes and corresponding scene prompt words, and the scene prompt database is pre-set by the operator according to actual demand.

[0146] For example, in the scene prompt database, when the application scene is a residential planning general plan, the scene prompt word is "high-rise apartment buildings, low-rise villas, pedestrian paths, underground parking lot entrances, and clearly classified roads (main roads, secondary roads, branch roads), environment represented by minimalist tree icons and green lawns, sunlight environment, etc.".

[0147] When the application scene is a commercial planning bird's eye view, the scene prompt word is "bird's eye view of the bustling central business district (CBD), fashionable glass steel structure skyscrapers, multi-level overpasses, main roads with traffic and streetlights, central squares, subway station entrances, etc.".

[0148] S43: Determine the size prompt word in combination with the picture size.

[0149] The size prompt word refers to a text description generated for the size of the picture size.

[0150] By converting the picture size into the format of the prompt word and taking it as the size prompt word, subsequent use is facilitated.

[0151] For example, when the picture size is 1024x1024, the size prompt word is "1024x1024 pixels, 1:1 aspect ratio".

[0152] S44: Check whether the demand style picture input by the user is included in the generated demand. If yes, perform S45; if no, perform S47.

[0153] The demand style picture refers to an image provided by the user in the generated demand, which is used to specify the style reference of the land space planning effect picture. The demand style picture provides an intuitive style template (such as example pictures of realistic style, hand-drawn style, and technology line drawing style) for the scene rendering model, helping the model accurately match the visual style expected by the user. It is a key reference basis for supplementing the text style description and improving the style matching degree.

[0154] By checking whether the demand style picture input by the user is included in the generated demand, subsequent generation of the effect picture is facilitated.

[0155] S45: image understanding is performed on the demand style diagram to form a demand style prompt word.

[0156] The demand style prompt word refers to a text description converted from visual style features (such as color tone, line style, and texture performance) extracted from the demand style diagram provided by the user.

[0157] When the demand contains the demand style diagram input by the user, it indicates that the demand style can be converted at this time, so the demand style diagram is input into a preset image conversion language model to identify each feature in the demand style diagram, and then the identified visual features are converted into a structured text description, and the semantic analysis result is integrated into a natural language prompt word according to the logic of the feature type, spatial layout, and key features, so as to obtain the demand style prompt word for subsequent use.

[0158] S46: The scene prompt word, the size prompt word, and the demand style prompt word are combined and used as a demand prompt word.

[0159] The complete text instruction formed by logically integrating the scene prompt word, the size prompt word, and the demand style prompt word is used as the demand prompt word, thereby improving the accuracy of the obtained demand prompt word.

[0160] S47: The scene prompt word and the size prompt word are combined and used as a demand prompt word.

[0161] When the demand does not contain the demand style diagram input by the user, it indicates that only the scene size can be converted at this time, so the complete text instruction formed by logically integrating the scene prompt word and the size prompt word is used as the demand prompt word, thereby improving the accuracy of the obtained demand prompt word.

[0162] S5: The image prompt word and the demand prompt word are fused to obtain a comprehensive guide word.

[0163] The comprehensive guide word refers to a text instruction used to guide the rendering adjustment of an image.

[0164] The complete text instruction formed by logically fusing the image prompt word reflecting the core information of the preprocessed image and the demand prompt word covering the scene adaptation, size specification, and style features is used as the comprehensive guide word, thereby facilitating subsequent use.

[0165] S6: The comprehensive guide word and the preprocessed image are input into a preset scene rendering model to generate and output a scene effect diagram.

[0166] The scene rendering model refers to a preset image generation model based on deep learning in the land space planning effect map generation system. The scene effect map refers to a final presented land space planning visualization image.

[0167] By inputting the comprehensive guide word and the preprocessed image into the preset scene rendering model, the preprocessed image is diffused and rendered according to the comprehensive guide word, so that the scene effect map is obtained and output, facilitating the generation of the land space planning effect map.

[0168] In order to further ensure the rationality of the scene rendering model, further separate analysis and calculation of the scene rendering model are required, which will be described in detail by the following steps.

