Methods, apparatus, electronic devices and media for constructing image retrieval databases

Abstract

This application discloses a method, apparatus, electronic device, and medium for constructing an image retrieval database. The method includes: acquiring the population density, current region type, and remote sensing image of the region to be processed; determining the current spatial grid size based on the population density and current region type; dividing the region to be processed into grid regions based on the current spatial grid size; matching the remote sensing image and the grid regions to obtain candidate grid regions that meet preset matching conditions; and extracting visual features of the remote sensing image in the candidate grid regions to construct an image retrieval database. Thus, the grid granularity is automatically adjusted according to the population density and region type of different regions to meet the diverse needs of different regions. Furthermore, matching the candidate grid regions with the remote sensing image accurately extracts the visual features in the candidate grid regions to establish the image retrieval database, thereby achieving efficient and accurate retrieval of remote sensing images.

Description

Technical Field

[0001] This application relates to the field of remote sensing image technology, and in particular to a method, apparatus, electronic device and medium for constructing an image retrieval database. Background Technology

[0002] Remote sensing imagery, as a crucial means of acquiring Earth's surface information, has been widely applied in various fields such as urban planning, environmental monitoring, disaster early warning, and resource surveys. With the rapid development of remote sensing technology and the increasing variety of satellite sensors, remote sensing data exhibits characteristics of multi-source, multi-temporal, multi-resolution, and multi-modal processing. Faced with the ever-growing volume of remote sensing imagery data, how to efficiently and accurately achieve image retrieval has become one of the most critical research directions in remote sensing information processing. Summary of the Invention

[0003] In view of this, one aspect of this application provides a method for constructing an image retrieval database, the method comprising:

[0004] Acquire the population density, current area type, and remote sensing image of the area to be processed;

[0005] Based on the population density and the current region type, determine the current spatial grid size; and based on the current spatial grid size, divide the region to be processed to obtain grid regions;

[0006] The remote sensing image to be processed and the grid region are matched to obtain candidate grid regions whose matching degree meets the preset conditions;

[0007] Extract the visual features of the remote sensing image to be processed from the candidate grid region;

[0008] An image retrieval database is constructed based on the aforementioned visual features.

[0009] Optionally, determining the current spatial grid size based on the population density and the current region type includes:

[0010] Obtain the current baseline grid size and the pre-constructed mapping relationship; wherein, the mapping relationship is the correspondence between region type and region weighting factor; the region type represents that the higher the population density, the larger the corresponding region weighting factor;

[0011] Based on the mapping relationship and the current region type, determine the weighting factor for the current region;

[0012] The current spatial grid size is determined by the current baseline grid size, the current region weighting factor, and the population density; wherein the population density is negatively correlated with the current spatial grid size, and the current region weighting factor is positively correlated with the current spatial grid size.

[0013] Optionally, obtain the current baseline mesh size, including:

[0014] Determine whether the population density is greater than a density threshold;

[0015] If it is greater than the given size, the first size will be used as the current reference mesh size.

[0016] If not greater than, determine whether the population density is greater than the density threshold of a preset multiple; the preset multiple is less than 1; if greater than, use the second size as the current reference grid size; if not greater than, use the third size as the current reference grid size; wherein, the first size is smaller than the second size, and the second size is smaller than the third size.

[0017] Optionally, matching the remote sensing image to be processed and the grid region to obtain candidate grid regions whose matching degree meets preset conditions includes:

[0018] Determine a first spatial consistency index between the remote sensing image to be processed and the grid region; and determine whether there is a target grid with a first spatial consistency index of 100%.

[0019] If it exists, the target mesh will be used as the candidate mesh region;

[0020] If it does not exist, proceed with the following steps:

[0021] The missing data in the grid region is reconstructed to obtain the reconstructed grid region;

[0022] A second spatial consistency index is determined between the reconstructed grid region and the remote sensing image to be processed; and regions with the second spatial consistency index greater than the index threshold are selected as candidate grid regions.

[0023] Optionally, constructing the image retrieval database based on the visual features includes:

[0024] Assign a unique spatial identifier to the grid region, and assign a unique image identifier to the image corresponding to the grid region;

[0025] Obtain the spatial attribute information of the candidate grid region; and preprocess the visual features to obtain a visual feature vector;

[0026] Based on the spatial identifier and the image identifier, construct the association between the visual feature vector and the spatial attribute information;

[0027] Based on the visual feature vectors, a visual index based on vector similarity is constructed; based on the spatial attribute information, a location index based on spatial topology is constructed; and based on the association relationship, a joint index of visual space is constructed to obtain the image retrieval database.

[0028] Optionally, before matching the remote sensing image to be processed with the grid region, the method includes:

[0029] Extract image metadata from the remote sensing image to be processed;

[0030] The image metadata is parsed; and based on the parsing results, it is determined whether the remote sensing image to be processed contains coordinate reference system information.

[0031] If included, based on the coordinate reference system information, read the spatial geographic information of the remote sensing image to be processed; and proceed to the step of matching the remote sensing image to be processed with the grid area.

[0032] Optionally, if the coordinate reference system information does not exist, the method includes:

[0033] Extract keywords from the filenames of the remote sensing images to be processed;

[0034] Matching is performed in a pre-built coordinate reference system database based on the keywords in the file name;

[0035] Based on the matching results, determine whether a matching target coordinate reference system exists in the coordinate reference system database;

[0036] If it exists, based on the target coordinate reference system, read the spatial geographic information and proceed to the step of matching the remote sensing image to be processed with the grid area;

[0037] If it does not exist, the remote sensing image to be processed is converted to the default coordinate reference system, and the process proceeds to the step of matching the remote sensing image to be processed with the grid region.

