A map resource recognition method based on deep learning

Abstract

The application provides a map resource recognition method based on deep learning, comprising collecting map resource image samples and labeling the map resource image samples to form a map image dataset; based on a set of model training parameters, training samples are obtained from the map image dataset in batches; the training samples obtained in batches are subjected to data enhancement to generate enhanced training samples in batches; based on the enhanced training samples in batches, a constructed deep learning model is trained until a map resource recognition model is obtained; a map resource image to be recognized is obtained and input into the map resource recognition model for map resource detection and recognition. The technical solution in the application significantly improves the accuracy, the degree of dependence on data, and the adaptability to complex scenes, and efficiently solves the technical problems existing in the existing map resource recognition method.

Description

Technical Field

[0001] This invention relates to the field of intelligent processing technology, and more specifically to a map resource identification method based on deep learning. Background Technology

[0002] In the field of map resource processing, practitioners have long been committed to finding accurate and efficient methods for map resource identification. Early on, rule-based identification techniques dominated. This method relied on manually writing complex identification rules based on map features, such as matching rules based on map color, line style, and symbol shape. However, real-world map styles are extremely diverse, with significant differences in representation between maps from different cartographic institutions and for different purposes. This makes it difficult for manually written rules to comprehensively cover all types of map features, resulting in poor identification accuracy. Furthermore, when new features or style changes occur on maps, a significant amount of manpower and time is required to revise and improve the rules, which is not only inefficient but also costly.

[0003] With the evolution of technology, map resource identification methods based on traditional machine learning have emerged. These methods utilize classifiers, such as Support Vector Machines (SVM) and decision trees, to identify map data. Compared to rule-based methods, they alleviate the burden of manually writing rules to some extent. However, these methods heavily rely on a large amount of high-quality labeled data to train the model. In practical applications, collecting sufficient and accurately labeled map data is a challenging task. On the one hand, labeling requires a significant investment of time and effort from professionals and is prone to errors due to human factors; on the other hand, the labeling standards for different types of map data are difficult to standardize, further increasing the difficulty. Furthermore, traditional machine learning models have inherent limitations in feature extraction. They primarily rely on manually designed feature extraction algorithms, which typically only capture relatively obvious and simple features in maps. They struggle to effectively extract and analyze subtle, hidden, and interconnected features in complex map scenes, significantly limiting the model's performance in complex map resource identification tasks.

[0004] In summary, existing map resource identification methods have significant shortcomings in terms of accuracy, efficiency, and adaptability to complex map scenarios. There is an urgent need for an innovative, more efficient, and accurate map resource identification technology to meet the growing demand for map data processing. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a deep learning-based map resource identification method to at least solve or alleviate the problems existing in the prior art.

[0006] To achieve the above objectives, according to one aspect of this application, a deep learning-based map resource identification method is provided, comprising:

[0007] Collect map resource image samples and annotate the map resource image samples to form a map image dataset;

[0008] Based on the set model training parameters, training samples are obtained in batches from the map image dataset;

[0009] Data augmentation is performed on the training samples acquired in batches to generate augmented training samples in batches.

[0010] Based on the batch-wise enhanced training samples, the constructed deep learning model is trained until the training is completed and the map resource recognition model is obtained.

[0011] The map resource image to be identified is acquired and input into the map resource identification model for map resource detection and identification.

[0012] The technical solution in this application has at least the following technical advantages:

[0013] ① Traditional rule-based methods struggle to fully cover features due to the diversity of map styles. In contrast, this method collects and labels a large number of map resource image samples to form a dataset, enabling the model to learn a rich variety of map features, greatly improving recognition accuracy and overcoming the limitations of rule writing.

[0014] ② Data augmentation plays a crucial role in addressing the challenges of traditional machine learning's reliance on large amounts of high-quality labeled data and the difficulty of labeling. By augmenting training samples acquired in batches, the diversity of training data is expanded without significantly increasing the actual labeling workload, reducing reliance on massive amounts of accurately labeled data, and mitigating the impact of human labeling errors.

[0015] ③ In terms of feature extraction, traditional machine learning, relying on manually designed algorithms, struggles to capture complex and hidden features. Deep learning models, however, can automatically learn various features in maps, including subtle, hidden, and interconnected features, effectively improving adaptability to complex map scenes and thus performing better in complex map resource identification tasks. Overall, this method achieves significant improvements in accuracy, data dependence, and adaptability to complex scenes, efficiently solving the technical challenges of existing map resource identification methods. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a deep learning-based map resource identification method according to an embodiment of this application. Detailed Implementation

[0017] Figure 1 This is a schematic diagram illustrating a deep learning-based map resource identification method according to an embodiment of this application. Figure 1 As shown, it includes:

[0018] Collect map resource image samples and annotate the map resource image samples to form a map image dataset;

[0019] Based on the set model training parameters, training samples are obtained in batches from the map image dataset;

[0020] Data augmentation is performed on the training samples acquired in batches to generate augmented training samples in batches.