[0169] Reference Figure 4 The construction method of the scene rendering model includes the following steps:

[0170] S61: Collect the space planning effect map data set.

[0171] The space planning effect map data set refers to the data set of the visualization effect map of the urban, rural, park and other space planning scenes. The visualization effect map includes the planning scheme map, the rendering effect map, the real scene comparison map and the like.

[0172] The visualization effect map is collected through academic databases, government and industry platforms, open source communities and the like, and the collected data is combined to form the space planning effect map data set, facilitating subsequent use.

[0173] S62: Classify the space planning effect map data set to obtain a single style data set.

[0174] The single style data set refers to the data set formed after division and screening.

[0175] The space planning effect map data set is classified to screen out data subsets with the same characteristics as the single style data set, facilitating subsequent use.

[0176] In order to further ensure the rationality of the single style data set, further separate analysis and calculation of the single style data set are required, which will be described in detail by the following steps.

[0177] The formation method of the single style data set includes the following steps:

[0178] S621: Retrieve a single planning effect map from the space planning effect map data set.

[0179] The single planning effect map refers to a single visualization effect map of the urban, rural, park and other space planning scenes.

[0180] By calling each rendering of the spatial planning rendering data set and taking it as a single planning rendering, subsequent use is facilitated.

[0181] S622: Identify the single planning rendering single thing coverage area value, single thing category and single thing color value.

[0182] Among them, the single thing coverage area value refers to the pixel area or actual geographic area occupied by a specific feature (such as a single building, a piece of green land, a road) in the image in the single planning rendering.

[0183] The single thing category refers to the specific feature category identified from the single planning rendering. Common types in the spatial planning scene include buildings (multi-story residential buildings, high-rise commercial buildings), green land (park green land, residential green land), roads (urban trunk roads, branch roads), water bodies (artificial lakes, rivers), etc.

[0184] The single thing color value refers to the color parameter presented by the surface of a specific feature in the single planning rendering. The single thing color value is represented by RGB (red, green, blue) three-color channel value (such as RGB(255,255,255) representing white) or HSV (hue, saturation, lightness) parameter, reflecting the visual color characteristics of the feature.

[0185] By identifying the categories of each thing in the single planning rendering to obtain the single thing category, calculating the area value occupied by the single thing as the single thing coverage area value, and extracting the color value of the single thing to obtain the single thing color value, subsequent use is facilitated.

[0186] S623: Calculate the area ratio value according to the single thing coverage area value.

[0187] Among them, the area ratio value refers to the percentage of the coverage area value of a specific feature in the single planning rendering to the total area of the rendering picture.

[0188] By calling the total area corresponding to the single planning rendering as the total area value of the rendering, calculating the ratio value between the single thing coverage area value and the total area value of the rendering, and taking the calculation result as the area ratio value, subsequent use is facilitated.

[0189] S624: Determine the single thing reference value in combination with the single thing color value, the area ratio value and the single thing category.

[0190] Among them, the single thing reference value refers to the parameter value after quantifying the existence of a specific feature in the spatial planning scene.

[0191] The single thing reference value is determined by combining the color value, the area ratio value and the single thing category, so as to facilitate subsequent classification.

[0192] In order to further ensure the rationality of the single thing reference value, it is necessary to make further separate analysis and calculation on the single thing reference value. The specific steps are as follows.

[0193] The method for determining the single thing reference value comprises the following steps:

[0194] S6241: retrieve the thing category reference color interval corresponding to the single thing category.

[0195] The thing category reference color interval refers to the color parameters presented by the single thing category under normal circumstances.

[0196] The single thing category is input into the preset category reference color database to match the thing category reference color interval, so as to facilitate subsequent use.

[0197] The category reference color database pre-stores a comparison table of different single thing categories and corresponding thing category reference color intervals, and the category reference color database is pre-set by the operator.

[0198] For example, the category reference color database can be set as follows: when the single thing category is residential area central green land, the thing category reference color interval is (20, 100, 20)~(50, 160, 50); when the single thing category is urban trunk road (lane), the thing category reference color interval is (50, 50, 50)~(80, 80, 80); and when the single thing category is sidewalk (paving), the thing category reference color interval is (180, 180, 180)~(220, 220, 220).

