Image processing method and device, equipment, storage medium and program product

Abstract

The invention discloses an image processing method and device, equipment, a storage medium and a program product, and relates to the technical field of big data. The method comprises the following steps: acquiring a first crowd thermodynamic distribution image of a first area in a first time period; determining first crowd thermodynamic distribution information of the first region in the first time period based on the first crowd thermodynamic distribution image through a multi-scale convolutional neural network by taking each geographic grid in the first region as a center; determining second crowd thermodynamic distribution information of the first region in a second time period according to the first crowd thermodynamic distribution information through a selective state space model; and constructing a second crowd thermodynamic distribution image of the first region in the second time period based on the second crowd thermodynamic distribution information. In this way, the accuracy of constructing the crowd thermodynamic distribution image in the future time period can be improved by considering that the crowd thermodynamic distribution has significant spatial-temporal heterogeneity and dynamic evolution characteristics.

Description

Technical Field

[0001] This application belongs to the field of big data technology, and in particular relates to an image processing method, apparatus, device, storage medium and program product. Background Technology

[0002] With increasingly frequent population flows and highly complex urban operations, population heat map data is becoming a crucial data asset for smart city governance, service operation decisions, and even public safety management. For service providers, population heat map data can effectively support the optimization of the transaction service acceptance environment in various offline urban scenarios, the adjustment of merchant outlet deployment strategies, the precise layout of service activities, and the improvement of resource allocation efficiency. This enables dynamic matching between payment infrastructure and population hotspots, thereby improving overall operational efficiency.

[0003] In related technologies, population thermal distribution maps of a certain area can be determined through single-point static observation. However, population thermal distribution itself has significant spatiotemporal heterogeneity and dynamic evolution characteristics. The aforementioned methods have limited spatial coverage and cannot simultaneously capture local fine-grained patterns and global macroscopic patterns, such as changes in population flow in a certain area and cross-regional population migration, which affects the accuracy of determining population thermal distribution images. Summary of the Invention

[0004] This application provides an image processing method, apparatus, device, storage medium, and program product that can improve the accuracy of determining the thermal distribution of crowds in images.

[0005] In a first aspect, embodiments of this application provide an image processing method, the method comprising: Obtain the thermal distribution image of the first population in the first region within the first time period. The first region includes N geographic grids, where N is an integer greater than 1. Centered on each geographic grid, a multi-scale convolutional neural network is used to determine the thermal distribution information of the first population in the first region within the first time period based on the thermal distribution image of the first population. Using a selective state-space model, the thermal distribution information of the second population in the first region during the second time period is determined based on the thermal distribution information of the first population. Based on the thermal distribution information of the second population, a thermal distribution image of the second population in the first region during the second time period is constructed.

[0006] Secondly, embodiments of this application provide an image processing apparatus, which may include: The acquisition module is used to acquire the thermal distribution image of the first population in the first region within the first time period. The first region includes N geographic grids, where N is an integer greater than 1. The module determines the thermal distribution information of the first population in the first region within the first time period, centered on each geographic raster and using a multi-scale convolutional neural network based on the thermal distribution image of the first population. The determination module is also used to determine the thermal distribution information of the second population in the first region during the second time period based on the thermal distribution information of the first population using a selective state-space model. The module is used to construct a thermal distribution image of the second population in the first region within a second time period based on the thermal distribution information of the second population.

[0007] Thirdly, embodiments of this application provide a computer device, which includes: a processor and a memory storing computer program instructions; When the processor executes computer program instructions, it implements the image processing method as described in the first aspect.

[0008] Fourthly, embodiments of this application provide a computer storage medium storing computer program instructions, which, when executed by a processor, implement the image processing method as described in the first aspect.

[0009] Fifthly, embodiments of this application provide a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled, and the processor is used to run programs or instructions to implement the image processing method as shown in the first aspect.

[0010] In a sixth aspect, embodiments of this application provide a computer program product stored in a storage medium, which is executed by at least one processor to implement the image processing method as described in the first aspect.

[0011] The image processing method, apparatus, device, storage medium, and program product of this application embodiment can acquire a thermal distribution image of a first population in a first region within a first time period. The first region includes N geographic grids, where N is an integer greater than 1. Centered on each geographic grid, a multi-scale convolutional neural network is used to determine the thermal distribution information of the first population in the first region within the first time period based on the thermal distribution image of the first population. A selective state-space model is used to determine the thermal distribution information of a second population in the first region within a second time period based on the thermal distribution information of the first population. Based on the thermal distribution information of the second population, a thermal distribution image of the second population in the first region within the second time period is constructed. In this way, by dividing the geographic grid and acquiring thermal images, a spatial discretization representation of population distribution is achieved, providing a structured data foundation for subsequent analysis. Based on a multi-scale convolutional neural network, population density features at different spatial scales are accurately extracted to generate high-resolution thermal distribution information, preserving local clustering details and global distribution trends. A selective state-space model is used to capture the spatiotemporal dynamic evolution of population distribution, and adaptive structural upgrades are achieved by combining the selective state-space model with the thermal model. This reduces computational complexity while ensuring prediction accuracy. Therefore, by considering the significant spatiotemporal heterogeneity and dynamic evolution characteristics of population thermal distribution, constructing population thermal distribution images for future time periods can improve the accuracy of determining population thermal distribution images. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of an image processing method according to an embodiment of the image processing method provided in this application; Figure 2 This is a schematic diagram of an image processing flow according to an embodiment of the image processing method provided in this application; Figure 3 This is a schematic diagram of a thermal distribution image of a crowd based on the image processing method provided in this application; Figure 4 This is a schematic diagram of a thermal distribution image of a crowd based on the image processing method provided in this application; Figure 5 This is a schematic diagram of a thermal distribution image of a crowd based on the image processing method provided in this application; Figure 6 This is a schematic diagram of the structure of an image processing apparatus provided in one embodiment of this application; Figure 7 This is a schematic diagram of the structure of a data query device provided in one embodiment of this application. Detailed Implementation

[0014] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0015] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0016] The acquisition, storage, use, and processing of data (including but not limited to features and information mentioned in this document) in the technical solution of this application all comply with the relevant provisions of national laws and regulations.

[0017] Population thermal distribution data exhibits significant spatiotemporal heterogeneity and dynamic evolution characteristics, making single-point static observations insufficient to meet the needs of practical services for forward-looking and continuous situational awareness. Therefore, accurately predicting future population thermal distribution has become a research topic in data-driven services. In recent years, with the availability of big data and the development of high-performance computing capabilities, a series of deep learning-based methods have emerged to address these issues, including convolutional neural networks, recurrent neural networks, graph neural networks, and their spatiotemporal fusion variants. These methods have achieved initial success in extracting spatial proximity and modeling time-series dependencies, but still face many challenges.

