Street space quality intelligent evaluation and image generation method and system based on deep learning

Abstract

The invention relates to the field of intelligent city planning, and discloses a street space quality intelligent evaluation and image generation method and system based on deep learning, and the method specifically comprises the following steps: movably collecting a street space image, and recording the geographic position information; performing perspective correction and multi-view image segmentation on the acquired street image data, and performing spatial matching with geographical location information to obtain a fused image database; and performing semantic segmentation on the image in the fused image database based on a deep learning model, calculating objective and subjective visual perception indexes and facility function indexes of the image, and generating a street space quality evaluation result. The method solves the problems that in the prior art, evaluation dimensions are limited, and an automatic response generation mechanism is lacked, and has the advantages of being intelligent, efficient and comprehensive in subjective and objective evaluation.

Description

Technical Field

[0001] The present invention relates to the field of smart city planning, and more specifically, to a method and system for intelligently evaluating and image-generating street space quality based on deep learning. Background Art

[0002] Street space is the most fundamental, extensive, and life-critical component of an urban public space system. It serves as a crucial interface for showcasing a city's landscape, functional complexity, and vitality, directly impacting its image, travel experience, and resident well-being. As the concept of "people-oriented" urban development continues to deepen, the comprehensive assessment and scientific improvement of street space quality has become a core component of numerous planning and design tasks, including urban renewal, county and town construction, and community development. For example, the "Hundreds, Millions, and Tens of Millions Project" explicitly proposes improving the quality of county-level public spaces using streets as a unit, and the "Urban Health Checkup" program also prioritizes street space as a key evaluation target.

[0003] However, traditional street space quality assessment methods rely heavily on manual on-site sampling and subjective qualitative analysis. These methods are limited in evaluation dimensions, inefficient, and lack unified quantitative standards and technical tools. Typically, these assessments involve on-site streetscape imagery, subjective interpretation, and design expression based on conceptual drawings. Lacking intelligent data support and in-depth analysis, these traditional methods struggle to meet the demands of large-scale, sophisticated urban renewal and planning decisions.

[0004] Against the backdrop of the rapid development of big data and artificial intelligence technologies, intelligent street space assessment methods based on street view imagery are gradually emerging. Deep learning-driven image semantic understanding algorithms can automatically identify spatial elements in street view imagery and quantify environmental characteristics, providing a viable path for constructing objective, comparable, and reusable quality evaluation systems. Emerging artificial intelligence generated content (AIGC) technology is also beginning to be applied to assist in urban space design, generating streetscape renovation renderings based on existing data, providing intuitive reference for design decisions. Related research has found initial application in scenarios such as urban landscape improvement, beautiful neighborhood construction, and community development. However, several bottlenecks remain: street view image acquisition relies on data from third-party platforms, resulting in untimely updates and limited coverage; comprehensive evaluation systems for visual perception and spatial function are still immature; deep learning assessment results are disconnected from the public's subjective perception; and there is a lack of automatic diagnosis and generative design response mechanisms for street space quality issues.

[0005] The prior art provides a street space quality evaluation method based on multi-feature fusion of street view images, comprising: obtaining street view images from multiple sampling points at multiple perspectives; performing quality labeling on each of the street view images; extracting multiple image features from each of the street view images; at each perspective, using at least one machine learning algorithm to perform model training based on a single image feature and a combination of multiple image features to construct multiple street space quality evaluation models; selecting the optimal evaluation model at each perspective based on model performance metrics; and using the optimal model to evaluate the street space quality of any area, thereby constructing a large-scale, high-precision street space quality measurement system.

[0006] However, existing technologies have problems such as limited evaluation dimensions and lack of an automated response generation mechanism. Therefore, how to invent an intelligent evaluation method for street space quality with multiple evaluation dimensions and an automated response generation mechanism is a technical problem that urgently needs to be solved in this technical field. Summary of the Invention

[0007] In order to solve the problems of limited evaluation dimensions and lack of automatic response generation mechanism in existing technologies, the present invention provides an intelligent evaluation and image generation method and system for street space quality based on deep learning, which is intelligent and efficient, and integrates subjective and objective evaluation.

[0008] In order to achieve the above-mentioned purpose of the present invention, the technical solutions adopted are as follows:

[0009] A deep learning-based intelligent assessment method for street space quality includes the following specific steps:

[0010] Mobile collection of street space images and recording of geographic location information;

[0011] Perform perspective correction and multi-view image segmentation on the collected street image data, and perform spatial matching with geographic location information to obtain a fused image database;

[0012] Based on the deep learning model, semantic segmentation is performed on the images in the fusion image database, and their objective and subjective visual perception indicators and facility function indicators are calculated to generate street space quality evaluation results.