[0199] S6242: compare the single thing color value with the thing category reference color interval to determine the color change value.

[0200] The color change value refers to the change value corresponding to the color deviation.

[0201] When the single thing color value is located in the thing category reference color interval, the change value is output as (0, 0, 0) at this time. When the single thing color value is not located in the thing category reference color interval, the nearest boundary value of the single thing color value and the thing category reference color interval is calculated to obtain R change value, G change value and B change value, so as to obtain the color change value, thereby facilitating subsequent use.

[0202] S6243: determine the area ratio influence value and the area color adjustment value according to the area ratio value.

[0203] The area proportion influence value refers to the influence degree value of the area proportion on the existence of the single thing category. The area color adjustment value refers to the reference value corresponding to the adjustment of the color change according to the area condition. The greater the area proportion value, the greater the area proportion influence value, and the greater the area color adjustment value.

[0204] The area proportion influence value is calculated by multiplying the area proportion value by the preset proportion influence coefficient, and the area color adjustment value is calculated by multiplying the area proportion value by the preset proportion adjustment coefficient.

[0205] The proportion influence coefficient is a coefficient for converting the area proportion value into the area proportion influence value, and the proportion adjustment coefficient is a coefficient for converting the area proportion value into the area color adjustment value. The proportion influence coefficient and the proportion adjustment coefficient are both preset by the operator according to actual needs.

[0206] For example, the proportion influence coefficient can be set to 1, and the proportion adjustment coefficient can be set to 0.8. When the area proportion value is 0.7, the area proportion influence value is 0.7, and the area color adjustment value is 0.56.

[0207] S6244: Determine the color influence value based on the area color adjustment value and the color change value.

[0208] The color influence value refers to the influence degree value of the color change on the existence of the single thing category.

[0209] The area color adjustment value is calculated by multiplying the area color adjustment value by the R change value, the G change value, and the B change value in the color change value, respectively. Then, the average of the three product values is calculated as the color influence value, which is convenient for subsequent use.

[0210] S6245: Determine the thing category reference value according to the single thing category.

[0211] The thing category reference value refers to the parameter value corresponding to the initial quantification of the existence of the single thing category.

[0212] The single thing category is input into the preset thing category database to match the thing category reference value, which is convenient for subsequent use.

[0213] The thing category database pre-stores a comparison table of different single thing categories and corresponding thing category reference values. The thing category database sets the corresponding thing category reference values for different single thing categories in advance by the operator according to actual needs.

[0214] S6246: Combine the color influence value, the thing category reference value, and the area proportion influence value to determine a comprehensive reference value, and take the comprehensive reference value as a single thing reference value.

[0215] The comprehensive reference value refers to a quantitative parameter value corresponding to the combination of the category, color, and area proportion of a single thing.

[0216] By calculating the sum value between the color influence value and the thing category reference value, and then calculating the product value between the sum value and the area proportion influence value, the final calculation result is taken as the comprehensive reference value, and the comprehensive reference value is taken as the single thing reference value, thereby improving the accuracy of the obtained single thing reference value.

[0217] S625: Screen single planning effect pictures according to the single thing reference values to form a single style data set.

[0218] The reference deviation value is calculated by calculating the difference between each single thing reference value, and the single planning effect picture corresponding to the reference deviation value that meets the preset reference deviation interval is selected to form a single style data set, thereby improving the accuracy of the obtained single style data set.

[0219] The reference deviation interval refers to a reference deviation interval for selecting similar styles, and the reference deviation interval is set by an operator according to actual needs in advance.

[0220] S63: Generate a single style description text for the single style data set.

[0221] The single style description text refers to a structured text formed by extracting the visual common features (color, line, texture, light and shadow, etc.) of each effect picture in the single style data set.

[0222] By extracting the thing category, color, line, texture, and light and shadow in each effect picture in the single style data set, and forming realistic, hand-drawn, and technical line drawing styles through color, line, texture, and light and shadow, and then combining with the thing category to form a single style description text, it is convenient for subsequent use.