[0018] Traditional methods, when modeling spatial dependencies, often rely on static, local adjacency structures, making it difficult to fully capture the dynamic, multi-scale population distribution patterns in urban spaces. Secondly, at the temporal modeling level, commonly used sequence modeling structures such as recurrent neural networks, while possessing some long-range dependency modeling capabilities, often ignore semantic asymmetry between different time points, easily leading to the smoothing of high-frequency changes and the over-memorization of low-frequency trends. Furthermore, existing works generally lack the ability to selectively model spatiotemporal coupling; that is, they cannot dynamically adjust the focus on historical population distribution information based on the context of the current prediction task, resulting in a forgetting problem in the model—the loss of early distribution patterns and the memorization of only recent distribution information.

[0019] Based on this, a method for predicting visitor traffic in a target area during a target time period is provided in related technologies. Specifically, the popularity of a target area can be determined based on historical visitor traffic, average dwell time, area, behavioral data, and feedback data. Generally, higher historical visitor traffic and longer average dwell time indicate that the services or experiences offered by the target area are more attractive, and tourists are more willing to stay and engage in activities there. To this end, this method collects the actual area of ​​each target area through a geographic information system; it collects behavioral data on each person's shopping, dining, entertainment, and rest activities within the target area through data sources such as mobile payment records to understand people's dwell time and consumption habits; in addition, it collects feedback data on the target area through multiple channels such as tourist satisfaction surveys, social media comments, and online reviews. Then, predictive analysis is performed based on historical visitor traffic and popularity to obtain the predicted visitor traffic in the target area during the target time period. Therefore, time series models are selected, such as the Autoregressive Integrated Moving Average (ARIMA) model, the exponential smoothing model, machine learning models, or ensemble learning methods combining multiple predictive analysis models. Using historical pedestrian traffic as input, reference pedestrian traffic for the target area within the target time period is obtained. Then, based on popularity, the reference pedestrian traffic is adjusted to obtain the predicted pedestrian traffic for the target area within the target time period.

[0020] It is evident that this method predicts based on historical pedestrian traffic and the popularity of the target area, relying on a manually constructed multi-factor weighted formula and a traditional time series forecasting model. This method has the following shortcomings in spatiotemporal modeling and dynamic adaptability: First, insufficient spatial dependency modeling. Existing technologies adopt a single-point analysis approach, treating the target area as independent units, lacking the ability to model the relationships between different areas in urban space and failing to capture multi-scale spatial dependencies. Second, limited handling of temporal dependencies. Temporal modeling relies on traditional time series methods, which have limited ability to capture long-range dependencies and struggle to distinguish between high-frequency changes and low-frequency trends, easily leading to distortions such as smoothing high-frequency signals and over-emphasizing low-frequency trends. Third, a lack of spatiotemporal coupling selection mechanisms. Existing methods cannot dynamically select the weight of historical information based on the context of the prediction task; key information from earlier periods is easily forgotten, retaining only recent patterns, affecting the performance of the prediction results.

[0021] A new spatial distribution prediction model for pedestrian traffic based on a multilayer perceptron, attention mechanism, and deep neural network is proposed in related technologies. This model comprises three main modules: a feature pattern extraction module, a spatial feature distance weight extraction module, and a feature fusion regression module. First, the feature pattern extraction module learns the time-series features of pedestrian traffic data and obtains the corresponding pattern matrix. Specifically, a multilayer perceptron is used with grid-based pedestrian traffic data as input to extract time-domain pattern features. Next, the spatial feature distance weight extraction module mines the weights of geographical environmental factors and obtains the corresponding spatial feature distance weights. An attention mechanism and deep neural network are used to capture spatial pattern features. Finally, the feature fusion regression module fuses the pattern matrix and spatial feature distance weights to calculate the final predicted spatial distribution data of pedestrian traffic.

[0022] However, this method has the following limitations in spatiotemporal modeling. First, it has limitations in spatial modeling, relying too heavily on static spatial distances and feature dimension distances calculated from geographic coordinates, and then generating spatial feature weights through attention mechanisms and deep neural networks. While this approach can capture some spatial correlations, it is a static, distance-based adjacency modeling method, lacking the ability to directly model dynamic, multi-scale spatial dependency patterns in urban population flows. Furthermore, this method does not explicitly introduce a multi-scale perception mechanism, making it difficult to simultaneously capture local fine-grained patterns such as street-level pedestrian flow changes and global macro-patterns such as cross-regional population migration trends. In actual population heat map prediction, this multi-scale spatial perception capability is particularly crucial for identifying the formation and dissipation of hotspots. Second, it has limitations in temporal modeling. The temporal features of this technology are mainly extracted through a multilayer perceptron, but it is essentially a static mapping and cannot retain and recursively utilize historical state information like state-space models or other sequential structures. Therefore, this method lacks the ability to characterize long-range dependencies and complex temporal patterns, especially when facing periodic changes, sudden events, or trend reversals, leading to distorted prediction results. Third, the spatiotemporal information fusion lacks dynamic selectivity. In existing technologies, temporal and spatial features are simply fused within a feature fusion regression module, with the fusion weights determined by the spatial feature weights and the pattern matrix. This static fusion mechanism cannot dynamically adjust the attention ratio of historical time steps based on the context of the current prediction task. Consequently, in long-term predictions or scenarios with drastic pattern changes, early key information is forgotten, and only recent patterns are retained, thus affecting the stability and foresight of the prediction.

[0023] To address this pain point and improve the accuracy of determining population thermal distribution images, this application provides an image processing method, apparatus, device, storage medium, and program product that can predict population thermal distribution based on the Spatio-Temporal Selective State Space Model (ST-SSSM). This method leverages the advantages of state space models in long-range dependency modeling, combining them with multi-scale convolutional neural networks. By explicitly modeling historical states at different time steps, it constructs a dynamic information flow path, thereby achieving a deep understanding and accurate prediction of population thermal evolution. Specifically, ST-SSSM uses spatial grids as modeling units, utilizes multi-scale two-dimensional convolutional kernels to perceive spatial structure, and introduces a state update mechanism in the time dimension to dynamically and selectively fuse state information from past moments to adapt to the differences in population activity patterns across different regions and time periods. Furthermore, the model structure is simple, facilitating parallel computation for efficient processing of large-scale thermal data and enabling rapid deployment in real-world urban data.

[0024] The following will be combined with the appendix Figures 1 to 7This application describes in detail the image processing methods, apparatus, computer equipment, storage media, and program products according to embodiments of the present application. It should be noted that these embodiments are not intended to limit the scope of the disclosure of this application.

[0025] First, the terminology used in the embodiments of this application will be explained. The terminology used in the implementation section of this application is only for explaining specific embodiments of this application and is not intended to limit this application.