[0038] Another aspect of this application provides an apparatus for constructing an image retrieval database, the apparatus comprising:

[0039] The target acquisition module is used to acquire the population density, current area type, and remote sensing image of the area to be processed.

[0040] The spatial grid size determination module is used to determine the current spatial grid size based on the population density and the current region type; and to divide the region to be processed based on the current spatial grid size to obtain a grid region;

[0041] The grid matching module is used to match the remote sensing image to be processed with the grid region to obtain candidate grid regions whose matching degree meets preset conditions;

[0042] A visual feature extraction module is used to extract the visual features of the remote sensing image to be processed in the candidate grid region;

[0043] The database construction module is used to construct an image retrieval database based on the visual features.

[0044] Another aspect of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor executes the computer program to implement the steps of the method for constructing the image retrieval database.

[0045] Another aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for constructing the image retrieval database.

[0046] The image retrieval database construction method, apparatus, electronic device, and medium provided in this application have the following beneficial effects: Based on the geographic semantic features of different regions, such as population density and region type, the spatial grid is dynamically divided; that is, the grid granularity is automatically adjusted to ensure fine division of high-population-density areas and efficient division of low-population-density areas. Precise grid planning is performed based on the actual geographic characteristics of different regions, thereby meeting the diverse needs of complex geographic scenarios. Furthermore, remote sensing images are matched based on the dynamically adjusted spatial grid, accurately extracting visual features from candidate grid areas, and establishing an image retrieval database based on these visual features. This provides accurate data support for remote sensing image retrieval, thereby achieving efficient and accurate retrieval of remote sensing images. Attached Figure Description

[0047] Figure 1 A flowchart illustrating a method for constructing an image retrieval database provided in an embodiment of this application;

[0048] Figure 2 A schematic diagram illustrating the principle of a method for constructing an image retrieval database provided in an embodiment of this application;

[0049] Figure 3 A schematic diagram illustrating the principle of a method for obtaining the current reference mesh size provided in an embodiment of this application;

[0050] Figure 4 A schematic diagram illustrating the principle of a method for constructing an image retrieval database according to another embodiment of this application;

[0051] Figure 5 A schematic diagram of the structure of an image retrieval database construction device provided in an embodiment of this application;

[0052] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0053] The reference numerals in the attached diagram are as follows: 50 is the target acquisition module, 51 is the spatial grid size determination module, 52 is the grid matching module, 53 is the visual feature extraction module, 54 is the database construction module, 60 is the memory, 61 is the processor, 62 is the display screen, 63 is the input / output interface, 64 is the communication interface, 65 is the power supply, 66 is the communication bus, 601 is the computer program, 602 is the operating system, and 603 is the data. Detailed Implementation

[0054] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0055] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0056] Figure 1 This is a flowchart illustrating a method for constructing an image retrieval database provided in an embodiment of this application, as shown below. Figure 1 As shown, the method includes:

[0057] S10: Acquire the population density, current area type, and remote sensing image of the area to be processed;

[0058] Current image spatial partitioning generally relies on fixed grid partitioning strategies, which divide remote sensing images according to predetermined rules, such as fixed-size tiles. While this method is simple to implement, it ignores the complexity and non-uniformity of ground feature distribution within the image. For example, urban areas may have high-density building clusters, while suburban areas may have sparse ground features. Fixed grids struggle to adaptively reflect this spatial heterogeneity, leading to the fragmentation of key feature regions or the redundant partitioning of insignificant areas. This not only affects the completeness of subsequent feature representation but also reduces the accuracy of remote sensing image matching and retrieval.

[0059] Figure 2 This is a schematic diagram illustrating the principle of a method for constructing an image retrieval database provided in an embodiment of this application. To solve the aforementioned technical problems, such as... Figure 2 As shown, in one optional embodiment, the method for constructing the image retrieval database provided in this application includes constructing an adaptive grid region.

[0060] Specifically, it is understood that different geographical regions have different population densities and different region types. To ensure that different regions have different requirements for delineation accuracy, in one optional embodiment, the population density, current region type, and remote sensing image of the region to be processed are acquired.

[0061] The area to be processed can be a region within a city or a suburban area. This application does not limit the rules for dividing the area or its size. The region type refers to the type of area to be processed. In specific embodiments, it can be divided according to administrative regions (provinces, cities, counties, communities) or according to urban and rural settlements; this application does not limit this. For example, in an optional embodiment, the region type can be divided into urban, suburban, rural, and mountainous areas, etc.

[0062] It should be noted that the remote sensing images to be processed refer to images related to the area to be processed, obtained from an open-source remote sensing image database. This application does not limit the acquisition method, quantity, or approach of the remote sensing images to be processed. Furthermore, it should be noted that, in an optional embodiment, the population density of the area to be processed can be obtained based on a GIS (Geographic Information System Database).

[0063] S11: Determine the current spatial grid size based on population density and current region type; and divide the region to be processed based on the current spatial grid size to obtain the grid region;

[0064] Furthermore, to meet the different regional segmentation accuracy requirements, the current spatial grid size suitable for segmenting the region to be processed is determined based on population density and the current region type. Understandably, in areas with high population density, higher segmentation accuracy is required, and a smaller current spatial grid size can be selected. Region type is also closely related to population density. If the region type indicates a higher level of development in the region to be processed, the population density is often higher, and a smaller current spatial grid size can be selected in this case.