[0223] S64: Label the single style data set with the single style description text to form a labeled data set.

[0224] The labeled data set refers to a data set corresponding to the text labeling of the effect picture.

[0225] By labeling each effect picture corresponding to the single style data set with the single style description text, a labeled data set is formed, which is convenient for subsequent use.

[0226] S65: Based on the SDXL model as the base model, low-rank adaptive training is carried out on the labeled data set to generate a single style rendering model.

[0227] The single style rendering model refers to a model for diffusing and rendering a single style effect drawing.

[0228] By taking the SDXL model as the base model, the labeled data set is converted into a training format readable by the SDXL model, and the training parameters are configured for low-rank adaptive training to generate a single style rendering model, which is convenient for subsequent use.

[0229] In order to further ensure the rationality of the single style rendering model, it is necessary to make further separate analysis and calculation on the single style rendering model. The specific steps are as follows.

[0230] The generation method of the single style rendering model includes the following steps:

[0231] S651: Retrieve the data quantity value from the labeled data set.

[0232] The data quantity value refers to the quantity value corresponding to the effect drawing in the labeled data set.

[0233] The effect drawing in the labeled data set is counted, and the counting result is taken as the data quantity value.

[0234] S652: Determine the learning rate adjustment value and the training times adjustment value according to the data quantity value.

[0235] The learning rate adjustment value refers to the adjustment value corresponding to the adjustment of the learning rate.

[0236] The training times adjustment value refers to the adjustment value corresponding to the adjustment of the training times.

[0237] The data quantity value is matched with the preset quantity adjustment database to obtain the learning rate adjustment value and the training times adjustment value, which is convenient for subsequent use.

[0238] The quantity adjustment database pre-stores a control table of different data quantity intervals and corresponding learning rate adjustment values and training times adjustment values. The quantity adjustment database is pre-set by the operator according to actual needs.

[0239] For example, the database size adjustment can be set as follows: when the number of data points is less than 300, the learning rate is adjusted to 3e-5 to 5e-5, and the number of training iterations is adjusted to 20-30 epochs. When the number of data points is between 300 and 2000, the learning rate is adjusted to 6e-5 to 1.2e-4, and the number of training iterations is adjusted to 8-15 epochs. When the number of data points is greater than 2000, the learning rate is adjusted to 1.2e-4 to 1.5e-4, and the number of training iterations is adjusted to 5-8 epochs.

[0240] S653: Retrieves the baseline learning rate and baseline training iterations from the SDXL model.

[0241] The baseline learning rate refers to the initial learning rate of the SDXL model. The baseline number of training iterations refers to the initial number of training iterations for the SDXL model.

[0242] The baseline learning rate and number of training iterations can be retrieved using the SDXL model for later use.

[0243] S654: Calculate the sum of the model baseline learning rate and the learning rate adjustment value, and use it as the learning rate correction value.

[0244] The learning rate adjustment value refers to the adjustment value corresponding to the learning rate after adjustment.

[0245] The sum of the model's baseline learning rate and the learning rate adjustment value is calculated, and the result is used as the learning rate correction value for convenient subsequent use.

[0246] S655: Calculate the sum of the baseline number of training iterations and the training iteration adjustment value, and use it as the training iteration correction value.

[0247] Among them, the training number correction value refers to the correction value corresponding to the adjustment of the training number.

[0248] The sum of the baseline number of training iterations and the training iteration adjustment value is calculated, and the result is used as the training iteration correction value for convenient subsequent use.

[0249] S656: Based on the learning rate correction value, training number correction value, and labeled dataset, the SDXL model is subjected to low-rank adaptive training to form a single training model, and the single training model is used as a single-style rendering model.

[0250] Here, a single-trained model refers to the model corresponding to a single training iteration. By using learning rate correction values ​​and training iteration correction values ​​as training parameters, and inputting the labeled dataset into the SDXL model for low-rank adaptation training, a single-trained model is formed. This single-trained model is then used as a single-style rendering model, improving the accuracy of the obtained single-style rendering model.

[0251] S66: Integrate single-style rendering models to obtain a comprehensive rendering model, and use the comprehensive rendering model as the scene rendering model.