[0026] The first region refers to a continuous spatial area on the Earth's surface divided according to natural or human characteristics. Its definition is based on natural elements such as topography, climate, and vegetation, or human elements such as language, culture, and economy. In the embodiments of this application, it can refer to areas with human activity, such as core areas, suburbs, and outer suburbs, divided by human elements such as topography, culture, and economy.

[0027] A heatmap, also known as a crowd thermal distribution image, is a data visualization tool that uses color changes to visually represent the density and distribution of a crowd.

[0028] A geographic raster is a data structure that divides a geographic space, such as a region, into a regular grid array according to latitude and longitude. Each geographic raster or cell is located by row and column coordinates, which can represent the latitude and longitude of the geographic raster.

[0029] Multiscale convolutional neural networks are deep learning models that process features at different scales in parallel or hierarchically, aiming to improve the performance of tasks sensitive to scale changes.

[0030] Selective state-space model is a dynamic time-domain model that can dynamically adjust its internal parameters according to the content of the input sequence, thereby selectively propagating or forgetting information and achieving efficient sequence modeling.

[0031] The first time period in this embodiment includes M time windows, and the thermal distribution information of the first population includes the thermal distribution information of the first population in the first region within each time window, where M is an integer greater than or equal to 1.

[0032] In this embodiment, the second time period occurs later than the first time period, and can be a future time.

[0033] The receptive field refers to the size of the region in the original input image that a pixel in the output feature map maps to. In this embodiment, it refers to the size of the region in the first population heat map represented by the first population heat map output by the MCNN. In multi-scale convolutional neural networks, the receptive field of each neuron determines the range of information it can perceive and respond to.

[0034] A convolutional kernel is a filter in a multi-scale convolutional neural network used to extract local features from the input data. It extracts features from the input data through element-wise multiplication and addition (cross-correlation) operations.

[0035] Scale refers to the resolution of an image or feature map, or the range of features extracted at different levels.

[0036] The kernel size refers to the width and height of the convolution kernel, such as 3×3 or 5×5, and determines the receptive field size and computational cost. Specifically, the kernel size directly determines the size of its corresponding receptive field. Larger kernels, such as 5×5 or 7×7, have a larger receptive field, covering a larger area of ​​the input image, thus capturing more global and macroscopic features, the overall shape and contour of objects. Smaller kernels, such as 1×1 or 3×3, have a smaller receptive field, focusing on capturing local, detailed features, edges, and textures.

[0037] Based on this, combined Figure 1 The image processing method provided in the embodiments of this application will be described in detail.

[0038] Figure 1 This is a flowchart of an image processing method provided in an embodiment of this application.

[0039] like Figure 1 As shown, this image processing method can be applied to computer devices and may specifically include the following steps: Step 110: Obtain the thermal distribution image of the first population in the first region within the first time period. The first region includes N geographic grids, where N is an integer greater than 1. Step 120: Using each geographic grid as the center, determine the thermal distribution information of the first population in the first region within the first time period based on the thermal distribution image of the first population using a multi-scale convolutional neural network. Step 130: Determine the thermal distribution information of the second population in the first region within the second time period based on the thermal distribution information of the first population using a selective state-space model. Step 140: Construct the thermal distribution image of the second population in the first region within the second time period based on the thermal distribution information of the second population.

[0040] In the embodiments of this application, such as Figure 2 As shown, in order to effectively model the complex spatial and temporal dependencies of population thermal data, a population thermal distribution prediction algorithm based on a spatiotemporal selective state space model is proposed. In the spatial dimension, a multi-scale convolutional neural network (MCNN) is constructed to extract the spatial feature representation of each region.

[0041] Specifically, the method centers on each geographic raster in the first region and applies convolutional kernels with different receptive fields. Through a multi-scale convolutional neural network 201, it extracts the population distribution features of the first region within the first time period at multiple scales, thereby achieving joint modeling of local and global patterns and enhancing representation capabilities. In the temporal dimension, the problem of predicting population heat distribution is analogous to sequence modeling tasks in natural language processing. Since the first time period can include M time windows, such as time window 1, time window 2, ..., time window M, the representation of the first population heat distribution information in each time window of the first region can be regarded as token embeddings. Therefore, the representation sequence of the same multi-scale convolutional neural network across multiple consecutive time windows can constitute a time-series text.

[0042] Based on this, predicting the thermal distribution image of the second population in the second time period is equivalent to performing next-token prediction. To achieve efficient modeling of the temporal context, a Selective State Space Model (Selective SSM) 202 is introduced. This model combines the long-range dependency modeling capability of state space modeling with the local information preservation advantage of the selection mechanism, and can fully capture the deep-level patterns of thermal value evolution over time. Thus, the thermal distribution information of the second population in the first region within the second time period can be determined by the Selective State Space Model 202, thereby improving the prediction accuracy. Finally, based on the thermal distribution information of the second population, a thermal distribution image 203 of the second population in the first region within the second time period can be constructed. Spatially, multi-scale convolutional networks can be used to enhance the regional feature representation capability, and temporally, the Selective State Space Model can be used to capture dynamic evolution patterns, effectively achieving accurate prediction of future population thermal distribution.

[0043] In this way, by dividing the geographic grid and acquiring thermal images, a spatial discretization representation of population distribution is achieved, providing a structured data foundation for subsequent analysis. Based on a multi-scale convolutional neural network, population density features at different spatial scales are accurately extracted to generate high-resolution thermal distribution information, preserving local clustering details and global distribution trends. A selective state-space model is used to capture the spatiotemporal dynamic evolution of population distribution, and adaptive structural upgrades are achieved by combining the selective state-space model with the thermal model. This reduces computational complexity while ensuring prediction accuracy. Therefore, by considering the significant spatiotemporal heterogeneity and dynamic evolution characteristics of population thermal distribution, constructing population thermal distribution images for future time periods can improve the accuracy of determining population thermal distribution images.

[0044] The steps described above are explained in detail below.

[0045] First, regarding step 110, in this embodiment of the application, N geographic grids can be obtained by following the steps below. Based on this, the image processing method may further include: The first region is divided into N geographic grids based on latitude and longitude, where N is an integer greater than 1. Each geographic grid corresponds to a spatial region within the first region, and the N geographic grids do not overlap.

[0046] In this embodiment of the application, the first region can be divided according to latitude and longitude. N geographic rasters, meaning N geographic rasters correspond to one spatial region in the first region. and These represent the number of grid cells in the latitude and longitude directions, respectively. Furthermore, in this embodiment, the geographic grid can be mapped onto an electronic map to obtain the spatial region corresponding to the geographic raster in the first region.

[0047] Secondly, step 120 involves a certain time window within the first time period. The population distribution in the entire first region can be represented as a two-dimensional heatmap matrix resembling an image. That is, a certain time window To fully capture the spatial correlation information between regions in the first-person thermal distribution image, a multi-scale convolutional neural network is provided to extract spatial feature representations at different perceptual scales. By introducing multiple convolutional kernels with different receptive fields, parallel convolution operations are performed on the input thermal map at multiple scales to enhance the ability to model local and global spatial patterns.