[0065] After determining the current spatial grid size, the area to be processed is divided based on this size, resulting in a large number of grid regions. These grid regions can be of any shape, such as rectangles, and this application does not limit their shape. However, to facilitate subsequent database construction calculations and processing, in a preferred embodiment, the grid region is a square, and the current spatial grid size refers to the side length of the square grid region.

[0066] In one alternative embodiment, an adaptive grid is generated in real time based on GIS attribute data, supporting grid computing caching and distributed computing (e.g., edge node execution).

[0067] S12: Match the remote sensing image to be processed with the grid area to obtain candidate grid areas that meet the preset matching conditions;

[0068] Based on the above embodiments, in order to construct the relationship between grid areas and remote sensing images to be processed, thereby establishing an image retrieval database to provide data support for subsequent image retrieval. Specifically, as follows... Figure 2 As shown, further, image-grid matching is performed, the remote sensing image to be processed is matched with the divided grid area, and candidate grid areas that meet the preset matching conditions are extracted based on the matching results.

[0069] In one optional embodiment, the matching degree calculation can be based on spatial relationships such as area coverage and latitude-longitude consistency, and this application does not limit this calculation. A higher matching degree indicates a closer proximity between the image and the corresponding geographic location of the grid area; that is, the more likely the image originates from the corresponding grid geographic location.

[0070] S13: Extract the visual features of the remote sensing image to be processed in the candidate grid region;

[0071] S14: Construct an image retrieval database based on visual features.

[0072] To achieve remote sensing image retrieval, it is necessary to establish a correspondence between remote sensing images and actual geographical locations; that is, to establish a correspondence between the remote sensing images to be processed and candidate grid areas. Specifically, such as... Figure 2As shown, visual features of the corresponding remote sensing images to be processed are extracted from the candidate grid regions to establish an image retrieval database based on these visual features. Therefore, in subsequent practical applications, image retrieval can be performed based on these visual features within this image retrieval database.

[0073] It should be noted that, in one optional embodiment, to improve data processing efficiency, the visual features of the remote sensing image to be processed in the candidate grid region can be extracted in parallel. Specifically, multiple GPU nodes can be deployed to perform feature extraction in parallel, supporting multiple models such as ResNet / ViT.

[0074] The image retrieval database construction method provided in this application is applicable to massive amounts of high spatiotemporal resolution data generated by high-resolution satellites, UAV remote sensing, etc. Based on the image retrieval database constructed in this application, it is possible to quickly locate changing areas in disaster emergency response and accurately identify illegal construction targets in smart city management, providing efficient and reliable retrieval support for the intelligent application of remote sensing images.

[0075] The method provided in this application can also be applied to urban refined governance scenarios. Specifically, it supports dynamic grid division with millimeter-level geographic entity perception, can automatically identify micro-change areas in the urban renewal process, and achieve sub-pixel-level detection of elements such as building outline evolution and road topology changes.

[0076] It can also be applied to disaster emergency response scenarios. That is, by aligning multimodal features with unified spatiotemporal benchmarks, it can complete three-dimensional dynamic comparative analysis of the impact range of disasters across time phases, and support real-time quantitative assessment of factors such as post-earthquake building damage and flood inundation range.

[0077] In addition, it can also be applied to agricultural resource monitoring scenarios. Specifically, by integrating multispectral time-series features and spatial autocorrelation models, it can achieve spatial clustering analysis of crop growth in large-scale farmland, and support precision agriculture applications such as growth cycle anomaly detection and yield prediction.

[0078] For remote sensing data assetization scenarios, a joint embedding space of visual, spatial, and semantic elements can be constructed, and high-quality image slices that meet specific spatial distribution patterns can be automatically selected, significantly improving the construction efficiency of remote sensing data cubes.

[0079] Therefore, the image retrieval database construction method provided in this application dynamically divides the spatial grid according to the geographic semantic features such as population density and region type of different areas. Specifically, it automatically adjusts the grid granularity to ensure fine division of high-population-density areas and efficient division of low-population-density areas. Precise grid planning is performed based on the actual geographic characteristics of different regions, thereby meeting the diverse needs of complex geographic scenarios. Furthermore, remote sensing images are matched based on the dynamically adjusted spatial grid, accurately extracting visual features from candidate grid areas. An image retrieval database is then established based on these visual features, providing accurate data support for remote sensing image retrieval and thus achieving efficient and accurate retrieval of remote sensing images.

[0080] In one alternative embodiment, determining the current spatial grid size based on population density and current region type includes:

[0081] Obtain the current baseline grid size and the pre-built mapping relationship; where the mapping relationship is the correspondence between region type and region weighting factor; the region type represents that the higher the population density, the larger the corresponding region weighting factor;

[0082] Determine the weighting factor for the current region based on the mapping relationship and the current region type;

[0083] The current spatial grid size is determined by the current baseline grid size, the current regional weighting factor, and the population density; among them, the population density is negatively correlated with the current spatial grid size, and the current regional weighting factor is positively correlated with the current spatial grid size.

[0084] In one optional embodiment, to accurately calculate the current spatial grid size of different regions to be processed, a correspondence between region types and region weighting factors is pre-constructed to obtain a mapping relationship. This mapping relationship can be stored in tabular form or in text form, and this application does not limit this.

[0085] In a specific embodiment, the region type represents a higher population density, that is, a higher corresponding population density, which in turn represents a higher level of development of the region to be processed, and thus a larger corresponding region weighting factor. In fact, it can be understood that there is a positive correlation between population density and region weighting factor.

[0086] For example, the population density of cities is higher than that of suburbs. When performing adaptive grid partitioning, the accuracy requirement for cities is higher than that for suburbs. Therefore, the regional weighting factor for cities is greater than that for suburbs.