[0013] Preferably, street space images are collected on the move and geographic location information is recorded, specifically by building a mobile sensing device including a panoramic camera, a GPS recorder, a vehicle-mounted suction cup bracket and a monitoring screen, continuously shooting panoramic street images at predetermined time intervals, and recording the geographic location and time information at the time of shooting to generate a street image dataset.

[0014] Furthermore, when collecting street space images, a street network diagram is constructed and the optimal collection route P is used. opt To collect, Popt satisfy:

[0015]

[0016] Among them, d(p i ,p i+1 ) is the distance between adjacent street points.

[0017] Furthermore, the collected street space images are subjected to perspective correction and multi-view image segmentation, and spatial matching is performed with the geographic location information to obtain a fused image. The specific steps are as follows:

[0018] Construct a street network diagram and further construct spatial coordinates;

[0019] The images in the street image data are perspective corrected and projected into spatial coordinates to achieve mapping:

[0020]

[0021] Where, (u, v) is the image pixel coordinate; (x, y) is the image coordinate after plane projection; f is the equivalent focal length; f x , f y are the scaling factors of the image in the X and Y directions respectively; c x , c y is the principal point coordinate of the camera, that is, the center point of the image;

[0022] The mapped image is divided into a set of images of four perspectives: front / back / left / right V′={I f ,I b ,I l ,I r};

[0023] The shooting time code of the image is matched with its GPS trajectory to generate a dataset that binds the image set to the geographic coordinates, forming a fused image database with spatial attributes.

[0024] Furthermore, the images in the fusion image database are semantically segmented based on a deep learning model, and their objective and subjective visual perception indicators and facility function indicators are calculated to generate street space quality evaluation results. The specific steps are as follows:

[0025] The deep semantic segmentation model SegFormer based on Transformer structure is used to collect street view images I i Perform semantic segmentation to identify the spatial element categories in the image and obtain pixel-level semantic label image S i , calculate S i Objective visual feature indicators in ;

[0026] Based on the crowdsourcing mechanism, a sample set of street view images is collected. Each image is scored independently by at least m participants, and a weighted average method is used to generate a subjective reference score:

[0027]

[0028] Among them, s ij Indicates the j-th reviewer's evaluation of image I i The scoring results, w j is the reviewer weight, For image I i The final reference score of

[0029] Introducing the ELO sorting algorithm, for any image pair (I a ,I b ) Compare their respective final reference scores and construct a win-lose relationship S∈{0,1}, where 1 represents I a >I b , 0 otherwise; dynamically update the image rating based on the win-loss relationship:

[0030]

[0031] Among them, R a 、R b For image I a , I b Current rating of E a is the expected winning rate, K is the update coefficient; after dynamic updating of all image pairs, the ELO perception level results of all images are obtained;

[0032] The actual ELO perception level results of all images are used as target labels, and an artificial neural network is established and trained as a subjective perception score prediction model with a regression prediction structure;

[0033] Extract S i The subjective visual features in the image are input into the trained subjective perception score prediction model to obtain the subjective visual perception index of the image;

[0034] Calculate S i The facility function index in the image is obtained; the facility function index is standardized and spatially associated with the coordinates of the corresponding image to obtain the facility function index of the corresponding image;

[0035] By integrating the objective visual perception indicators, subjective visual perception indicators, and facility function indicators of all images, a normalization process and weighted calculation method are used to construct a comprehensive quality scoring model for street units, generating a street space quality evaluation result that integrates subjective and objective perceptions with physical facility characteristics.

[0036] Furthermore, the objective visual characteristic indicators specifically include green view ratio, sky openness, interface enclosure, and motorization degree;

[0037] Green view rate GVI indicates the proportion of green vegetation in the image:

[0038]

[0039] Among them, A green A represents the number of pixels in the street view image that are identified as the “green plant” semantic category. total Indicates the total number of pixels in the image;

[0040] Sky openness SO indicates the proportion of sky in the image:

[0041]

[0042] Among them, A sky is the number of pixels in the sky area of ​​the image;

[0043] The interface enclosure (IE) represents the degree of enclosure of the street space and is defined as the proportion of streetscape image structures and column elements within the street unit:

[0044]

[0045] Among them, A building and A pillar are the number of “building” and “column” type pixels in the image respectively;

[0046] The motorization degree (MD) reflects the degree to which a street space is dependent on motor vehicles and friendly to pedestrians. It is defined as the proportion of vehicle-car-car roads minus the proportion of pedestrian-car-car roads:

[0047]

[0048] Among them, A vehicle 、A road 、A sidewalk 、A pedestrian Represents the number of pixels in vehicle, motorway, sidewalk and pedestrian areas respectively.