[0048] Based on this, in some embodiments of this application, step 120 may specifically include steps 1201 and 1202, as shown below.

[0049] Step 1201: Centered on each geographic grid, extract the thermal features of the first region at multiple scales from the thermal distribution image of the first population using multiple convolutional kernels with different receptive fields in a multi-scale convolutional neural network. The size of each convolutional kernel corresponds to one scale. Each receptive field corresponds to at least one convolutional kernel.

[0050] For example, suppose we want to analyze the thermal distribution of people in a central urban area, such as the area around People's Square, within 1 square kilometer during the morning rush hour (7:00-9:00) on a weekday. This area, i.e., the first area, is divided into 10-meter × 10-meter geographic grids, for a total of 100 × 100 = 10,000 grids. The population density data of each geographic grid refers to the population thermal value or the population density values ​​that constitute the first population thermal distribution image, which can be understood as pixel value = population density, ranging from 0 to 200 people / grid.

[0051] Based on this, taking a geographic grid near the exit of People's Square subway station (coordinates X=50, Y=50) and its corresponding image center pixel as the center, we need to extract its features at three scales, i.e., three sets of convolutional kernels of different sizes, corresponding to three spatial scales. The small scale, used to obtain local features, can use a 3×3 convolutional kernel with a receptive field of 3×3 grids (30m×30m area), which can be used to capture close-range crowd gatherings around the geographic grid, such as instantaneous crowd congestion at the subway station exit. The medium scale, used to obtain regional features, can use a 7×7 convolutional kernel with a receptive field of 7×7 grids (70m×70m area), which can be used to capture street-level crowd diffusion within the geographic grid, such as the pedestrian distribution at five intersections around the subway station. The large scale, used to obtain global features, can use a 15×15 convolutional kernel with a receptive field of 15×15 grids (150m×150m area), which can be used to capture the crowd trends within the surrounding commercial area, such as the overall pedestrian flow interaction in shopping malls and bus stops around the square.

[0052] Centered on the geographic grid (50, 50), a convolution operation is performed on the thermal distribution image of the first crowd. The output of the 3×3 convolution kernel is the crowd density gradient of the 3×3 area surrounding the geographic grid. For example, if the center pixel value is 200, the values ​​of the four surrounding grids are 180-200, showing a local high-density clustering feature. The output of the 7×7 convolution kernel is the crowd distribution pattern of the 7×7 area surrounding the geographic grid. For example, the values ​​of the two grids in the northeast direction drop sharply to 50, corresponding to road construction causing pedestrians to detour. The output of the 15×15 convolution kernel is the crowd flow trend of the 15×15 area surrounding the geographic grid. For example, the pixel values ​​increase in the southwest direction, corresponding to the shopping mall's morning market attracting crowds.

[0053] By repeating the above process for all 10,000 geographic grids, we can eventually obtain the population thermal characteristics of the first region at multiple scales, namely the local aggregation, medium-range diffusion, and global trend information of the first region.

[0054] Step 1202: Based on the thermal characteristics of the population in the first region at multiple scales, determine the thermal distribution information of the first population in the first region within the first time period.

[0055] Here, feature stitching can be performed on thermal characteristics of populations at all scales, and their dimensions can be rearranged to obtain thermal distribution information of the first population in a region within a first time period. For example, it can be referenced... Figure 3 The population thermal distribution information in the first area is generated as follows: red = extremely high density area (subway station exit), yellow = medium to high density area (business district roads), and blue = low density area (parks and green spaces).

[0056] In this embodiment of the application, step 1201 may specifically include steps 12011 and 12012.

[0057] Step 12011: Centered on each geographic grid, the thermal distribution image of the first population is convolved in parallel using multiple convolution kernels with different receptive fields in a multi-scale convolutional neural network to obtain the spatial thermal feature information of the first region obtained by convolving with at least one convolution kernel at each scale within the first time period.

[0058] In this embodiment of the application, the configurable convolution kernel size is: ,in Indicates the first Kernel size at various scales This indicates the number of convolutional kernels. Each convolutional kernel is a single... A two-dimensional matrix of size 1.5. For the first time interval... Heatmap of each time window For a given kernel size Apply a 2D convolution operation with edge padding, where the amount of edge padding is... In this embodiment of the application, the parallel convolution involved in this step can be achieved by the following formula (1): Among them, each of the above This represents the spatial thermal characteristics of the population obtained by convolving the first region within the first time period or the s-th time window of the first time period with at least one convolution kernel at the s-th scale.

[0059] Step 12012: The spatial population thermal feature information obtained by convolving the first region within the first time period with at least one convolution kernel at each scale in the multi-scale is spliced ​​along the channel dimension to obtain the population thermal feature information of the first region at multiple scales.

[0060] In this embodiment, the outputs of all scales are concatenated along the channel dimension using the following formula (2). The output is the output of step 12011, that is, the spatial crowd thermal feature information of the first region within the first time period or the s-th time window of the first time period obtained by convolving the first region with at least one convolution kernel at each scale in the multi-scale, thus obtaining the crowd thermal feature information of the first region at multiple scales. : In this embodiment of the application, step 1202 may specifically include steps 12021 and 12022.

[0061] Step 12021: Rearrange the channel dimensions in the population thermal feature information of the first region extracted at multiple scales to obtain the joint population thermal feature information of each geographic raster in the first region at multiple scales in the first time period.

[0062] In this embodiment of the application, as shown in formula (3), rearrangement refers to moving the channel dimension to the end to obtain the joint population thermal feature information of each geographic raster in the first region at S scales within the first time period or t time windows of the first time period. : Step 12022: Based on the joint population thermal feature information of each geographic raster in the first region at multiple scales in the first time period, generate the first population thermal distribution information of the first region in the first time period.

[0063] Here, the joint population thermal characteristics of each geographic raster in the first region at multiple scales within the first time period can be determined as the first population thermal distribution information of the first region within the first time period. For example, it is possible to... Indicates the first time period The first time window The joint population thermal feature information at multiple scales of the network grid will be used as a lexical embedding for subsequent selective state-space modeling.

[0064] Therefore, the multi-scale convolutional neural network in this embodiment has the following advantages: First, it integrates local and global information: small-scale convolutional kernels are good at capturing local population density changes, such as the clustering of a certain block, while large-scale convolutional kernels help extract global trends, such as the diffusion effect of a large commercial district. The combination of the two can provide a more comprehensive spatial understanding. Second, its lightweight structure facilitates parallel processing. All scale convolutional operations are independent of each other, supporting parallel computing, resulting in high computational efficiency and suitability for large-scale urban data modeling. In actual implementation, the model simultaneously reads continuous data within the first time period. A sequence of heatmaps for each time window is processed, and all convolutional kernels are applied simultaneously to obtain a four-dimensional tensor with multi-dimensional scale spatial features. The output structure is then regular, facilitating temporal modeling, with each grid position corresponding to a... A three-dimensional feature vector is suitable as input for subsequent time series modeling modules.