[0087] In one optional embodiment, the region types include urban, suburban, and rural areas, and the region weighting factor corresponding to urban areas is 1.0±0.1, the region weighting factor corresponding to suburban areas is 0.5±0.1, and the region weighting factor corresponding to rural areas is 0.2±0.1.

[0088] Understandably, based on the pre-built mapping relationship, the current region type can be found in the mapping relationship to correspond to the current region weighting factor. Furthermore, the current spatial grid size is determined by the current baseline grid size, the current region weighting factor, and the population density. In one optional embodiment, the specific calculation formula is formula (1):

[0089] (1)

[0090] in, Given the current spatial grid size, For the current baseline grid size, As the weighting factor for the current region, For adjustment coefficients, Density threshold For population density. In an optional embodiment, the adjustment coefficient... It can prevent the elimination of a zero adjustment factor and can be set to 0.01. In an optional embodiment, the current reference mesh size... The value ranges from 500 meters to 5000 meters.

[0091] According to formula (1), in the specific embodiment, the population density With current spatial grid size There is a negative correlation, that is, population density The larger the size, the larger the corresponding current spatial grid size. The smaller the value, the lower the weighting factor for the current region. With current spatial grid size There is a positive correlation, that is, the current regional weighting factor When it is larger, the current spatial grid size The larger.

[0092] It should be noted that the density threshold A pre-set population density threshold, and the density threshold With current spatial grid size They are positively correlated. In specific embodiments, the appropriate threshold can be selected based on actual business needs. For example, for scenarios requiring high accuracy, a smaller density threshold can be chosen. This improves mesh generation accuracy. In one optional embodiment, a density threshold... The value range can be set from 500 people / square kilometer to 2000 people / square kilometer.

[0093] Furthermore, it should be noted that in one optional embodiment, when the population density change rate reaches a preset value, adaptive grid space partitioning is triggered, thereby achieving dynamic updates to the image retrieval database and ensuring its accuracy. However, it is worth noting that in specific embodiments, population density changes may occur in certain areas. To conserve computing resources, grid re-partitioning and image retrieval database updates can be performed only on the target areas where the population density change rate reaches the preset value.

[0094] Figure 3 This is a schematic diagram illustrating the principle of a method for obtaining the current reference mesh size provided in an embodiment of this application. Based on the above embodiment, as an optional embodiment, such as... Figure 3 As shown, obtaining the current baseline mesh size includes:

[0095] Determine whether the population density is greater than the density threshold;

[0096] If it is larger, the first dimension will be used as the current reference mesh size;

[0097] If not greater than, determine whether the population density is greater than a preset multiple of the density threshold; the preset multiple is less than 1; if greater than, use the second size as the current reference grid size; if not greater than, use the third size as the current reference grid size; wherein, the first size is smaller than the second size, and the second size is smaller than the third size.

[0098] Based on the above embodiments, and according to formula (1), the current reference grid size is closely related to the current spatial grid size, that is, they are positively correlated. Therefore, in order to further improve the accuracy of the current spatial grid size, in an optional embodiment, a current reference grid size more suitable for the area to be processed is obtained based on population density and density threshold.

[0099] Specifically, in specific embodiments, such as Figure 3 As shown, it determines whether the population density is greater than a density threshold. If it is, the first size is used as the current reference grid size. Otherwise, it further determines whether the population density is greater than a preset multiple of the density threshold, for example, whether the population density is greater than one-third of the density threshold. If it is, the second size is used as the current reference grid size. If it is still not greater, the third size is used as the current reference grid size.

[0100] It should be noted that the first dimension is smaller than the second dimension, and the second dimension is smaller than the third dimension. For example, the first dimension is 1000 meters, the second dimension is 2000 meters, and the third dimension is 5000 meters. Understandably, the higher the population density, the smaller the corresponding dimension. That is, population density is negatively correlated with the current baseline grid size.

[0101] As an optional embodiment, the remote sensing image to be processed and the grid region are matched to obtain candidate grid regions whose matching degree meets preset conditions, including:

[0102] Determine the first spatial consistency index between the remote sensing image to be processed and the grid area; and determine whether there is a target grid with a first spatial consistency index of 100%.

[0103] If it exists, the target grid will be used as a candidate grid region;

[0104] If it does not exist, proceed with the following steps:

[0105] Reconstruct the missing data in the grid area to obtain the reconstructed grid area;

[0106] Determine the second spatial consistency index between the reconstructed grid region and the remote sensing image to be processed; and select regions with a second spatial consistency index greater than the index threshold as candidate grid regions.

[0107] like Figure 2 As shown in the specific embodiment, in order to improve the matching accuracy between the remote sensing image to be processed and the grid area, and thus ensure the data support of the subsequent image retrieval database, the image grid matching method provided in this application is a two-level matching strategy.

[0108] Specifically, in a specific embodiment, a first spatial consistency index between the remote sensing image to be processed and the grid area is first calculated. The first spatial consistency index is used to reflect the spatial consistency between the remote sensing image to be processed and the grid area. The first spatial consistency index may include, but is not limited to, area coverage and latitude and longitude consistency. This application does not limit this.

[0109] Furthermore, in the two-level matching strategy, the core idea of ​​the first-level strategy is to obtain as much complete coverage as possible, that is, to use completely consistent grid regions as candidate grid regions. It can be understood that a first spatial consistency index can be calculated between different grid regions and their corresponding sub-regions of the remote sensing image to be processed; thus, multiple first spatial consistency indices can be obtained.