[0049] Furthermore, the facility function indicators specifically include facility density, facility mix, walkability, and spatial vitality;

[0050] Among them, the facility density FD represents the scale and number of POIs within the street unit, reflecting the service convenience and commercial vitality of the street; when extracting, the buffer zone A is constructed based on the street centerline. i , count the number of points of interest in the area N POI,i :

[0051]

[0052] Where i represents the street unit number; the higher the FD value, the stronger the concentration of facilities in the street area;

[0053] The facility mix index (FMI) represents the types of POIs within a street unit, reflecting the service diversity and commercial vitality of the street. When extracting, the Shannon entropy index is used to calculate the functional mix index:

[0054]

[0055] Among them, p ik It represents the proportion of POIs of type k in the street i area, where K is the total number of facility types. The higher the FMI value, the more complex the spatial functions of the area are, and it can adapt to various daily life and activity needs.

[0056] Walking accessibility WI represents the straight-line distance from a street unit road network node to all other street nodes, reflecting the street's walking accessibility and traffic convenience. During extraction, the shortest path Euclidean distance d from street unit i to each point in the target facility set F is calculated based on the sDNA spatial network analysis tool. ij , and find its mean:

[0057]

[0058] Among them, |F| represents the number of facilities, d ij is the network path distance from street i to facility point j. The smaller the WI value, the better the pedestrian accessibility of the street area, which is more suitable for creating a convenient and pleasant street environment.

[0059] The SVD of spatial vitality is used to measure the intensity of crowd activity in the neighborhood of street unit i. During extraction, the resident population data is crawled and the heat level of people in the street space is used to characterize the vitality of the street space:

[0060]

[0061] Among them, SVD i represents the spatial vitality of the i-th street unit, h j represents the population heat value of the jth grid point around street i, taken from Baidu Huiyan's resident population heat map, reflecting the activity intensity of the location; d ij It represents the spatial distance between street unit i and heat point j, using Euclidean distance; n represents the total number of heat points participating in the calculation within the buffer range of street unit i.

[0062] A method for generating an intelligent image of street space quality based on deep learning, based on the intelligent evaluation method of street space quality based on deep learning, comprises the following specific steps:

[0063] Build and train the AIGC street creative design platform, which includes a street image generation network and a street space semantic guidance network;

[0064] The images in the fusion image database and the street space quality evaluation results are input into the AIGC street creative design platform to generate the corresponding streetscape transformation design images.

[0065] Preferably, the AIGC street creative design platform including the street image generation network and the street space semantic guidance network is constructed and trained, and the specific steps are as follows:

[0066] A street image generation network based on a diffusion model is developed; a street space semantic guidance network is developed to transform street space quality evaluation results into semantic control conditions in the image generation process;

[0067] The AIGC street creative design platform is constructed by integrating street image generation network and street space semantic guidance network;

[0068] Collect street images and their corresponding semantic description texts to construct a training dataset D:

[0069] D={(I i ,T i )}

[0070] Among them, I i Represents the current street image; T i Indicates the corresponding transformation target semantic description text;

[0071] D is input into the AIGC Street Creative Design Platform for image-text joint training. During the training, the street space quality is evaluated, and the quantitative indicators of the quality evaluation results are structured and encoded into a semantic control vector C. i , establish the input image I from the status quo i To generate image I' i The mapping function:

[0072] G(I i ,C i )→I i '

[0073] Among them, G is the generated model after training, C i Prompt conditions for street design, I i ' is the output street transformation intention map.

[0074] An intelligent image generation system for street space quality based on deep learning, including a motion sensing acquisition module, an image preprocessing module, a space quality evaluation module, and a design image generation module:

[0075] The mobile sensing acquisition module is used to collect street space images and record geographic location information;

[0076] The image preprocessing module is used to perform perspective correction and multi-view image segmentation on the collected street image data, and perform spatial matching with the geographic location information to obtain a fused image database;

[0077] The space quality evaluation module is used to perform semantic segmentation on images in the fusion image database based on a deep learning model, and calculate its objective and subjective visual perception indicators and facility function indicators to generate street space quality evaluation results;

[0078] The design image generation module is used to build and train the AIGC street creative design platform, which includes a street image generation network and a street space semantic guidance network; the images in the fusion image database and their street space quality evaluation results are input into the AIGC street creative design platform to generate corresponding streetscape renovation design images.

[0079] The beneficial effects of the present invention are as follows:

[0080] The present invention discloses an intelligent assessment method for street space quality based on deep learning. By collecting street space images and performing spatial matching with geographic location information, and calculating their objective and subjective visual perception indicators and facility function indicators, the street space quality evaluation results are generated. The present invention solves key problems existing in existing street space quality assessment and design practices, such as strong dependence on data acquisition, limited assessment dimensions, high subjectivity, and lack of an automated response mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is a specific flow chart of the intelligent evaluation method of street space quality based on deep learning of the present invention.