[0065] It should be noted that the multiple convolutional kernels include a first convolutional kernel and a second convolutional kernel, with the kernel size of the first convolutional kernel being smaller than that of the second convolutional kernel; the population thermal feature information of the first region at multiple scales includes first population thermal feature information and second population thermal feature information; wherein, the first population thermal feature information is used to characterize the degree of variation in population aggregation within each geographic grid captured by the first convolutional kernel; the population thermal feature information is used to characterize the global population thermal change trend of the first region captured by the second convolutional kernel.

[0066] Therefore, multi-scale convolutional neural networks can be used to extract local and global spatial features simultaneously under different receptive fields. This can fully capture the dynamic and multi-scale characteristics of population distribution in urban space, overcome the limitations of traditional single-point analysis models in spatial dependency modeling, and improve the adaptability and generalization ability of prediction results under different regional conditions.

[0067] Next, regarding step 130, in this embodiment of the application, based on the foregoing content, it is known that in the first time period... Geographic raster grid within time window The thermal value of the population above can be This step aims to solve problems that were previously known. Given the thermal distribution information of the first population within a given time window, predict the second time window, i.e., the... The thermal distribution information of the second population can be expressed by the following formula (4). The thermal distribution information of the first population in the first time window, and the expression of the first population in the second time window by the following formula (5). The second population thermal distribution information.

[0068] This problem essentially falls under the category of typical spatiotemporal sequence prediction, requiring the simultaneous modeling of spatial correlation and temporal dependence.

[0069] Therefore, in this embodiment, the second time period occurs later than the first time period. Accordingly, step 130 may specifically include steps 1301 to 1305, as detailed below.

[0070] Step 1301: Take the joint population thermal feature information of each geographic grid in the first population thermal distribution information at multiple scales in the first time period as a word and input it into the selective state space model.

[0071] For example, it can be As a lexical input selective state-space model.

[0072] Step 1302: Using a selective state-space model, establish the time-dimensional dependency of joint population thermal characteristics information for each geographic grid at multiple scales.

[0073] 1303. Based on dependencies, generate context vectors for the critical time windows of each geographic raster that are adjacent to the second time period in the first time period.

[0074] Step 1304 involves performing a linear transformation on the context vector of the critical time window to obtain joint population thermal feature information for each geographic raster at multiple scales within the second time period. Here, the dynamic evolution features from the first time period are incorporated.

[0075] Step 1305: Based on the joint population thermal feature information of each geographic grid in the first region at multiple scales in the second time period, generate the second population thermal distribution information of the first region in the second time period.

[0076] For example, the joint population thermal characteristics of each geographic raster at multiple scales in the second time period can be summarized to obtain the second population thermal distribution information of the first region in the second time period.

[0077] The training and structure of the selective state-space model in the embodiments of this application are described below.

[0078] For a single region, the time series formed by the evolution of its population heat values ​​over time exhibits a high degree of temporal dependence. This dependence is similar to word order dependence in natural language, where the current state is often strongly influenced by preceding data. Therefore, this application employs an autoregressive approach to introduce the recently popular selective state-space model (Mamba) to enhance the modeling capability of population heat time series patterns. The Mamba model demonstrates superior performance in long-term sequence modeling. It can not only capture population dynamics within local time periods but also effectively remember and utilize early historical distribution features as the sequence grows, thereby significantly alleviating the information forgetting problem common in traditional sequence models.

[0079] Based on this, the first The first time window, the first The spatial features extracted by the multi-scale convolutional neural network at the geographic grid are: We use it as a word embedding input selective state-space model to model contextual dependencies in the time dimension and predict the heat value distribution at the next time step. Specifically, the core update mechanism of the Mamba model is shown in Equations (6) and (7) below: in, It is to learnable parameters Combined with time scale parameters The forgetting matrix obtained by discretization; These are learnable shortcut parameters; parameters , , All are generated from the input, and can be calculated using the following formula (8): in, , , This is the projection matrix. The Softplus function operates element-wise, and its form is: .

[0080] The above formula is essentially composed of four components: first, the input signal. First, use Hadamard product and... The combination can be viewed as an input gate, determining which new information needs to be stored by the network; secondly, the implicit state of the previous moment. and Multiplication acts as a forgetting gate, controlling the decay of the old state; thirdly, the current hidden state. Through projection matrix A linear mapping to the output is equivalent to an output gate; inspired by this, an input-based... The four components work together to provide learning shortcuts and enhance the memory of recently inputted information. It is this synergy that enables Mamba to maintain the continuity of historical information while responding quickly to sudden changes when dealing with long sequences, thereby effectively improving prediction performance.

[0081] The macroscopic structural design of the Mamba module includes: (1) Root Mean Square Normalization (RMSNorm) submodule: normalizes the input and stabilizes the training process; (2) Spatial Awareness Module (SSM): carries the main function of temporal modeling; (3) Residual Connection: strengthens gradient flow and alleviates the gradient vanishing problem in deep networks. In addition, the state space module sequentially integrates various techniques such as linear projection, SiLU activation, selective state space model and gating mechanism to jointly improve the expressive power and dynamic modeling ability of the model.

[0082] Therefore, in terms of time modeling, this proposal draws on the autoregressive approach from natural language processing, introducing a selective state-space model to model the dynamic evolution of population dynamics. This model can dynamically integrate historical state information from different moments in time, adaptively adjusting the focus on different time segments and effectively avoiding the loss of early pattern information. In long-term forecasting, it can simultaneously maintain the ability to capture both high-frequency changes and low-frequency trends, improving the foresight and stability of the forecast.

[0083] Regarding step 140, in some embodiments of this application, step 140 may specifically include: Based on the network structure of the first population thermal distribution image, the joint population thermal feature information of each geographic grid in the second population thermal distribution information at multiple scales in the second time period is rearranged to obtain the second population thermal distribution image of the first region in the second time period.

[0084] In this embodiment of the application, it is based on a first time period. Image sequence of thermal distribution of the first population within a time window The second time period is generated. Thermal distribution images of the second population within a time window Based on the foregoing, spatial encoding has been extracted using a multi-scale convolutional neural network, and then temporal modeling is performed using a selective state-space model. This ultimately yields joint population thermal characteristics of each geographic raster at multiple scales within the second time period. It contains the former The dynamic evolution characteristics of the first step. To predict the... The second population thermal distribution image for each time window can be obtained from the context vector of each geographic raster. Applying a shared linear transformation layer can be done using the following formula (9) to transform... Rearranging the data into a two-dimensional heatmap structure yields the final thermal distribution image of the second population: in, , These are learnable parameters.