[0110] Therefore, in a further embodiment, it is determined whether there are 100% target grids among all the calculated first spatial consistency indices. That is, it is determined whether there are grid areas that completely match the remote sensing image to be processed. If so, these target grids are taken as candidate grid areas.

[0111] In another optional embodiment, if a 100% target grid does not exist, a second-level strategy can be implemented to ensure sufficient data volume in the image retrieval database. Specifically, missing data in the grid area can be reconstructed to obtain a reconstructed grid area. It is understood that the inability to obtain a 100% target grid may be due to factors such as insufficient pixel values ​​in certain areas of the remote sensing image to be processed. In this case, missing data in the grid area can be reconstructed, i.e., the missing data can be filled in.

[0112] In one optional embodiment, the second-level strategy may involve calculating the real-time spatial relationship between remote sensing images and grid regions using the PostGIS-based ST_Intersects spatial predicate, and selecting grid regions with coverage rates greater than a coverage threshold as candidate grid regions. ST_Intersects is one of the OGC standard spatial predicates used to determine whether two geometric objects have any intersecting parts. Furthermore, in one optional embodiment, the coverage threshold ranges from 0.7 to 1, and the appropriate coverage threshold can be selected according to actual needs.

[0113] Furthermore, the second spatial consistency index between the reconstructed grid region and the remote sensing image to be processed is calculated again, so that regions with a second spatial consistency index greater than the index threshold are considered as candidate grid regions. It should be noted that if no region with a second spatial consistency index greater than the index threshold exists, the current remote sensing image to be processed is discarded to ensure the quality of the image retrieval database.

[0114] In one optional embodiment, an image retrieval database is constructed based on visual features, including:

[0115] Assign a unique spatial identifier to the grid region, and assign a unique image identifier to the image corresponding to the grid region;

[0116] The spatial attribute information of the candidate grid region is obtained; and the visual features are preprocessed to obtain the visual feature vector.

[0117] Based on spatial identifiers and image identifiers, construct the association between visual feature vectors and spatial attribute information;

[0118] Based on visual feature vectors, a visual index based on vector similarity is constructed; based on spatial attribute information, a location index based on spatial topology is constructed; and based on association relationships, a joint index of visual space is constructed to obtain an image retrieval database.

[0119] Currently, at the feature representation and retrieval mechanism level, existing methods mostly employ independent feature extraction techniques, such as extracting texture, color, shape, or spectral features separately, and performing similarity matching based on a single modality. These methods neglect the inherent correlation between visual features and spatial attributes in remote sensing imagery, failing to construct an effective collaborative retrieval mechanism. For example, the same ground feature may exhibit similar spatial distribution patterns and contextual relationships in different images, but relying solely on pixel-level feature matching makes it difficult to capture this high-level semantic information. Furthermore, the lack of comprehensive modeling of spatial topological relationships, orientation, scale variations, and other factors limits the system's generalization ability across modalities and temporal phases.

[0120] Therefore, in order to solve the above-mentioned technical problems, in one optional embodiment, such as Figure 2 As shown, visual features are combined with spatial location features to provide a dual-driven retrieval mechanism based on both visual content semantic similarity and spatial topological correlation.

[0121] Specifically, in specific embodiments, in order to achieve rapid identification and retrieval of grid regions and their corresponding images, a unique spatial identifier (i.e., grid ID) is assigned to each grid region, and a unique image identifier (i.e., image ID) is assigned to the corresponding image. It is understood that, based on the above embodiments, each candidate grid region possesses a unique spatial identifier and a corresponding unique image identifier.

[0122] Simultaneously, based on spatial identifiers, spatial attribute information is extracted from candidate grid regions. This spatial attribute information may include, but is not limited to, grid center coordinates, spatial consistency indices, and region type. In an optional embodiment, the region type can be in the form of a number or text; this application does not limit this. Furthermore, in an optional embodiment, a deep convolutional neural network can be used to extract visual features from the candidate grid regions. Further, after preprocessing such as normalization of the visual features, a visual feature vector is obtained.

[0123] In one alternative embodiment, the deep convolutional neural network can be an improved model based on the ResNet34 deep convolutional neural network, in which the last fully connected layer is replaced with a 512-dimensional feature output layer, and a multimodal feature fusion strategy is adopted to fuse the output features of the conv3_x and conv4_x layers. Here, conv3_x and conv4_x are two convolutional layers of different depths, which can extract visual features at different levels of abstraction (e.g., conv3_x may capture more detailed edges and textures, while conv4_x captures more abstract semantic information).

[0124] In a specific embodiment, pre-trained image features are used to extract the backbone network. The feature output layer is dynamically adapted to the grid size, and its dimensions are adjusted according to the grid spatial resolution. In an optional embodiment, features from intermediate layers of different depths within the neural network can be fused, and channel attention weighting can be applied to the fused features to improve the accuracy of visual feature extraction.

[0125] Furthermore, based on the spatial identifier and image identifier, a correlation between visual feature vectors and spatial attribute information is constructed. Thus, a composite key containing both spatial and image identifiers can be assigned to each candidate grid region, and the correlation between the spatial attribute information and visual feature vectors of the candidate grid regions can be quantitatively stored, for example, in a Milvus vector database.

[0126] It should be noted that, in one optional embodiment, the composite key can be generated by concatenating the grid ID and the image file hash value. That is, the unique identifier of each grid cell in the database is used as a composite primary key, which is composed of two parts concatenated, with a fixed format of grid ID_image hash value.