[0021] Based on the batch-wise enhanced training samples, the constructed deep learning model is trained until the training is completed and the map resource recognition model is obtained.

[0022] The map resource image to be identified is acquired and input into the map resource identification model for map resource detection and identification.

[0023] Preferably, in one embodiment, when collecting map resource image samples and labeling them to form a dataset, the number of collected map resource image samples is N, and each image sample can be represented as a three-dimensional matrix. Where i = 1, 2, ..., N, H represents the height of the image, W represents the width of the image, and C represents the number of channels of the image (for example, C = 3 for an RGB image).

[0024] For each image sample I i Labeling is performed, and the labeling information can be represented as a vector. Where K is the number of map resource categories. i Each element y in i,j Image I i The existence status of the j-th type of map resource (e.g., 1 for existence, 0 for non-existence).

[0025] Then map image dataset It can be represented as:

[0026] Preferably, in one embodiment, when acquiring training samples in batches, model training parameters are set, including batch size B and number of training epochs E. In each training epoch, the dataset is... The training samples are divided into multiple batches. Let the training sample set of the b-th batch in the e-th round of training be... but: Where e = 1, 2, ..., E,

[0027] Preferably, in one embodiment, during data augmentation, data augmentation is performed on the training samples acquired in batches, and the data augmentation function is set to... It can perform operations such as rotating, flipping, and scaling images. For batch processing... Each sample (I) k ,y k After data augmentation, the augmented sample (I) is obtained. k′ ,y k′ ),in y k′ =y k (Because the annotation information does not change with the geometric transformation of the image).

[0028] The batch-wise augmented training sample set is:

[0029] Preferably, in one embodiment, when training the deep learning model, the constructed deep learning model is denoted as f(I; θ), where I is the input image and θ are the model parameters. In each round of training, for each batch... Calculate the loss function of the model. A commonly used loss function is the cross-entropy loss function: Among them, f j (I k′ ;θ) represents the model f(I k′ The probability of the j-th class is output by θ.

[0030] Update the model parameters θ using optimization algorithms (such as stochastic gradient descent, SGD):

[0031] Where α is the learning rate. The loss function L(θ) with respect to the parameter θ old The gradient.

[0032] Repeat the above process until the number of training rounds reaches E, at which point the trained map resource recognition model f(I; θ) is obtained. * ), where θ * These are the final model parameters.

[0033] Preferably, in one embodiment, during map resource detection and identification, the map resource image I to be identified is acquired. test It is then input into the map resource recognition model f(I; θ) * In the process, the output y is obtained. test =f(I test ;θ * ). y test Each element y in test,j Image I testThe predicted probability of the j-th type of map resource is calculated. Based on the predicted probability, map resource detection and identification can be performed; for example, the category with the highest probability can be selected as the final identification result.

[0034] Preferably, in a specific application scenario, the solution provided by the above preferred embodiments can achieve the following technical advantages:

[0035] 1. Traditional rule-based map resource identification methods struggle to handle diverse map styles because they rely on limited, manually summarized rules that fail to cover all map features. While traditional machine learning methods can learn data features to some extent, their ability to represent and utilize data is limited when dealing with complex map images, and insufficient manually labeled data severely impacts model performance. This application represents map image samples as three-dimensional matrices using a formula. The annotation information is represented as a vector This data representation can completely and accurately describe map images and their corresponding resource categories. This data representation provides rich input for deep learning models, enabling them to learn various complex features and patterns in map images, thereby improving the accuracy and robustness of recognition. Simultaneously, representing the dataset as... This facilitates subsequent batch processing and model training.

[0036] 2. Traditional methods are inefficient when processing large-scale data, as they cannot process all data at once due to memory limitations. Furthermore, traditional methods lack dynamic utilization and learning capabilities for the data, making them difficult to adapt to the diversity and variability of the data. This application divides the dataset into multiple batches by setting the batch size B and the number of training epochs E. Training is then performed. This batch training method effectively reduces memory usage and improves training efficiency. Simultaneously, by iterating through different batches in each training round, the model learns from different perspectives of the data, enhancing its generalization ability and enabling it to better adapt to various map images.