[0085] Furthermore, this application embodiment also provides a process for determining the samples for training the model. Based on this, after step 140, the image processing may further include: Based on the thermal distribution image of the second population and the actual thermal distribution image of the third population in the first region during the second time period; The thermal distribution images of the second and third populations are used as training samples to train the target model. The target model includes at least one of the following: multi-scale convolutional neural network and selective state space model.

[0086] Therefore, this application proposes a selective spatiotemporal state-space model for predicting population thermal distribution. It effectively integrates spatial neighborhood perception and dynamic time selection capabilities, constructing a multi-scale convolutional neural network. By applying convolutional kernels with different receptive fields, it extracts population distribution features from various regions at multiple scales, achieving a joint characterization of local and global patterns in the spatial domain. Furthermore, by analogy to the problem of predicting population thermal distribution in natural language processing, the selective state-space model Mamba is introduced in an autoregressive manner. Leveraging its superior performance in long sequence modeling, it captures the dynamic evolution of population thermal distribution in the temporal domain, accurately predicting the thermal value at the next moment.

[0087] It should be noted that the image processing method provided in this application embodiment can play an important role in various scenarios such as smart city governance, service operation decision-making, and public safety management. Based on this, it can be applied to at least one of the following application scenarios: financial payment institutions or third-party payment institutions can realize scenarios such as optimizing merchant outlet layout, dynamically adjusting payment acceptance environment, and accurately launching service activities; location services and map platforms can also be used in scenarios such as traffic prediction and merchant site selection analysis; merchants can use it in scenarios such as customer flow prediction, tenant combination optimization, and site selection for operational activities.

[0088] The following is combined with Figures 3 to 5 The image processing method provided in the embodiments of this application will be described in detail.

[0089] First, taking Shanghai as the first region, with a time range of June 1st to June 7th, 2025, as an example, we will illustrate this. The geographical coordinates of Shanghai range from 30°41'42.41" to 31°51'39.74" N and 120°52'5.34" to 121°58'15.25" E, encompassing the main urban areas and surrounding suburbs of Shanghai. The population heatmap data for constructing the first population heatmap comes from at least one of the following sources: communication base stations, Wi-Fi hotspots, and network location services. By fusing multi-source data, it accurately reflects the population density in different areas at different times. The higher the population heatmap value, the higher the degree of population concentration at that location during the corresponding time period.

[0090] To achieve unified spatial analysis, this embodiment employs Geohash encoding to spatially discretize the study area and selects Geohash 7 level for grid division. At this level, the spatial resolution of each geographic raster is approximately 150m × 150m. Ultimately, the first region is divided into 825 × 799 geographic rasteres, forming a two-dimensional thermal matrix of size 825 × 799. This two-dimensional thermal matrix corresponds to the thermal distribution image of the first population. Figure 3 The image shows a heat map of the Jing'an Temple commercial area on June 1st.

[0091] During the experiment, the thermal distribution images of the first population from June 1st to 6th, 2025, can be used as training samples, and the thermal distribution image of the first population on June 7th, 2025, can be used as test samples to verify the generalization ability of the model. The prediction target is the population thermal value of each geographic grid on the test day. During the training process of the target model using the training samples, the Mean Square Error (MSE) function can be used for model training, and the Mean Absolute Error (MAE) can be used as the evaluation metric.

[0092] After model training and testing, the final MAE on the test set on June 7, 2025, was 16.5923. This result indicates that the model has good accuracy in predicting urban population distribution heatmaps.

[0093] Further visualization analysis results, such as Figure 4 The model shown predicts the thermal distribution image of the population, and... Figure 5 The comparison between the real-world crowd heatmap images shows that in the crowd heatmap images, color saturation / depth visually represents the density of people in a region. Different colors correspond to different density levels, with the core logic being that the darker the color (or the higher the saturation), the greater the crowd density. Based on this, red represents the most densely populated areas, typically corresponding to commercial centers, transportation hubs, popular tourist attractions, and large event venues—the core hotspots for crowd gathering. Yellow represents areas with medium crowd density, such as secondary commercial areas, residential areas, and cultural venues, where crowd activity is at a moderate level—neither the most crowded nor the most deserted. Green represents areas with low crowd density, such as residential areas, parks, and remote streets, where crowd activity is relatively sparse, belonging to low-density or secondary hotspot areas. Blue represents areas with extremely low crowd density, such as suburbs, undeveloped areas, and deserted areas at night—cool zones for crowd activity. It can be observed that the target model can accurately identify high-density pedestrian areas in the city, such as the area around People's Square and Nanjing East Road, the area around Nanjing West Road and Zhangyuan, the intersection of Huangpi South Road and Huaihai Middle Road, and the Yu Garden area—core business districts. These results validate the target model's ability to capture crowd gathering trends in complex urban spatial environments. This will provide valuable forward-looking reference for subsequent business operations such as optimizing the payment environment and planning service activities, further promoting the efficient coupling of payment infrastructure and urban pedestrian flow dynamics.

[0094] Therefore, this application proposes a prediction algorithm, ST-SSSM, based on a spatiotemporally selective state-space model. This method integrates multi-scale convolutional neural networks and state-space modeling mechanisms, jointly capturing local and global patterns in the spatial domain and the dynamic evolution of population thermal patterns in the temporal domain, thereby improving the model's ability to model complex population dynamics and its predictive performance. Its advantages are threefold: First, it proposes a spatiotemporally selective state-space model to achieve dynamic spatiotemporal feature fusion. By combining the long-range dependency modeling capability of the state-space model with a dynamic selection mechanism, it adaptively adjusts the focus on historical information, effectively alleviating the problem of traditional models forgetting early spatiotemporal patterns and achieving accurate prediction of population thermal evolution. Second, it applies multi-scale convolutional neural networks to enhance spatial pattern characterization. It uses multi-scale two-dimensional convolutional kernels to simultaneously extract local and global spatial features, fully capturing the dynamic and multi-scale characteristics of urban population distribution, overcoming the limitations of traditional single-point analysis in spatial dependency modeling. Third, it applies the state-space model in the temporal domain to capture the dynamic evolution of population thermal patterns. By analogy to the task of predicting population thermal distribution, we introduce the Mamba selective state space structure to efficiently model the thermal evolution of population over long periods of time in an autoregressive manner, taking into account both the ability to capture high-frequency changes and low-frequency trends.

[0095] Based on the same inventive concept, this application also provides an image processing apparatus. (Specifically combined with...) Figure 6 Please provide a detailed explanation.

[0096] Figure 6 This is a schematic diagram of the structure of an image processing apparatus provided in one embodiment of this application.

[0097] In some embodiments of this application, Figure 6 The image processing device shown can be installed in a computer device.