[0082] Figure 2 This is a flow chart of the street space image acquisition line design proposed in an embodiment of the present invention.

[0083] Figure 3 2 is a schematic diagram of the street panoramic image distortion correction proposed in an embodiment of the present invention.

[0084] Figure 4 3 is a schematic diagram of coordinate matching between GPS coordinates and image capture locations based on time codes, as proposed in an embodiment of the present invention.

[0085] Figure 5It is a technical roadmap for street space quality evaluation modeling proposed in the embodiment of the invention.

[0086] Figure 6 This is a graph of the street space semantic segmentation result based on the Segformer model proposed in an embodiment of the present invention.

[0087] Figure 7 4 is a diagram showing the calculation results of the objective indicators at the visual perception level proposed in an embodiment of the present invention.

[0088] Figure 8 4 is a diagram showing the calculation results of the subjective indicators at the visual perception level proposed in an embodiment of the present invention.

[0089] Figure 9 4 is a diagram showing the calculation results of the facility function level indicators proposed in the embodiment of the present invention.

[0090] Figure 10 This is a result diagram of the comprehensive evaluation calculation of visual perception level and facility function level indicators proposed in an embodiment of the present invention.

[0091] Figure 11 It is a flowchart of an intelligent image generation method for street space quality based on deep learning.

[0092] Figure 12 This is the result diagram of the street design image generation based on AIGC proposed in an embodiment of the present invention.

[0093] Figure 13 This is a schematic diagram of an intelligent image generation system for street space quality based on deep learning in the present invention. DETAILED DESCRIPTION

[0094] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0095] Example 1

[0096] like Figure 1 As shown in FIG, a method for intelligently evaluating street space quality based on deep learning includes the following specific steps:

[0097] Mobile collection of street space images and recording of geographic location information;

[0098] Perform perspective correction and multi-view image segmentation on the collected street image data, and perform spatial matching with geographic location information to obtain a fused image database;

[0099] Based on the deep learning model, semantic segmentation is performed on the images in the fusion image database, and their objective and subjective visual perception indicators and facility function indicators are calculated to generate street space quality evaluation results.

[0100] In a specific embodiment, street space images are collected on the move and geographic location information is recorded. Specifically, a mobile sensing device including a panoramic camera, a GPS recorder, a vehicle-mounted suction cup bracket, and a monitoring screen is built to continuously capture panoramic street images at a predetermined time interval of 5 seconds, and the geographic location and time information at the time of shooting are recorded to generate a street image dataset.

[0101] In a specific embodiment, Figure 2 As shown in the figure, when collecting street space images, a street network diagram is constructed and the optimal collection route P is used. opt The specific steps for collection are:

[0102] Starting from the road network data, we consider streets as edges and intersections as nodes to construct a street network graph;

[0103] Check whether the odd-degree vertices of the road network are equal to 0;

[0104] If so, Fleury's algorithm will be used to solve the Euler path to ensure that each street is covered at least once. Otherwise, Dijkstra's algorithm will be used to find the shortest distance between the odd-degree vertices. opt Connect two groups of two; add duplicate edges based on the optimal grouping, and return to check whether the odd-degree vertices of the road network are equal to 0;

[0105] Get the best route

[0106] In a specific embodiment, perspective correction and multi-view image segmentation are performed on the collected street space image, and spatial matching is performed with the geographic location information to obtain a fused image. The specific steps are as follows:

[0107] Construct a street network diagram and further construct spatial coordinates;

[0108] The images in the street image data are perspective corrected and projected into spatial coordinates to achieve mapping:

[0109]

[0110] Where, (u, v) is the image pixel coordinate; (x, y) is the image coordinate after plane projection; f is the equivalent focal length; f x , f y are the scaling factors of the image in the X and Y directions respectively; c x , c y is the principal point coordinate of the camera, that is, the center point of the image;

[0111] like Figure 3 As shown, after projection transformation and image correction, the image is split into six perspective image sets: front, back, left, right, top, and bottom, retaining the front / back / left / right four perspectives V′={If ,I b ,I l ,I r}, obtain the distortion-free street space image, and then, match the GPS trajectory according to the shooting timestamp to achieve accurate spatial matching between the image and the geographic coordinates; after the image is spatially annotated, it is uploaded to the GIS platform to form a street image database with spatial attributes, such as Figure 4 shown.

[0112] The shooting time code of the image is matched with its GPS trajectory to generate a dataset that binds the image set to the geographic coordinates, forming a fused image database with spatial attributes.