[0037] 3. Traditional machine learning methods heavily rely on large amounts of high-quality labeled data. However, in practical applications, obtaining sufficient and accurately labeled map data is very difficult, costly, and prone to human error. This application introduces data augmentation functions. Augmentation processing is performed on the training samples to generate augmented samples. Data augmentation can expand the diversity of training data without increasing the actual annotation workload, allowing the model to learn more map image features of different forms. For example, by rotating, flipping, and scaling, map images at different viewpoints and scales can be simulated, improving the model's adaptability to various map images, reducing dependence on large amounts of labeled data, and mitigating the impact of human annotation errors.

[0038] 4. Traditional rule-based methods lack the ability to learn and adapt to data, making it difficult to cope with changes in map styles. Traditional machine learning methods have limited feature extraction capabilities, and their optimization processes are relatively simple, making it difficult to find optimal model parameters. This application uses a deep learning model f(I;θ) for training and employs a cross-entropy loss function L(θ) to measure the difference between the model's predictions and the actual annotations. This loss function-based optimization approach allows the model to automatically learn complex features and patterns in map images, improving recognition accuracy. Simultaneously, using optimization algorithms such as stochastic gradient descent (SGD) to update the model parameters θ enables efficient finding of optimal model parameters on large-scale datasets, allowing the model to continuously optimize and improve, adapting to different map resource recognition tasks.

[0039] 5. Traditional methods often suffer from poor recognition performance when faced with complex map scenes, making it difficult to accurately detect and identify various resources within the map. This application addresses this issue by using the map image to be identified... test The input is fed into the trained map resource recognition model f(I; θ) * In the process, the predicted result y is obtained. test Deep learning models can automatically extract complex features from map images and detect and identify map resources based on predicted probabilities, exhibiting higher accuracy and reliability in complex map scenarios. For example, they can more accurately identify different types of geographic markers, buildings, and other resources on maps.

[0040] Optionally, map resource image samples are collected, including:

[0041] Based on the established spatial distribution strategy and shooting specifications, map resource image samples are collected from the map resource image sample library. The spatial distribution strategy includes: uniform distribution rules of sample points for different land cover regions. The shooting specifications include: the number of sample points on a single map resource image sample.

[0042] Preferably, in one embodiment, when the spatial distribution strategy and shooting specifications are expressed mathematically in detail, it is assumed that there are M different types of land parcel regions, labeled as T1, T2, ..., T M For each land cover type, the area T is... m (m=1,2,…,M) has a complex geometric shape in geographic space, and its area A m It can be calculated using integration. Assume the land parcel region T... m Boundary curves on a two-dimensional plane can be represented by the parametric equation r. m (t)=(x m (t),y m (t)), t∈[a m ,b mIf ] represents the area, then its area is:

[0043] For each type of land parcel region T m Based on its importance and data requirements, a weighting coefficient ω is set. m ,and The total number of sample points to be collected is N total Then the land parcel region T m Number of sample points n to be collected m for:

[0044] To achieve a more accurate and uniform distribution of sample points within each land cover region, a non-uniform density function ρ is introduced. m (x,y) is used to describe the land cover region T. m The density of the distribution of internal sampling points. (This refers to the region T.) m Divided into k m Each small grid has a ΔR value. m,l (l=1,2,…,k m The area of ​​) is ΔA m,l In the small grid ΔR m,l Number of samples placed inside n m,l for:

[0045]

[0046] Preferably, in one embodiment, when setting the shooting specifications, the number of sample points on a single map resource image sample is set to S, while considering the type distribution of the sample points. Let the expected number of sample points of the s-th type in a single image sample be S. s ,and

[0047] Preferably, in one embodiment, when collecting samples from the map resource image sample library, the total number of images in the map resource image sample library is N, and each image can be represented as a three-dimensional matrix. Where i = 1, 2, ..., N, H represents the height of the image, W represents the width of the image, and C represents the number of channels of the image (e.g., C = 3 for an RGB image). Let P be the number of map resource image samples collected, and the collected sample set be... Where s j (j = 1, 2, ..., P) is the index of the image selected from the sample library.

[0048] For each image I i Define a sample feature vector v i =(v i,1 ,v i,2 ,…,v i,M ), where vi,m Image I i Central landform patch region T m The number of sample points. Simultaneously, define a sample point type distribution vector u. i =(u i,1 ,u i,2 ,…,u i,M ), where u i,s Image I i The number of samples of type s in the sample.