[0098] like Figure 6 As shown, the image processing device 60 may specifically include: The acquisition module 601 is used to acquire the thermal distribution image of the first population in the first region within the first time period. The first region includes N geographic grids, where N is an integer greater than 1. The determination module 602 is used to determine the thermal distribution information of the first population in the first region within the first time period, based on the thermal distribution image of the first population, using a multi-scale convolutional neural network with each geographic raster as the center. The determining module 602 is also used to determine the thermal distribution information of the second population in the first region during the second time period based on the thermal distribution information of the first population using a selective state-space model. Module 603 is used to construct a thermal distribution image of the second population in the first region within a second time period based on the thermal distribution information of the second population.

[0099] In this embodiment, a thermal distribution image of a first population in a first region within a first time period can be obtained. The first region includes N geographic grids, where N is an integer greater than 1. Centered on each geographic grid, a multi-scale convolutional neural network is used to determine the thermal distribution information of the first population in the first region within the first time period based on the thermal distribution image of the first population. A selective state-space model is used to determine the thermal distribution information of a second population in the first region within a second time period based on the thermal distribution information of the first population. Based on the thermal distribution information of the second population, a thermal distribution image of the second population in the first region within the second time period is constructed. In this way, by dividing the geographic grid and acquiring thermal images, a spatial discretization representation of population distribution is achieved, providing a structured data foundation for subsequent analysis. Based on a multi-scale convolutional neural network, population density features at different spatial scales are accurately extracted to generate high-resolution thermal distribution information, preserving local clustering details and global distribution trends. A selective state-space model is used to capture the spatiotemporal dynamic evolution of population distribution, and adaptive structural upgrades are achieved by combining the selective state-space model with the thermal model. This reduces computational complexity while ensuring prediction accuracy. Therefore, by considering the significant spatiotemporal heterogeneity and dynamic evolution characteristics of population thermal distribution, constructing population thermal distribution images for future time periods can improve the accuracy of determining population thermal distribution images.

[0100] The image processing apparatus 60 in the embodiments of this application will be described in detail below.

[0101] In one or more optional embodiments, the determining module 602 may be specifically used to extract the thermal feature information of the first region at multiple scales from the thermal distribution image of the first population, with each geographic grid as the center, through multiple convolutional kernels with different receptive fields in a multi-scale convolutional neural network, wherein the convolutional kernel size of one convolutional kernel corresponds to one scale. Based on the thermal characteristics of the population in the first region at multiple scales, the thermal distribution information of the first population in the first region during the first time period is determined.

[0102] In one or more optional embodiments, the image processing device 60 in this application embodiment may further include a convolution module, which is used to perform parallel convolution on the thermal distribution image of the first population with each geographic grid as the center and through multiple convolution kernels of different receptive fields in a multi-scale convolutional neural network, to obtain spatial population thermal feature information of the first region obtained by convolution with at least one convolution kernel at each scale in the first time period. In this embodiment, the image processing device 60 may further include a stitching module, which is used to stitch together the spatial crowd thermal feature information obtained by convolving the first region within a first time period through at least one convolution kernel at each scale in multiple scales along the channel dimension to obtain the crowd thermal feature information of the first region at multiple scales.

[0103] In one or more optional embodiments, the image processing device 60 in this application embodiment may further include a rearrangement module, which is used to rearrange the channel dimensions in the thermal feature information of the crowd in the first region extracted at multiple scales, so as to obtain the joint thermal feature information of the crowd at multiple scales for each geographic grid in the first region in the first time period. In this embodiment, the image processing device 60 may further include a generation module, which is used to generate first population thermal distribution information of the first region within the first time period based on the joint population thermal feature information of each geographic raster in the first region at multiple scales within the first time period.

[0104] In one or more optional embodiments, the plurality of convolution kernels include a first convolution kernel and a second convolution kernel, wherein the kernel size of the first convolution kernel is smaller than the kernel size of the second convolution kernel; The thermal characteristics of the population in the first region at multiple scales include thermal characteristics of the first population and thermal characteristics of the second population; Among them, the first population thermal feature information is used to characterize the degree of change in population aggregation within each geographic grid captured by the first convolutional kernel; the population thermal feature information is used to characterize the global population thermal change trend of the first region captured by the second convolutional kernel.

[0105] In one or more optional embodiments, the first time period includes M time windows, and the thermal distribution information of the first population includes the thermal distribution information of the first population in the first region within each time window.

[0106] In one or more optional embodiments, the determining module 602 can also be used to, when the occurrence time of the second time period is later than the occurrence time of the first time period, use the joint population thermal feature information of each geographic grid in the first population thermal distribution information at multiple scales in the first time period as a word and input it into the selective state space model. In this embodiment, the image processing device 60 may further include a building module, used to establish the time-dimensional dependency of joint population thermal feature information of each geographic grid at multiple scales through a selective state-space model. In this embodiment, the image processing device 60 may further include a generation module, which generates a context vector for each geographic raster of a critical time window adjacent to the second time period in the first time period, based on the dependency relationship. In this embodiment, the image processing device 60 may further include a transformation module for performing a linear transformation on the context vector of the critical time window to obtain joint population thermal feature information of each geographic grid at multiple scales in the second time period. The generation module is also used to generate second population thermal distribution information of the first region in the second time period based on the joint population thermal feature information of each geographic raster in the first region at multiple scales in the second time period.

[0107] In one or more optional embodiments, the construction module 603 may be specifically used to rearrange the joint population thermal feature information of each geographic grid in the second population thermal distribution information at multiple scales in the second time period according to the network structure of the first population thermal distribution image, so as to obtain the second population thermal distribution image of the first region in the second time period.

[0108] In one or more optional embodiments, the image processing device 60 in this application embodiment may further include a division module for dividing the first region according to latitude and longitude to obtain N geographic grids; wherein, one geographic grid corresponds to one spatial region in the first region, and the N geographic grids do not overlap.

[0109] In one or more optional embodiments, the image processing device 60 in this application embodiment may further include a training module, used to train the image based on the thermal distribution image of the second population and the actual thermal distribution image of the third population in the first region during the second time period. The thermal distribution images of the second and third populations are used as training samples to train the target model. The target model includes at least one of the following: multi-scale convolutional neural network and selective state space model.

[0110] Based on the same inventive concept, this application also provides a computer device. (Specifically combined with...) Figure 7 Please provide a detailed explanation.

[0111] Figure 7 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application.

[0112] like Figure 7 As shown, the computer device may include at least one of the following as described in the embodiments of this application: an electronic device, a server. The computer device may include a processor 701 and a memory 702 storing computer program instructions.