[0127] like Figure 2 As shown, in a specific embodiment, the obtained image retrieval database includes three dimensions of retrieval methods: visual retrieval, location-based retrieval, and joint retrieval. Therefore, when constructing the image retrieval database, a visual index based on vector similarity is built based on visual feature vectors, and a location index based on spatial topology is built based on spatial attribute information. Simultaneously, a joint index of visual space is built based on association relationships.

[0128] Therefore, the image retrieval database constructed in this application can provide a multi-dimensional retrieval interface. In the joint index interface, it supports combined query conditions based on both grid spatial attributes and visual feature similarity. In an optional embodiment, the combined query conditions may include grid identifier conditions and visual vector similarity threshold conditions connected by logical operators.

[0129] In one alternative embodiment, the visual retrieval interface can perform nearest neighbor search based on feature vectors of visual similarity, while the location retrieval interface can perform query based on grid attribute conditions of spatial location.

[0130] Therefore, the image retrieval database construction method provided in this application embodiment combines the multimodal feature fusion mechanism with the visual feature extraction and spatial topology analysis of deep convolutional networks to achieve dual-dimensional retrieval of content and location, enabling the system to flexibly respond to the query needs of different application scenarios.

[0131] Figure 4This is a schematic diagram illustrating the principle of a method for constructing an image retrieval database according to another embodiment of this application. Before matching the remote sensing image to be processed with the grid area, the method includes:

[0132] Extract image metadata from the remote sensing image to be processed;

[0133] Parse the image metadata; and based on the parsing results, determine whether the remote sensing image to be processed contains coordinate reference system information;

[0134] If included, based on coordinate reference system information, read the spatial geographic information of the remote sensing image to be processed; and proceed to the step of matching the remote sensing image to be processed with the grid area.

[0135] Currently, in terms of multi-source data fusion and coordinate system unification, remote sensing images often originate from different sensor platforms (e.g., optical, radar, and hyperspectral), with varying projection methods, coordinate systems (e.g., WGS84, UTM, and Albers), and georeferenced information. Current processing workflows typically rely on a single, linear coordinate system transformation process, where all images are first uniformly reprojected to a standard coordinate system before subsequent processing. This process is not only computationally expensive, especially in large-scale or high-resolution image processing, but also prone to introducing geometric distortion and interpolation errors during frequent coordinate transformations, affecting spatial positioning accuracy. Furthermore, this workflow lacks support for real-time requirements, making it difficult to meet the needs of applications with high timeliness requirements, such as emergency response and dynamic monitoring.

[0136] Therefore, to solve the above-mentioned technical problems, this application provides a unified processing method for multi-coordinate systems. Specifically, as follows: Figure 4 As shown, in one optional embodiment, the Coordinate Reference System (CRS) information is preferentially parsed from the image metadata. That is, the CRS information used by the image is read and parsed from the metadata of the remote sensing image. The image metadata describes the attributes of the image itself and is usually stored in the image file in text form.

[0137] The extracted image metadata is parsed. If the remote sensing image to be processed already contains CRS information, the spatial geographic information (including but not limited to latitude and longitude) of the remote sensing image to be processed can be directly read based on the CRS information. After the spatial geographic information is read, it can be further matched with the grid area based on the read spatial geographic information.

[0138] Therefore, the image retrieval database construction method provided in this application prioritizes the use of the CRS information inherent in the image itself, avoiding geometric distortion caused during the unified projection transformation process, thereby reducing the accuracy of the image retrieval database.

[0139] Based on the above embodiments, as an optional embodiment, if coordinate reference system information does not exist, the method includes:

[0140] Extract keywords from the filenames of the remote sensing images to be processed;

[0141] Matching is performed in a pre-built coordinate reference system database based on keywords in the filename;

[0142] Based on the matching results, determine whether a matching target coordinate reference system exists in the coordinate reference system database;

[0143] If it exists, based on the target coordinate reference system, read the spatial geographic information and proceed to the step of matching the remote sensing image to be processed with the grid area;

[0144] If it does not exist, the remote sensing image to be processed is converted to the default coordinate reference system, and the process proceeds to the step of matching the remote sensing image to be processed with the grid area.

[0145] It is understandable that in the above embodiments, it may be impossible to obtain the CRS information of the image based on the image metadata. To solve this technical problem, as an optional embodiment based on the above embodiments, the file name keywords of the remote sensing image to be processed are further extracted, and the file name keywords are matched with a pre-built CRS database, which can also be understood as matching with a predefined CRS template.

[0146] It should be noted that the CRS types supported in the CRS database include, but are not limited to, EPSG:4326, EPSG:32651, and CGCS2000. When performing keyword matching in filenames, the corresponding projected coordinate system can be matched when UTM is included, and the geographic coordinate system can be matched when WGS84 is included; this application does not limit this.

[0147] If a matching target CRS exists, the spatial geographic information of the remote sensing image to be processed is read based on the target CRS. If no matching target CRS exists, the remote sensing image to be processed can be converted into a default CRS.

[0148] Therefore, this application provides a three-level coordinate recognition mechanism: image metadata parsing, filename matching, and default CRS transformation. Addressing the heterogeneity issue of multi-source remote sensing data, the multi-level coordinate recognition and two-level spatial matching strategy significantly improve data integration efficiency, while the distributed architecture ensures real-time processing capabilities for large-scale, high-resolution imagery.

[0149] In the above embodiments, the method for constructing an image retrieval database has been described in detail. This application also provides an embodiment of an apparatus for constructing an image retrieval database.

[0150] Figure 5 This is a schematic diagram of the structure of an image retrieval database construction device provided in an embodiment of this application, as shown below. Figure 5 As shown, the device includes:

[0151] The target acquisition module 50 is used to acquire the population density of the area to be processed, the current area type, and the remote sensing image to be processed.