[0049] Introducing a comprehensive evaluation function To determine image I i Whether the data collection requirements are met, this function takes into account the number of sample points, the distribution of sample point types, and the spatial uniformity of the sample points.

[0050] First, define the sample number deviation function D. num (v i ):

[0051] Define the sample type distribution deviation function D type (u i ):

[0052] Define the spatial distribution uniformity function D of the sample points dist (I i This is measured by the difference between the actual distribution of sample points in an image and the ideal uniform distribution. It is assumed that the coordinates of the sample points under the ideal uniform distribution are... The coordinates of the actual sample points are but:

[0053]

[0054] Comprehensive evaluation function for: Where α, β, γ are weighting coefficients, and α + β + γ = 1.

[0055] Select function Defined as: Here, ∈ is a pre-set threshold.

[0056] In the above preferred embodiments, the actual physical meaning of each parameter is explained as follows: M: Number of land parcel types. T m : The m-th land cover region, m = 1, 2, ..., M. A m ω represents the area of ​​the m-th land cover region, calculated through integration. m : Weight coefficient of the m-th land parcel region. N total The total number of sample points to be collected. n mρ: The number of sample points to be collected in the m-th land cover area. m (x,y): The non-uniform density function of the sample point distribution within the m-th land type patch area. k m ΔR: The number of small grids used to divide the m-th land cover region. m,l ΔA represents the l-th subgrid within the m-th land cover region. m,l : The area of ​​the l-th small grid in the m-th land cover region. n m,l : The number of sample points placed in the l-th small grid of the m-th land cover region. S: The number of sample points on a single map resource image sample. s : The expected number of type s samples in a single image sample. N: The total number of images in the map resource image sample library. P: The number of map resource image samples collected. s j : The index of the j-th acquired image in the sample database, j = 1, 2, ..., P. i Image I i The feature vector of the sample points. v i,m Image I i Central landform patch region T m The number of sample points. i Image I i The sample type distribution vector. u i,s Image I i The number of samples of type s. α, β, γ: Weighting coefficients of the comprehensive evaluation function. ∈: Threshold of the selection function. The i-th image in the map resource image sample library. A collection of map resource image samples collected. The selection function is used to determine the image I. i Does it meet the data collection requirements? A comprehensive evaluation function is used to evaluate image I. i The number, type distribution, and spatial distribution uniformity of the sample points. D num (v i ): Sample quantity deviation function. D type (u i ): Sample type distribution deviation function. D dist (I i ): Spatial uniformity function of sample points.

[0057] Preferably, based on the scheme provided in the above embodiments, during data collection, each image I in the map resource image sample library is traversed. i For each image, first calculate its sample feature vector v. i and sample type distribution vector u i Then, calculate the sample number deviation function D according to the above formula.num (v i ), Sample type distribution deviation function D type (u i ) and the spatial distribution uniformity function D of the sample points dist (I i Next, the comprehensive evaluation function is calculated. Finally, based on the selection function Determine if the image meets the criteria. If it does, add the image to the collected sample set. Finally, the collected sample set is output.

[0058] Preferably, in a specific application scenario, the technical solution provided by the above embodiments can achieve the following technical benefits.

[0059] 1. Traditional map resource image sample collection simply involves random sampling based on approximate regional divisions, without fully considering the area, importance, and precise distribution of sample points across different land use types. This results in uneven distribution of collected samples across different land use types, failing to accurately reflect the actual geographical situation and thus affecting the accuracy of subsequent map resource identification models. This application calculates the area A of land use type patch regions through integration. m And introduce a weighting coefficient ω m To determine the number of sample points (n) needed to be collected for different land cover regions. m This method allows for more precise allocation of sample points based on the actual conditions of different land types, resulting in more representative samples. Furthermore, this application introduces a non-uniform density function ρ. m The distribution density of sample points within each land cover type region is described by (x, y). This takes into account the complexity within each land cover type, where different regions have different feature densities. In this way, more sample points can be collected in feature-rich areas, improving the coverage of regional features and helping the model learn more comprehensive information.

[0060] 2. Traditional data acquisition methods only focus on the total number of sample points, ignoring the distribution of sample point types. This results in an excessive number of sample points of some types and an insufficient number of sample points of others, leading to an imbalance in the model's learning of different types of map resources during training, affecting the accuracy and generalization ability of recognition. This application specifies the expected number S of the s-th type of sample points in a single image sample. s Furthermore, strict control is exercised during the data collection process. This ensures that the collected samples are balanced in terms of sample point type, enabling the model to learn the characteristics of various types of map resources more comprehensively and improving the model's ability to identify different types of map resources.