[0113] like Figure 5 As shown, in a specific embodiment, semantic segmentation is performed on images in a fusion image database based on a deep learning model, and objective and subjective visual perception indicators and facility function indicators are calculated to generate street space quality evaluation results. The specific steps are as follows:

[0114] like Figure 6 As shown, the deep semantic segmentation model SegFormer based on the Transformer structure is used to collect the street view image I i Perform semantic segmentation to identify the spatial element categories in the image and obtain pixel-level semantic label image S i , calculate S i The objective visual feature indicators in Figure 7 As shown;

[0115] Based on the crowdsourcing mechanism, experts and scholars with backgrounds in urban design, environmental psychology, and ordinary citizens are invited to conduct a survey on a representative set of street view image samples. Conduct subjective perception scoring. In this example, experts account for 60% of the weight, and citizen volunteers account for 40%. The scoring dimensions include experience-oriented indicators such as "imagery", "humanization", "cleanliness", and "transparency". Each image is independently scored by at least m participants, and a weighted average method is used to generate a subjective reference score:

[0116]

[0117] Among them, s ij Indicates the j-th reviewer's evaluation of image I i The scoring results, w j is the reviewer weight, For image I i The final reference score of

[0118] Introducing the ELO sorting algorithm, for any image pair (I a ,I b) Compare their respective final reference scores and construct a win-lose relationship S∈{0,1}, where 1 represents I a >I b , 0 otherwise; dynamically update the image rating based on the win-loss relationship:

[0119]

[0120] Among them, R a 、R b For image I a , I b Current rating of E a is the expected winning rate, K is the update coefficient; after dynamic updating of all image pairs, the ELO perception level results of all images are obtained;

[0121] The actual ELO perception level results of all images are used as target labels, and a subjective perception score prediction model with an artificial neural network as a regression prediction structure is established and trained; the model training goal is to minimize the mean square error loss between the predicted value output by the model and the actual ELO perception level result;

[0122] Extract S i The subjective visual features in the image are input into the trained subjective perception score prediction model to obtain the subjective visual perception index of the image, such as Figure 8 As shown;

[0123] Calculate S i The facility function index in the image is standardized and then spatially associated with the coordinates of the corresponding image to obtain the facility function index of the corresponding image, such as Figure 9 As shown;

[0124] By integrating the objective visual perception indicators, subjective visual perception indicators, and facility function indicators of all images, a normalization process and weighted calculation method are used to construct a comprehensive quality scoring model for street units, generating a street space quality evaluation result that integrates subjective and objective perceptions with physical facility characteristics.

[0125] In this embodiment, after obtaining the spatial quality evaluation results, heat map expression and classification are performed on the GIS platform, such as Figure 10 As shown in the figure, the analysis results can be used to identify differences in spatial quality, explore areas with weak quality and their causes, and provide quantitative input conditions for the subsequent generation of street design images based on AIGC, thus achieving an efficient closed loop from assessment and diagnosis to design intervention.

[0126] In a specific embodiment, the objective visual characteristic indicators specifically include green viewing rate, sky openness, interface enclosure, and motorization degree;

[0127] Green view rate GVI indicates the proportion of green vegetation in the image:

[0128]

[0129] Among them, A green A represents the number of pixels in the street view image that are identified as the “green plant” semantic category. total Indicates the total number of pixels in the image;

[0130] Sky openness SO indicates the proportion of sky in the image:

[0131]

[0132] Among them, A sky is the number of pixels in the sky area of ​​the image;

[0133] The interface enclosure (IE) represents the degree of enclosure of the street space and is defined as the proportion of streetscape image structures and column elements within the street unit:

[0134]

[0135] Among them, A building and A pillar are the number of “building” and “column” type pixels in the image respectively;

[0136] The motorization degree (MD) reflects the degree to which a street space is dependent on motor vehicles and friendly to pedestrians. It is defined as the proportion of vehicle-car-car roads minus the proportion of pedestrian-car-car roads:

[0137]

[0138] Among them, A vehicle 、A road 、A sidewalk 、A pedestrian Represents the number of pixels in vehicle, motorway, sidewalk and pedestrian areas respectively.

[0139] In a specific embodiment, the facility function indicators specifically include facility density, facility mix, walkability, and spatial vitality;

[0140] Among them, the facility density FD represents the scale and number of POIs within the street unit, reflecting the service convenience and commercial vitality of the street; when extracting, the buffer zone A is constructed based on the street centerline. i , count the number of points of interest in the area N POI,i :

[0141]

[0142] Where i represents the street unit number; the higher the FD value, the stronger the concentration of facilities in the street area;

[0143] The facility mix index (FMI) represents the types of POIs within a street unit, reflecting the service diversity and commercial vitality of the street. When extracting, the Shannon entropy index is used to calculate the functional mix index:

[0144]

[0145] Among them, p ik It represents the proportion of POIs of type k in the street i area, where K is the total number of facility types. The higher the FMI value, the more complex the spatial functions of the area are, and it can adapt to various daily life and activity needs.