[0152] The spatial grid size determination module 51 is used to determine the current spatial grid size based on population density and current area type; and to divide the area to be processed based on the current spatial grid size to obtain grid areas;

[0153] The grid matching module 52 is used to match the remote sensing image to be processed with the grid area to obtain candidate grid areas whose matching degree meets the preset conditions.

[0154] The visual feature extraction module 53 is used to extract the visual features of the remote sensing image to be processed in the candidate grid area;

[0155] Database building module 54 is used to build an image retrieval database based on visual features.

[0156] Furthermore, the image retrieval database construction apparatus provided in this application embodiment also includes:

[0157] The first acquisition module is used to acquire the current baseline grid size and the pre-built mapping relationship; wherein, the mapping relationship is the correspondence between region type and region weighting factor; the region type represents that the higher the population density, the larger the corresponding region weighting factor;

[0158] The weighting factor determination module is used to determine the weighting factor of the current region based on the mapping relationship and the current region type.

[0159] The size determination module is used to determine the current spatial grid size using the current baseline grid size, the current region weighting factor, and the population density; where the population density is negatively correlated with the current spatial grid size, and the current region weighting factor is positively correlated with the current spatial grid size.

[0160] The baseline grid size determination module is used to determine whether the population density is greater than a density threshold. If it is greater, the first size is used as the current baseline grid size. If it is not greater, the module determines whether the population density is greater than a preset multiple of the density threshold. The preset multiple is less than 1. If it is greater, the second size is used as the current baseline grid size. If it is not greater, the third size is used as the current baseline grid size. Wherein, the first size is smaller than the second size, and the second size is smaller than the third size.

[0161] The first processing module is used to determine the first spatial consistency index between the remote sensing image to be processed and the grid area; and to determine whether there is a target grid with a first spatial consistency index of 100%; if there is, the target grid is used as a candidate grid area; if not, the following steps are performed: reconstructing the missing data in the grid area to obtain the reconstructed grid area; determining the second spatial consistency index between the reconstructed grid area and the remote sensing image to be processed; and using the area with the second spatial consistency index greater than the index threshold as a candidate grid area.

[0162] The identifier allocation module is used to assign a unique spatial identifier to the grid area and a unique image identifier to the image corresponding to the grid area.

[0163] The second acquisition module is used to acquire spatial attribute information of candidate grid regions and preprocess visual features to obtain visual feature vectors.

[0164] The association construction module is used to construct the association between visual feature vectors and spatial attribute information based on spatial identifiers and image identifiers;

[0165] The index building module is used to construct a visual index based on vector similarity based on visual feature vectors; to construct a location index based on spatial topology based on spatial attribute information; and to construct a joint index of visual space based on association relationships, thereby obtaining an image retrieval database.

[0166] The image metadata extraction module is used to extract image metadata from the remote sensing images to be processed.

[0167] The image metadata parsing module is used to parse image metadata; and based on the parsing results, to determine whether the remote sensing image to be processed contains coordinate reference system information; if it does, to read the spatial geographic information of the remote sensing image to be processed based on the coordinate reference system information; and to call the grid matching module.

[0168] The filename keyword extraction module is used to extract keywords from the filenames of remote sensing images to be processed;

[0169] The coordinate system matching module is used to match the file name keywords in a pre-built coordinate reference system database;

[0170] The second processing module is used to determine whether a matching target coordinate reference system exists in the coordinate reference system database based on the matching results; if it exists, it reads spatial geographic information based on the target coordinate reference system and calls the grid matching module; if it does not exist, it converts the remote sensing image to be processed to the default coordinate reference system and calls the grid matching module.

[0171] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 6 As shown, the electronic device includes: a memory 60 for storing computer programs;

[0172] The processor 61 is configured to execute a computer program to implement the steps of the image retrieval database construction method as described in the above embodiments.

[0173] The electronic devices provided in this embodiment may include, but are not limited to, laptops or desktop computers.

[0174] The processor 61 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 61 may be implemented using at least one of the following hardware forms: Digital Signal Processor (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 61 may also include a main processor and a coprocessor. The main processor, also known as the Central Processing Unit (CPU), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 61 may integrate a Graphics Processing Unit (GPU), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 61 may also include an Artificial Intelligence (AI) processor, which is used to handle computational operations related to machine learning.

[0175] The memory 60 may include one or more computer-readable storage media, which may be non-transitory. The memory 60 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In this embodiment, the memory 60 is used to store at least the following computer program 601, which, after being loaded and executed by the processor 61, is capable of implementing the relevant steps of the image retrieval database construction method disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 60 may also include an operating system 602 and data 603, etc., and the storage method may be temporary storage or permanent storage. The operating system 602 may include Windows, Unix, Linux, etc. The data 603 may include, but is not limited to, the relevant data involved in the image retrieval database construction method.

[0176] In some embodiments, the electronic device may further include a display screen 62, an input / output interface 63, a communication interface 64, a power supply 65, and a communication bus 66.

[0177] Those skilled in the art will understand that Figure 2 The structures shown do not constitute a limitation on electronic devices and may include more or fewer components than those shown.

[0178] The electronic device provided in this application includes a memory and a processor. When the processor executes the program stored in the memory, it can implement the image retrieval database construction method in the above embodiments.