[0113] Specifically, the processor 701 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0114] Memory 702 may include mass storage for data or instructions. For example, and not limitingly, memory 702 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 702 may include removable or non-removable (or fixed) media. Where appropriate, memory 702 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 702 is non-volatile solid-state memory. In a particular embodiment, memory 702 includes solid-state storage (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0115] The processor 701 implements any of the image processing methods described in the above embodiments by reading and executing computer program instructions stored in the memory 702.

[0116] In one example, the computer device may also include a communication interface 703 and a bus 710. Wherein, as... Figure 7 As shown, the processor 701, memory 702, and communication interface 703 are connected through bus 710 and complete communication with each other.

[0117] The communication interface 703 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0118] Bus 710 includes hardware, software, or both, that couples components of a flow control device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 710 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0119] The computer device can execute the image processing method described in the embodiments of this application, thereby achieving the combination Figures 1 to 6 The image processing method and apparatus described herein.

[0120] Furthermore, in conjunction with the image processing methods in the above embodiments, this application embodiment can provide a computer-readable storage medium for implementation. This computer-readable storage medium stores computer program instructions; when executed by a processor, these computer program instructions implement any of the image processing methods in the above embodiments.

[0121] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0122] The functional blocks shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0123] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0124] The above are merely specific embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. An image processing method, characterized in that, include: Obtain a thermal distribution image of the first population in a first region within a first time period. The first region includes N geographic grids, where N is an integer greater than 1. Centered on each of the geographic grids, a multi-scale convolutional neural network is used to determine the thermal distribution information of the first population in the first area within the first time period based on the thermal distribution image of the first population. Using a selective state-space model, the thermal distribution information of the second population in the first area during the second time period is determined based on the thermal distribution information of the first population. Based on the thermal distribution information of the second population, a thermal distribution image of the second population in the first region during the second time period is constructed.

2. The method according to claim 1, characterized in that, The step of determining the thermal distribution information of the first population in the first area within the first time period, centered on each of the geographic grids and using a multi-scale convolutional neural network based on the thermal distribution image of the first population, includes: Centered on each of the geographic grids, the thermal features of the first region at multiple scales are extracted from the thermal distribution image of the first population through multiple convolutional kernels with different receptive fields in a multi-scale convolutional neural network. The size of one convolutional kernel corresponds to one scale. Based on the thermal characteristics of the population in the first region at multiple scales, the thermal distribution information of the first population in the first region during the first time period is determined.

3. The method according to claim 2, characterized in that, The step of extracting population thermal feature information of the first region at multiple scales from the first population thermal distribution image using multiple convolutional kernels with different receptive fields in a multi-scale convolutional neural network, centered on each of the geographic grids, includes: Centered on each of the geographic grids, the thermal distribution image of the first population is convolved in parallel using multiple convolution kernels with different receptive fields in a multi-scale convolutional neural network, to obtain spatial population thermal feature information of the first region obtained by convolving with at least one convolution kernel at each scale within the first time period. The spatial population thermal feature information obtained by convolving the first region within the first time period with at least one convolution kernel at each of the multiple scales is then spliced ​​along the channel dimension to obtain the population thermal feature information of the first region at multiple scales.

4. The method according to claim 2 or 3, characterized in that, The step of determining the thermal distribution information of the first population in the first time period based on the thermal characteristics of the first population in the first region at multiple scales includes: The channel dimensions in the population thermal feature information of the first region extracted at the multiple scales are rearranged to obtain the joint population thermal feature information of each geographic grid in the first region at multiple scales within the first time period. Based on the joint population thermal feature information of each geographic grid in the first region at multiple scales within the first time period, the first population thermal distribution information of the first region within the first time period is generated.

5. The method according to claim 2, characterized in that, The plurality of convolutional kernels includes a first convolutional kernel and a second convolutional kernel, wherein the kernel size of the first convolutional kernel is smaller than the kernel size of the second convolutional kernel; The thermal characteristics of the population in the first region at multiple scales include thermal characteristics of the first population and thermal characteristics of the second population; Wherein, the first population thermal feature information is used to characterize the degree of change in population aggregation within each geographic grid captured by the first convolutional kernel; the population thermal feature information is used to characterize the global population thermal change trend of the first region captured by the second convolutional kernel.

6. The method according to any one of claims 1 to 5, characterized in that, The first time period includes M time windows, and the thermal distribution information of the first population includes the thermal distribution information of the first population in the first region within each time window.

7. The method according to claim 1, characterized in that, The second time period occurs later than the first time period; determining the thermal distribution information of the second population in the first region within the second time period using a selective state-space model based on the thermal distribution information of the first population includes: The joint population thermal feature information of each geographic grid in the first population thermal distribution information at multiple scales within the first time period is used as a word unit and input into the selective state space model. The selective state-space model is used to establish the time-dimensional dependency of joint population thermal characteristics information of each geographic grid at multiple scales. Based on the dependencies, generate a context vector for each geographic raster of the critical time window adjacent to the second time period within the first time period; A linear transformation is performed on the context vector of the critical time window to obtain the joint population thermal feature information of each geographic raster at multiple scales within the second time period; Based on the joint population thermal feature information of each geographic raster in the first region at multiple scales in the second time period, the second population thermal distribution information of the first region in the second time period is generated.

8. The method according to claim 7, characterized in that, The step of constructing a thermal distribution image of the second population in the first region within the second time period based on the thermal distribution information of the second population includes: Based on the network structure of the first population thermal distribution image, the joint population thermal feature information of each geographic grid in the second population thermal distribution information at multiple scales in the second time period is rearranged to obtain the second population thermal distribution image of the first region in the second time period.

9. The method according to claim 1, characterized in that, The method further includes: The first region is divided according to latitude and longitude to obtain N geographic grids; wherein, one geographic grid corresponds to one spatial region in the first region, and the N geographic grids do not overlap.

10. The method according to claim 1, characterized in that, The method further includes: Based on the thermal distribution image of the second population and the actual thermal distribution image of the third population in the first region during the second time period; The thermal distribution images of the second population and the third population are used as training samples to train the target model, which includes at least one of the following: the multi-scale convolutional neural network and the selective state space model.

11. An image processing apparatus, characterized in that, include: The acquisition module is used to acquire a thermal distribution image of the first population in a first region within a first time period. The first region includes N geographic grids, where N is an integer greater than 1. The determination module is used to determine the thermal distribution information of the first population in the first area within the first time period, based on the thermal distribution image of the first population, using a multi-scale convolutional neural network with each of the geographic grids as the center. The determining module is further configured to determine the thermal distribution information of the second population in the first area during the second time period based on the thermal distribution information of the first population using a selective state-space model. The construction module is used to construct a thermal distribution image of the second population in the first region during the second time period based on the thermal distribution information of the second population.

12. A computer device, characterized in that, The computer device includes: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements the image processing method as described in any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the image processing method as described in any one of claims 1-10.

14. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the image processing method as described in any one of claims 1-10.

Images