[0146] Walking accessibility WI represents the straight-line distance from a street unit road network node to all other street nodes, reflecting the street's walking accessibility and traffic convenience. During extraction, the shortest path Euclidean distance d from street unit i to each point in the target facility set F is calculated based on the sDNA spatial network analysis tool. ij , and find its mean:

[0147]

[0148] Among them, |F| represents the number of facilities, d ij is the network path distance from street i to facility point j. The smaller the WI value, the better the pedestrian accessibility of the street area, which is more suitable for creating a convenient and pleasant street environment.

[0149] The SVD of spatial vitality is used to measure the intensity of crowd activity in the neighborhood of street unit i. During extraction, the resident population data is crawled and the heat level of people in the street space is used to characterize the vitality of the street space:

[0150]

[0151] Among them, SVD i represents the spatial vitality of the i-th street unit, h j represents the population heat value of the jth grid point around street i, taken from Baidu Huiyan's resident population heat map, reflecting the activity intensity of the location; d ij It represents the spatial distance between street unit i and heat point j, using Euclidean distance; n represents the total number of heat points participating in the calculation within the buffer range of street unit i.

[0152] In this embodiment, all indicators and their weights in the street space quality evaluation index system are shown in Table 1:

[0153] Table 1 Street space quality evaluation index system

[0154]

[0155]

[0156] Example 2

[0157] like Figure 11 As shown, a method for generating an intelligent image of street space quality based on deep learning, based on the intelligent evaluation method of street space quality based on deep learning, includes the following specific steps:

[0158] Build and train the AIGC street creative design platform, which includes a street image generation network and a street space semantic guidance network;

[0159] The images in the fusion image database and the street space quality evaluation results are input into the AIGC street creative design platform to generate the corresponding streetscape transformation design images.

[0160] In a specific embodiment, the AIGC street creative design platform including a street image generation network and a street space semantic guidance network is constructed and trained. The specific steps are as follows:

[0161] A street image generation network based on a diffusion model is developed; a street space semantic guidance network is developed to transform street space quality evaluation results into semantic control conditions in the image generation process;

[0162] The AIGC street creative design platform is constructed by integrating street image generation network and street space semantic guidance network;

[0163] Collect street images and their corresponding semantic description texts to construct a training dataset D:

[0164] D={(I i ,T i )}

[0165] Among them, I i Represents the current street image; T i Represents the semantic description text of the corresponding renovation target, such as "repair the facades of buildings along the street", "optimize the parking order of non-motor vehicles", "improve the road paving material", "increase the width of street green belts and sidewalks";

[0166] D is input into the AIGC Street Creative Design Platform for image-text joint training. During the training, the street space quality is evaluated, and the quantitative indicators of the quality evaluation results are structured and encoded into a semantic control vector C. i , establish the input image I from the status quo i To generate image I i 'Mapping function:

[0167] G(I i ,C i )→I i '

[0168] Among them, G is the generated model after training, C i Prompt conditions for street design, I i ' is the output street transformation intention map.

[0169] Finally, if Figure 12 The generated street renovation images shown can be widely used in various planning and design scenarios, including urban block micro-renewal, urban health assessments, county-level public space quality improvement, and support for the "Hundreds, Millions, and Tens of Millions" project. This system can automatically generate streetscape images with a high degree of realism and design logic, while maintaining the original spatial context while integrating design elements such as traffic organization optimization, building facade enhancements, and street greening improvements. These images can serve as important reference material for design proposals, assisting planners in making decisions on tasks such as street micro-renewal, community renovation, and urban health assessments.

[0170] Example 3

[0171] like Figure 13 As shown in the figure, an intelligent image generation system for street space quality based on deep learning includes a motion perception acquisition module, an image preprocessing module, a space quality evaluation module, and a design image generation module:

[0172] The mobile sensing acquisition module is used to integrate mobile acquisition equipment (including panoramic cameras, GPS recorders, vehicle-mounted suction cup brackets and monitoring terminal sensing devices), combined with path planning based on the China Postal Route Problem optimization algorithm, to achieve efficient and autonomous acquisition of street space images.

[0173] The image preprocessing module performs distortion correction and multi-view splitting on panoramic images, converting them into a flat format and retaining the four perspectives of front, back, left, and right. Furthermore, coordinate matching is performed by combining image capture timestamps with GPS trajectory information to generate a street image database with geographic attributes.

[0174] The spatial quality assessment module performs distortion correction and multi-view splitting on panoramic images, converting them into a flat format while retaining the four perspectives of front, back, left, and right. It also combines image capture timestamps with GPS trajectory information for coordinate matching, generating a street image database with geographic attributes.