[0061] 3. Traditional sample selection is based on simple rules, such as whether the number of sample points meets a certain range, without comprehensively considering the type distribution and spatial uniformity of the sample points. This leads to unreasonable spatial distribution or uneven type distribution of the collected samples, affecting the model's performance. This application introduces a comprehensive evaluation function... The sample number deviation function D was taken into account. num (v i ), Sample type distribution deviation function D type (u i ) and the spatial distribution uniformity function D of the sample points dist (I i This comprehensive evaluation method can fully assess whether each image sample meets the acquisition requirements, ensuring that the acquired samples have good quality in terms of sample quantity, type distribution, and spatial distribution. Additionally, the selection function... Based on the results of the comprehensive evaluation function, a pre-set threshold is used to determine whether an image is selected. This precise filtering mechanism can eliminate unqualified samples, improve the quality of collected samples, and provide higher-quality data for subsequent map resource recognition models, thereby enhancing the accuracy and stability of the models.

[0062] 4. Traditional sample collection methods, lacking scientific strategies and precise evaluation mechanisms, result in inconsistent sample data quality, containing significant noise and unrepresentative data. This makes it difficult for map resource recognition models to learn effective features during training, thus affecting model performance and generalization ability. This application, through the aforementioned innovative spatial distribution strategy, shooting specifications, and sample collection selection function, can collect more representative, type-balanced, and spatially uniform sample data. This high-quality data provides richer and more accurate information for the map resource recognition model, enabling it to better learn the features and patterns of map resources, thereby improving the model's recognition accuracy, generalization ability, and stability. Simultaneously, it reduces the risk of overfitting during model training, improving the model's performance in practical applications.

[0063] Optionally, the map resource image samples are labeled, including:

[0064] The labeled patches on the map resource image sample are identified, and their geographic coordinates are associated to determine the latitude and longitude and land use code corresponding to the patches, and a patch attribute table record is generated.

[0065] Preferably, in one embodiment, when determining the mathematical model that associates labeled patches with geographic coordinates, the map resource image sample set is assumed to be... in Let H represent the i-th image sample, H be the image height, W be the image width, and C be the number of image channels (e.g., C = 3 for RGB images).

[0066] For each image sample I i Image segmentation algorithms (such as deep learning-based semantic segmentation models) are used to segment the image into multiple regions, which are then labeled as image patches. Let image I... i The set of labeled patches obtained by mid-segmentation is Where M i Image I i The number of marked patches.

[0067] There exists a geographic reference system (such as WGS-84), and a projection transformation relationship exists between the map resource image samples and this geographic reference system. Let image I... i The geographic coordinates (latitude and longitude) corresponding to the pixel coordinates (x, y) are: The transformation can be achieved using a projection transformation function T:

[0068]

[0069] For the labeled map features Its boundary consists of a series of pixel coordinate points Composition, in which L ij It is image P ij The number of boundary pixels. The image patch P can be obtained through the projection transformation function T. ij Geographic coordinates of the boundary

[0070]

[0071] Preferably, in one embodiment, when determining the latitude, longitude, and land use code corresponding to a patch, in order to determine patch P... ij The corresponding latitude and longitude can be used to calculate the centroid of its boundary geographic coordinates. Let map patch P... ij The centroid latitude and longitude are The calculation formula is as follows:

[0072]

[0073] Let the set of land class codes be For each labeled map patch P ij The land parcel P is classified using a classifier (such as a machine learning-based classification model) to determine its land use code. Let the classification function be F, then the land parcel P... ij land category code c ij c ij =F(P ij ).

[0074] Preferably, in one embodiment, when generating the patch attribute table record, the patch attribute table record can be represented by a tuple, for image I i The labeled patch P in ij Its attribute table records R ij for: Where i represents the image sample number, and j represents the patch number in the image. and These are the latitude and longitude of the centroid of the image patch, c ij This is the land class code for the map patch.

[0075] Collection of attribute table records for all image samples for:

[0076]

[0077] In the preferred embodiments described above, the physical meanings of each parameter are explained as follows: N: Number of map resource image samples. H: Height of the image sample. W: Width of the image sample. C: Number of channels of the image sample. M i Image I i The number of marked polygons. L ij Image P ij Number of boundary pixels. K: Number of land class codes. I i : The i-th map resource image sample. Image I i The set of labeled polygons in the image. T: Projection transformation function, converting image pixel coordinates to geographic coordinates. F: Classification function, used to determine the land use code of the labeled polygons. R ij Image I i The attribute table record of the j-th labeled polygon. A collection of attribute records for all image samples.