[0179] It should be noted that although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

Claims

1. A method for constructing an image retrieval database, characterized in that, The method includes: Acquire the population density, current area type, and remote sensing image of the area to be processed; Based on the population density and the current region type, determine the current spatial grid size; and based on the current spatial grid size, divide the region to be processed to obtain grid regions; The remote sensing image to be processed and the grid area are matched to obtain candidate grid areas that meet the matching degree preset conditions; the matching degree is a spatial consistency index, and the spatial consistency index includes area coverage and latitude and longitude consistency; the spatial consistency index is used to reflect the spatial consistency between the remote sensing image to be processed and the grid area, and the higher the matching degree, the higher the spatial consistency. Extract the visual features of the remote sensing image to be processed from the candidate grid region; Based on the aforementioned visual features, an image retrieval database is constructed; Obtain the current baseline grid size and the pre-constructed mapping relationship; wherein, the mapping relationship is the correspondence between region type and region weighting factor; the region type represents that the higher the population density, the larger the corresponding region weighting factor; Based on the mapping relationship and the current region type, determine the weighting factor for the current region; The current spatial grid size is determined by the current baseline grid size, the current region weighting factor, and the population density; wherein the population density is negatively correlated with the current spatial grid size, and the current region weighting factor is positively correlated with the current spatial grid size.

2. The method for constructing an image retrieval database as described in claim 1, characterized in that, Obtain the current baseline mesh size, including: Determine whether the population density is greater than a density threshold; If it is greater than the given size, the first size will be used as the current reference mesh size. If not greater than, determine whether the population density is greater than the density threshold of a preset multiple; The preset multiple is less than 1; if it is greater than 1, the second size is used as the current reference grid size. If it is not greater than, the third dimension is used as the current reference grid size; wherein, the first dimension is smaller than the second dimension, and the second dimension is smaller than the third dimension.

3. The method for constructing an image retrieval database as described in claim 1, characterized in that, The step of matching the remote sensing image to be processed with the grid region to obtain candidate grid regions whose matching degree meets preset conditions includes: Determine a first spatial consistency index between the remote sensing image to be processed and the grid region; and determine whether there is a target grid with a first spatial consistency index of 100%. If it exists, the target mesh will be used as the candidate mesh region; If it does not exist, proceed with the following steps: The missing data in the grid region is reconstructed to obtain the reconstructed grid region; A second spatial consistency index is determined between the reconstructed grid region and the remote sensing image to be processed; and regions with the second spatial consistency index greater than the index threshold are selected as candidate grid regions.

4. The method for constructing an image retrieval database as described in claim 1, characterized in that, The construction of the image retrieval database based on the visual features includes: Assign a unique spatial identifier to the grid region, and assign a unique image identifier to the image corresponding to the grid region; Obtain the spatial attribute information of the candidate grid region; and preprocess the visual features to obtain a visual feature vector; Based on the spatial identifier and the image identifier, construct the association between the visual feature vector and the spatial attribute information; Based on the visual feature vectors, a visual index based on vector similarity is constructed; based on the spatial attribute information, a location index based on spatial topology is constructed; and based on the association relationship, a joint index of visual space is constructed to obtain the image retrieval database.

5. The method for constructing an image retrieval database as described in claim 1, characterized in that, Before matching the remote sensing image to be processed with the grid region, the method includes: Extract image metadata from the remote sensing image to be processed; The image metadata is parsed; and based on the parsing results, it is determined whether the remote sensing image to be processed contains coordinate reference system information. If included, based on the coordinate reference system information, read the spatial geographic information of the remote sensing image to be processed; and proceed to the step of matching the remote sensing image to be processed with the grid area.

6. The method for constructing an image retrieval database as described in claim 5, characterized in that, If the coordinate reference system information does not exist, the method includes: Extract keywords from the filenames of the remote sensing images to be processed; Matching is performed in a pre-built coordinate reference system database based on the keywords in the file name; Based on the matching results, determine whether a matching target coordinate reference system exists in the coordinate reference system database; If it exists, based on the target coordinate reference system, read the spatial geographic information and proceed to the step of matching the remote sensing image to be processed with the grid area; If it does not exist, the remote sensing image to be processed is converted to the default coordinate reference system, and the process proceeds to the step of matching the remote sensing image to be processed with the grid region.

7. An apparatus for constructing an image retrieval database, characterized in that, The device includes: The target acquisition module is used to acquire the population density, current area type, and remote sensing image of the area to be processed. The spatial grid size determination module is used to determine the current spatial grid size based on the population density and the current region type; and to divide the region to be processed based on the current spatial grid size to obtain a grid region; The grid matching module is used to match the remote sensing image to be processed with the grid area to obtain candidate grid areas whose matching degree meets preset conditions; the matching degree is a spatial consistency index, and the spatial consistency index includes area coverage and latitude and longitude consistency; the spatial consistency index is used to reflect the spatial consistency between the remote sensing image to be processed and the grid area, and the higher the matching degree, the higher the spatial consistency. A visual feature extraction module is used to extract the visual features of the remote sensing image to be processed in the candidate grid region; A database construction module is used to construct an image retrieval database based on the visual features; The first acquisition module is used to acquire the current baseline grid size and the pre-constructed mapping relationship; wherein, the mapping relationship is the correspondence between region type and region weighting factor; the region type represents that the higher the population density, the larger the corresponding region weighting factor; The weighting factor determination module is used to determine the weighting factor of the current region based on the mapping relationship and the current region type. The size determination module is used to determine the current spatial grid size using the current reference grid size, the current region weighting factor, and the population density; wherein the population density is negatively correlated with the current spatial grid size, and the current region weighting factor is positively correlated with the current spatial grid size.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for constructing the image retrieval database according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method for constructing the image retrieval database according to any one of claims 1 to 6.

Images