[0175] The design image generation module is used to automatically generate street renovation design images based on street space quality evaluation results. This module, centered on the AIGC model and combined with the semantic control module, converts evaluation results into semantic conditions for image generation. This enables the generation of problem-oriented street space design intent maps, providing an efficient and intuitive visual reference for neighborhood renewal and design decision-making.

[0176] Obviously, the above embodiments of the present invention are merely examples for the purpose of illustrating the present invention, and are not intended to limit the embodiments of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A deep learning-based intelligent assessment method for street space quality, characterized by: The specific steps include: Mobile collection of street space images and recording of geographic location information; Perform perspective correction and multi-view image segmentation on the collected street image data, and perform spatial matching with geographic location information to obtain a fused image database; Based on the deep learning model, semantic segmentation is performed on the images in the fusion image database, and their objective and subjective visual perception indicators and facility function indicators are calculated to generate street space quality evaluation results.

2. The deep learning-based intelligent street space quality assessment method according to claim 1, characterized in that: Mobile street space imagery is collected and geographic location information is recorded. Specifically, a mobile sensing device consisting of a panoramic camera, a GPS recorder, a vehicle-mounted suction cup bracket, and a monitoring screen is built to continuously capture panoramic street images at predetermined time intervals. The geographic location and time information at the time of capture are recorded to generate a street image dataset.

3. The deep learning-based intelligent street space quality assessment method according to claim 2, characterized in that: When collecting street space images, a street network diagram is constructed and the optimal collection route P is used. opt To collect, P opt satisfy: Among them, d(p i ,p i+1 ) is the distance between adjacent street points.

4. The deep learning-based intelligent street space quality assessment method according to claim 3, characterized in that: The collected street space images are subjected to perspective correction and multi-view image segmentation, and spatial matching is performed with the geographic location information to obtain a fused image. The specific steps are as follows: Based on the street network diagram, spatial coordinates are further constructed; The images in the street image data are perspective corrected and projected into spatial coordinates to achieve mapping: Where, (u, v) is the image pixel coordinate; (x, y) is the image coordinate after plane projection; f is the equivalent focal length; f x , f y are the scaling factors of the image in the X and Y directions respectively; c x , c y is the principal point coordinate of the camera, that is, the center point of the image; The mapped image is divided into a set of images of four perspectives: front / back / left / right V′={I f ,I b ,I l ,I r }; The shooting time code of the image is matched with its GPS trajectory to generate a dataset that binds the image set to the geographic coordinates, forming a fused image database with spatial attributes.

5. The intelligent street space quality assessment method based on deep learning according to claim 1, characterized in that: Based on the deep learning model, semantic segmentation is performed on the images in the fusion image database, and their objective and subjective visual perception indicators and facility function indicators are calculated to generate street space quality evaluation results. The specific steps are as follows: The deep semantic segmentation model SegFormer based on the Transformer structure is used to perform semantic segmentation on the collected street view image Ii, identify the spatial element categories in the image, obtain the pixel-level semantic label image Si, and calculate the objective visual feature indicators in Si; Based on the crowdsourcing mechanism, a sample set of street view images is collected. Each image is scored independently by at least m participants, and a weighted average method is used to generate a subjective reference score: Among them, s ij represents the jth reviewer's evaluation of image I i The scoring results, w j is the reviewer weight, For image I i The final reference score of Introducing the ELO sorting algorithm, for any image pair (I a ,I b ) Compare their respective final reference scores and construct a win-lose relationship S∈{0,1}, where 1 represents I a >I b , 0 otherwise; dynamically update the image rating based on the win-loss relationship: Among them, R a 、R b For image I a , I b Current rating of E a is the expected winning rate, K is the update coefficient; after dynamic updating of all image pairs, the ELO perception level results of all images are obtained; The actual ELO perception level results of all images are used as target labels, and an artificial neural network is established and trained as a subjective perception score prediction model with a regression prediction structure; Extract the subjective visual features in Si and input them into the trained subjective perception score prediction model to obtain the subjective visual perception index of the image; Calculate the facility function index in Si; standardize the facility function index and spatially associate it with the coordinates of the corresponding image to obtain the facility function index of the corresponding image; By integrating the objective visual perception indicators, subjective visual perception indicators, and facility function indicators of all images, a normalization process and weighted calculation method are used to construct a comprehensive quality scoring model for street units, generating a street space quality evaluation result that integrates subjective and objective perceptions with physical facility characteristics.