[0078] Specifically, when executed in an application scenario, it reads a set of map resource image samples. For each image sample I i Image segmentation algorithms are used to determine the set of labeled patches. Then, for each labeled patch P ij The boundary pixel coordinates are converted to geographic coordinates using the projection transformation function T, and the centroid latitude and longitude are calculated. Next, the land use code c of the map patch is determined using the classification function F. ij Finally, a map attribute table is generated to record R. ij and add to the collection After processing, the output is a set of attribute table records for the map features.

[0079] Preferably, the technical solution provided by the above-mentioned preferred embodiments has the following technical advantages:

[0080] 1. Traditional methods for determining labeled map features often rely on manual drawing, which is not only inefficient but also susceptible to human error, leading to significant errors and inconsistencies in the labeling results. For complex map images, manual labeling is even more difficult, and accuracy is hard to guarantee. This application utilizes image segmentation algorithms (such as deep learning-based semantic segmentation models) to automatically segment images into labeled features, improving both efficiency and accuracy. This method can quickly process large numbers of map images and learn complex features within the images, resulting in more precise and consistent labeling. For example, when processing map images containing multiple land features, the deep learning model can accurately identify the boundaries of different land features, thus obtaining more reasonable labeled features.

[0081] 2. Traditional geographic coordinate association methods require complex manual measurements and calculations, which are not only cumbersome but also prone to errors. Furthermore, traditional methods struggle to handle large-scale map image data, failing to meet the demands for efficient annotation. This application introduces a projection transformation function T to automatically convert image pixel coordinates to geographic coordinates, significantly improving the efficiency and accuracy of geographic coordinate association. It can quickly calculate the geographic coordinates of each labeled patch boundary and handle large-scale map image data. This automated processing method reduces manual intervention, lowers the probability of errors, and makes the geographic coordinates of labeled patches more accurate.

[0082] 3. Traditional methods determine the latitude and longitude of map features through simple estimation or manual measurement. This method is not precise enough and cannot accurately reflect the geographical location of the map features. This application determines the latitude and longitude of map features by calculating the centroid of the geographic coordinate points at the map feature boundary. This method is more scientific and accurate. Centroid calculation comprehensively considers all geographic coordinate points at the map feature boundary, making the determined latitude and longitude more accurately represent the geographical location of the map feature. In subsequent map analysis and applications, accurate latitude and longitude information can improve the accuracy and reliability of the analysis results.

[0083] 4. Traditional land use code determination typically relies on human experience, which is not only inefficient but also susceptible to subjective influences, leading to inaccurate classification. This application uses a classifier (such as a machine learning-based classification model) to determine the land use codes of land parcels, improving both accuracy and efficiency. Machine learning models can learn the features of a large number of land parcels and their corresponding land use codes, thus accurately classifying new land parcels. This method reduces the subjectivity of human judgment and improves the consistency and reliability of land use code classification.

[0084] 5. Traditional methods of generating map feature attribute table records require manual data entry, which is not only inefficient but also prone to errors. This application addresses this by automatically generating map feature attribute table records R. ij And stored in a collection This system enables automated data management and storage. It can quickly generate accurate attribute table records for map features and facilitate subsequent data retrieval and analysis. This automated processing method improves the efficiency and accuracy of data processing and reduces errors from manual operations.

[0085] Optionally, the method further includes: estimating high-frequency and low-frequency textures based on multi-scale target and pixel segmentation boundaries identified from map resources; and determining the model training parameters based on the high-frequency and low-frequency textures.

[0086] Preferably, in one embodiment, during multi-scale target and pixel segmentation boundary analysis, the map resource image is assumed to be... Where H is the image height, W is the image width, and C is the number of image channels. Multi-scale feature extraction is performed on the image using convolutional kernels or feature extractors of different scales. Let the scale levels be s = 1, 2, ..., S. The feature map extracted at scale s is... Where H s W s These are the height and width of the feature map at scale s, respectively, D s It is the number of channels in the feature map.

[0087] The image is segmented using a pixel segmentation algorithm (such as semantic segmentation networks like U-Net) to obtain the pixel segmentation results. Each pixel value represents the category to which that pixel belongs. The segmentation boundary can be obtained by performing edge detection on the segmentation results. Let the segmentation boundary map be... The boundary pixel value is 1, and the non-boundary pixel value is 0.