[0252] Among them, the integrated rendering model refers to the model that is formed by integrating various individual style models.

[0253] By performing compatibility checks on each individual style rendering model and then integrating the weights of multiple individual style rendering models into a comprehensive LoRA weight file, it is possible to trigger the corresponding style through prompt words or keywords, thereby forming a comprehensive rendering model. This comprehensive rendering model is then used as the scene rendering model, improving the accuracy of the obtained scene rendering model.

[0254] To further ensure the rationality of the generated and output scene renderings, it is necessary to perform further separate analysis and calculations after the scene renderings are generated and output. The specific steps are explained in detail below.

[0255] After generating and outputting the scene rendering, the following steps are also included:

[0256] S71: Collect the image selection area and modification prompts input by the user.

[0257] The image selection area refers to the area corresponding to the scene effect image selected by the user.

[0258] Modification prompts refer to text commands entered by the user that describe how to make changes.

[0259] The image selection area is obtained by the user using tools such as the rectangular selection tool, polygonal selection tool, and brush selection tool. The modification prompt is obtained through user input.

[0260] S72: Select from the scene rendering based on the image selection area to form a scene selection rendering and a selection remaining rendering.

[0261] The scene selection rendering refers to the image corresponding to the selected area. The remaining selection rendering refers to the image outside the selected area.

[0262] By selecting the portion corresponding to the selected area in the scene effect image and using it as the scene selection effect image, and using the portion of the scene effect image other than the scene selection effect image as the remaining selection effect image, it is convenient for subsequent use.

[0263] S73: Input the scene selection effect image and modification prompts into the preset scene rendering model to generate the selection adjustment effect image.

[0264] The adjustment effect picture refers to an effect picture corresponding to the selected region after adjustment and modification.

[0265] The scene selection effect picture and the modification prompt word are input into a preset scene rendering model, so as to regenerate an effect picture according to the modification prompt word for a part corresponding to the scene selection effect picture to obtain a selected adjustment effect picture, facilitating subsequent use.

[0266] S74: The selected adjustment effect picture and the selected remaining effect picture are fused to form a region adjustment effect picture and output.

[0267] The region adjustment effect picture refers to an overall effect picture corresponding to the selected region after adjustment.

[0268] The selected adjustment effect picture and the selected remaining effect picture are preprocessed to eliminate size and format differences, the selected adjustment effect picture is embedded in a corresponding position in the selected remaining effect picture to realize seamless alignment in space, the splicing boundary is locally optimized to form a region adjustment effect picture and output, the demand of a user for adjusting a specific region of an effect picture according to new demand or new idea in the process of land space planning is met, and the flexibility and interactivity of the generation method are enhanced.

[0269] Based on the same inventive concept, the embodiments of the present application provide a land space planning effect picture intelligent generation system, comprising:

[0270] The acquisition module is configured to acquire an original image, generation demand, a spatial planning effect picture dataset, an image selection region and a modification prompt word.

[0271] The memory stores a program for realizing the land space planning effect picture intelligent generation method.

[0272] The processor loads and executes the program stored in the memory.

[0273] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional modules is taken as an example for illustration, and in actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. The specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0274] The above merely describes the preferred embodiments of the present application, and the protection scope of the present application is not limited to the above-described embodiments. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall also be considered as falling within the protection scope of the present application.