6. The deep learning-based intelligent street space quality assessment method according to claim 5, characterized in that: The objective visual characteristic indicators specifically include green view ratio, sky openness, interface enclosure, and motorization degree; Green view rate GVI indicates the proportion of green vegetation in the image: Among them, A green A represents the number of pixels in the street view image that are identified as the "green plant" semantic category. total Indicates the total number of pixels in the image; Sky openness SO indicates the proportion of sky in the image: Among them, A sky is the number of pixels in the sky area of ​​the image; The interface enclosure (IE) represents the degree of enclosure of the street space and is defined as the proportion of streetscape image structures and column elements within the street unit: Among them, A building and A pillar are the number of "building" and "column" pixels in the image respectively; The motorization degree (MD) reflects the degree to which a street space is dependent on motor vehicles and friendly to pedestrians. It is defined as the proportion of vehicle-car-car roads minus the proportion of pedestrian-car-car roads: Among them, A vehicle 、A road 、A sidewalk 、A pedestrian Represents the number of pixels in vehicle, motorway, sidewalk and pedestrian areas respectively.

7. The intelligent street space quality assessment method based on deep learning according to claim 5, characterized in that: The facility function indicators specifically include facility density, facility mix, walkability, and spatial vitality; Among them, the facility density FD represents the scale and number of POIs within the street unit, reflecting the service convenience and commercial vitality of the street; when extracting, the buffer zone A is constructed based on the street centerline. i , count the number of points of interest in the area N POI,i : Where i represents the street unit number; the higher the FD value, the stronger the concentration of facilities in the street area; The facility mix index (FMI) represents the types of POIs within a street unit, reflecting the service diversity and commercial vitality of the street. When extracting, the Shannon entropy index is used to calculate the functional mix index: Among them, p ik It represents the proportion of POIs of type k in the street i area, where K is the total number of facility types. The higher the FMI value, the more complex the spatial functions of the area are, and it can adapt to various daily life and activity needs. Walking accessibility WI represents the straight-line distance from a street unit road network node to all other street nodes, reflecting the street's walking accessibility and traffic convenience. During extraction, the shortest path Euclidean distance d from street unit i to each point in the target facility set F is calculated based on the sDNA spatial network analysis tool. ij , and find its mean: Among them, |F| represents the number of facilities, d ij is the network path distance from street i to facility point j. The smaller the WI value, the better the pedestrian accessibility of the street area, which is more suitable for creating a convenient and pleasant street environment. The SVD of spatial vitality is used to measure the intensity of crowd activity in the neighborhood of street unit i. During extraction, the resident population data is crawled and the heat level of people in the street space is used to characterize the vitality of the street space: Among them, SVD i represents the spatial vitality of the i-th street unit, h j represents the population heat value of the jth grid point around street i, taken from Baidu Huiyan's resident population heat map, reflecting the activity intensity of the location; d ij It represents the spatial distance between street unit i and heat point j, using Euclidean distance; n represents the total number of heat points participating in the calculation within the buffer range of street unit i.

8. A method for generating intelligent images of street space quality based on deep learning, characterized by: The intelligent assessment method for street space quality based on deep learning according to any one of claims 1 to 7 comprises the following specific steps: Build and train the AIGC street creative design platform, which includes a street image generation network and a street space semantic guidance network; The images in the fusion image database and the street space quality evaluation results are input into the AIGC street creative design platform to generate the corresponding streetscape transformation design images.

9. The method for generating intelligent images of street space quality based on deep learning according to claim 8, characterized in that: Build and train the AIGC street creative design platform, which includes a street image generation network and a street space semantic guidance network. The specific steps are as follows: A street image generation network based on a diffusion model is developed; a street space semantic guidance network is developed to transform street space quality evaluation results into semantic control conditions in the image generation process; The AIGC street creative design platform is constructed by integrating street image generation network and street space semantic guidance network; Collect street images and their corresponding semantic description texts to construct a training dataset D: D={(I i ,T i )} Among them, I i Represents the current street image; T i Indicates the corresponding transformation target semantic description text; D is input into the AIGC Street Creative Design Platform for image-text joint training. During the training, the street space quality is evaluated, and the quantitative indicators of the quality evaluation results are structured and encoded into a semantic control vector C. i , establish the input image I from the status quo i To generate image I' i The mapping function: G(I i ,C i )→I’ i Among them, G is the generated model after training, C i Prompt conditions for street design, I' i The output is the street reconstruction intention map.

10. A deep learning-based intelligent image generation system for street space quality, characterized by: Including motion perception acquisition module, image preprocessing module, and space quality evaluation module: The mobile sensing acquisition module is used to collect street space images and record geographic location information; The image preprocessing module is used to perform perspective correction and multi-view image segmentation on the collected street image data, and perform spatial matching with the geographic location information to obtain a fused image database; The space quality evaluation module is used to perform semantic segmentation on images in the fusion image database based on a deep learning model, and calculate their objective and subjective visual perception indicators and facility function indicators to generate street space quality evaluation results.

Images