[0088] Preferably, in one embodiment, when estimating high-frequency and low-frequency textures, image I is transformed to the frequency domain using a two-dimensional discrete Fourier transform (2D-DFT):

[0089] Where (u,v) are frequency domain coordinates, F I (u,v) is the frequency domain representation of image I.

[0090] Define a cutoff frequency f c The frequency domain is divided into low-frequency and high-frequency components. The low-frequency component F... L (u,v) and the high-frequency component F H (u,v) are respectively:

[0091]

[0092] Calculate the low-frequency texture energy E L and high-frequency texture energy E H :

[0093] Preferably, in one embodiment, when determining model training parameters based on texture,

[0094] Let the model training parameters include the learning rate α, batch size B, number of training epochs E, etc., and represent these parameters as a parameter vector θ = (α, B, E).

[0095] Define a parameter adjustment function G based on the low-frequency texture energy E. L and high-frequency texture energy E H To adjust the model training parameters:

[0096] θ′=G(E L E H ,θ).

[0097] Specifically, the learning rate α can be adjusted according to the following rules: Where β is an adjustment coefficient.

[0098] Batch size B and number of training epochs E can also be adjusted similarly based on texture energy, for example:

[0099] Where γ and δ are adjustment coefficients.

[0100] In the above preferred scheme, the physical meanings of each parameter are explained as follows: H: Image height. W: Image width. C: Number of image channels. S: Number of scale levels for multi-scale feature extraction. s W s : Height and width of the feature map at scale s. D s f: Number of channels in the feature map at scale s. c Frequency domain cutoff frequency. E L Low-frequency texture energy. E H : High-frequency texture energy. α: Initial learning rate. B: Initial batch size. E: Initial number of training epochs. β, γ, δ: Adjustment coefficients. F s : Feature map extracted at scale s. S: Pixel segmentation result. B: Segmentation boundary map. F I (u,v): Frequency domain representation of image I. F L (u,v): Low-frequency domain portion. F H (u,v): High-frequency domain component. G: Parameter adjustment function, which adjusts the model training parameters based on texture energy.

[0101] Preferably, when the above scheme is executed, the map resource image I is read, and multi-scale feature extraction is performed to obtain F. s Simultaneously, pixel segmentation is performed to obtain S and B. Then, image I is transformed into the frequency domain, and the low-frequency texture energy E is calculated. L and high-frequency texture energy E H Finally, the parameter adjustment function G is used according to E. L and E H Adjust the model training parameters θ and output the adjusted parameters θ′.

[0102] Preferably, in a specific application scenario, the above-mentioned preferred solution has the following technical advantages:

[0103] 1. Traditional map resource identification methods often extract features only at a single scale, making it difficult to capture the feature information of targets of different sizes on the map. Furthermore, when determining target boundaries, they employ relatively simple edge detection algorithms, which have poor accuracy in segmenting boundaries for complex map scenes, failing to accurately distinguish different map resources. This application utilizes convolutional kernels or feature extractors of different scales to perform multi-scale feature extraction on map resource images, obtaining feature maps F at different scales. s This approach comprehensively captures the features of targets of different sizes on the map. For example, at a small scale, it can focus on detailed features, while at a large scale, it can grasp the overall layout information. This allows the model to learn richer and more comprehensive features during training, improving its ability to recognize map resources of different scales. Pixel segmentation algorithms (such as semantic segmentation networks like U-Net) are used to obtain pixel segmentation results S, and edge detection is used to determine the segmentation boundary map B. Compared to traditional edge detection algorithms, these semantic segmentation networks can combine contextual information to more accurately delineate the boundaries of different map resources, providing a more precise foundation for subsequent texture analysis and model training.

[0104] 2. Traditional methods for processing map images rarely conduct in-depth analysis of image texture, or only use simple statistical methods to describe texture features, failing to fully extract the texture information within the image. However, texture information in maps is crucial for distinguishing different types of map resources (such as urban areas, farmland, and forests). This application transforms the map image into the frequency domain, obtaining the frequency domain representation F of the image through a two-dimensional discrete Fourier transform. I (u,v), and based on the cutoff frequency f c Divide the low-frequency part F L (u,v) and the high-frequency component F H(u,v). This frequency domain analysis method can clearly separate low-frequency and high-frequency textures in an image. Low-frequency textures typically correspond to the overall structure and macroscopic features of the image, while high-frequency textures reflect the details and variations. This is achieved by calculating the low-frequency texture energy E. L and high-frequency texture energy E H This quantifies the richness of different types of textures in an image. This texture energy information can provide the model with more information about the map resource features, helping the model better understand the map's content.