Claims

1. A method for intelligently generating a land space planning effect picture, characterized in that, Comprise: S1: collect user input of the original image and generate demand; S2: generate a preprocessed image according to the original image; S3: through the preprocessed image to carry out image understanding, and then form image prompt word; S4: based on the generation demand call demand prompt word; S5: the image prompt word and the demand prompt word fusion, get comprehensive guide word; S6: the comprehensive guide word and the preprocessed image input preset scene rendering model, generate and output scene effect drawing; After generating and outputting the scene effect drawing, it further comprises: S71: collect user input of image selection area and modification prompt word; S72: according to the image selection area from the scene effect drawing to form scene selection effect drawing and selection remaining effect drawing; S73: the scene selection effect drawing and the modification prompt word input preset scene rendering model, generate selection adjustment effect drawing; S74: the selection adjustment effect drawing and the selection remaining effect drawing fusion, form area adjustment effect drawing and output; The construction method of the scene rendering model comprises: S61: collect space planning effect drawing data set; S62: classify the space planning effect drawing data set to obtain single style data set; S63: generate single style description text for the single style data set; S64: the single style data set is labeled with the single style description text to form a labeled data set; S65: based on the SDXL model as the base model, the labeled data set is trained to generate a single style rendering model; S66: integrate the single style rendering model to obtain a comprehensive rendering model, and the comprehensive rendering model is used as the scene rendering model; The forming method of the single style data set comprises: S621: retrieve single planning effect drawing from the space planning effect drawing data set; S622: identify single thing coverage area value, single thing type and single thing color value in the single planning effect drawing; S623: calculate area ratio value according to the single thing coverage area value; S624: determine single thing reference value in combination with the single thing color value, the area ratio value and the single thing type; S625: according to each single thing reference value, the single planning effect drawing is screened to form a single style data set.

2. The method of claim 1, wherein The determination method of the preprocessed image comprises: S21: based on the original image to retrieve image type; S22: determine whether the image type belongs to the preset CAD base map type; S23: if yes, remove useless elements from the original image and extract contour line to obtain black and white draft drawing; S24: check the completeness of the black and white draft drawing to form a CAD adjustment drawing, and the CAD adjustment drawing is used as the preprocessed image; S25: if no, the original image is preprocessed to form a picture adjustment drawing, and the picture adjustment drawing is used as the preprocessed image.

3. The method of claim 2, wherein, After using the CAD adjustment drawing as the preprocessed image, it further comprises: S241: input the CAD adjustment diagram into a preset ControlNet neural network architecture to obtain a conditional output result and an unconditional output result; S242: calculate the conditional output result and the unconditional output result by using a preset inference formula to form a sketch control condition diagram, and replace the preprocessed image with the sketch control condition diagram.

4. The method of claim 1, wherein The generation method of the demand prompt word comprises: S41: extract an application scenario and a picture size from the generated demand; S42: determine a scene prompt word according to the application scenario; S43: determine a size prompt word in combination with the picture size; S44: check whether the demand style graph input by a user is contained in the generated demand; S45: if yes, carry out image understanding on the demand style graph to form a demand style prompt word; S46: combine the scene prompt word, the size prompt word and the demand style prompt word and take them as the demand prompt word; S47: if no, combine the scene prompt word and the size prompt word and take them as the demand prompt word.

5. The method of claim 1, wherein, The determination method of the single thing reference value comprises: S6241: call a thing category reference color interval corresponding to the single thing category; S6242: compare the single thing color value with the thing category reference color interval to determine a color change value; S6243: determine an area proportion influence value and an area color adjustment value according to the area proportion value; S6244: determine a color influence value based on the area color adjustment value and the color change value; S6245: determine a thing category reference value according to the single thing category; S6246: combine the color influence value, the thing category reference value and the area proportion influence value to determine a comprehensive reference value, and take the comprehensive reference value as the single thing reference value.

6. The method of claim 1, wherein The generation method of the single style rendering model comprises: S651: call a data quantity value from the labeled data set; S652: determine a learning rate adjustment value and a training frequency adjustment value according to the data quantity value; S653: call a model reference learning rate and a reference training frequency from the SDXL model; S654: calculate a sum of the model reference learning rate and the learning rate adjustment value as a learning rate correction value; S655: calculate a sum of the reference training frequency and the training frequency adjustment value as a training frequency correction value; S656: perform low-rank adaptive training on the SDXL model based on the learning rate correction value, the training frequency correction value and the labeled data set to form a single training model, and take the single training model as the single style rendering model.

7. A land space planning effect picture intelligent generation system, characterized in that, Comprise: The acquisition module is used to acquire an original image, a generation demand, a spatial planning effect diagram data set, an image selection region and a modification prompt word; The memory stores a program for implementing the land space planning effect diagram intelligent generation method according to any one of claims 1 to 6; The processor loads and executes the program stored in the memory.

Images