[0105] 3. Traditional model training parameter settings are usually fixed or adjusted based on experience, without considering the characteristics of the map image itself. This fixed parameter setting method cannot adapt to the diversity of different map images, resulting in poor training performance, slow convergence speed, or overfitting on some map images. This application defines a parameter adjustment function G, based on the low-frequency texture energy E. L and high-frequency texture energy E H This allows for dynamic adjustment of model training parameters, such as the learning rate α, batch size B, and number of training epochs E. For example, when the high-frequency texture energy E... H Relative low-frequency texture energy E L A higher learning rate indicates rich detail in the image. In this case, the learning rate can be appropriately increased to accelerate the model's learning speed of detailed features. Simultaneously, adjusting the batch size and number of training epochs allows the model to better adapt to the image's texture features. This adaptive parameter adjustment method improves the model's training efficiency and performance, enabling it to achieve good recognition results on different map images.

[0106] This application dynamically adjusts training parameters based on image texture, enabling the model to better learn the essential features of map resources, reduce overfitting to specific images, and thus improve the model's generalization ability, allowing it to accurately identify map resources when faced with new map images.

[0107] Optionally, the data augmentation of the training samples acquired in batches includes:

[0108] The attention-based region augmentation model calculates the texture complexity of the training samples and determines the structuring elements on the training samples. Based on the texture complexity and structuring elements, data augmentation parameters are matched to perform data augmentation on the batch-acquired training samples.

[0109] Preferably, in one embodiment, in the attention-based region augmentation model, the training sample set acquired in batches is set as follows: in Let H represent the i-th training sample image, where H is the image height, W is the image width, and C is the number of image channels.

[0110] To more accurately capture important regions in the training samples, a multi-layer attention mechanism is employed. An L-layer attention function is defined. Each attention function will input the feature map F i,l-1 (When l = 1, F) i,0 =X i Mapped to an attention graph Where H l and W l These are the height and width of the attention map at layer l.

[0111]

[0112] The attention map of each layer is used to weight the feature map of the next layer to obtain a new feature map F. i,l :

[0113] F i,l =F i,l-1 Upsample(M) i,l )

[0114] Where ⊙ denotes element-wise multiplication, and Upsample is the upsampling function used to adjust the size of the attention map to match the feature size. Figure 1 To. The final attention map M i This can be obtained by fusing the attention maps of each layer, for example:

[0115]

[0116] Where ω l It is the fusion weight of the attention maps of each layer, and

[0117] Texture complexity calculation considers not only Local Binary Pattern (LBP) but also incorporates features from the Gray-Level Co-occurrence Matrix (GLCM). For training sample X... i Divide it into multiple non-overlapping local regions {R}. i1 ,R i2 ,…,R iN}, where each local region h and w are the height and width of the local region, respectively.

[0118] For each local region R ij Calculate its LBP features L ij and GLCM features G ij Define a feature fusion function F(L) ij G ij ), fusing LBP features and GLCM features: F ij =F(L ij Gij )=β1L ij +β2G ij , where β1 and β2 are the weights of feature fusion, and β1+β2=1.

[0119] Based on the fused characteristics F ij The statistical information (such as variance, entropy, etc.) is used to measure the texture complexity T of the local region. ij The entire training sample X i Texture complexity T i It can be defined as the weighted sum of the texture complexity of all local regions: Where (x) j ,y j ) is a local region R ij The central position.

[0120] Multi-scale morphological operations are used to determine structuring elements. Sets of structuring elements at different scales are defined. For the structural element S at each scale k For training sample X i Perform morphological operations (such as erosion, dilation, etc.) to obtain the morphologically processed image X. i,k .

[0121] Combining attention map M i Define the struct element extraction function It will train sample X i Attention Map M i The set of structuring elements at different scales is taken as input, and the output is a set of structuring elements. Each structural element S ik It can be described using parameters such as its location, size, shape, and scale.

[0122] Preferably, in one embodiment, a data augmentation parameter matching function is defined during data augmentation parameter matching. It will reduce texture complexity T i and struct set As input, the output is a set of data augmentation parameters α. i =(α i1 ,α i2 ,…,α iL To match parameters more accurately, a parameter matching model P is introduced, implemented through a neural network: Where θ P These are the parameters of the parameter matching model.

[0123] Data augmentation parameters can include rotation angle, scaling factor, brightness adjustment factor, contrast adjustment factor, etc. For example, rotation angle α i1It can be determined based on texture complexity and the orientation information of structuring elements: in It is a set of structural elements The main directional information is f1, which is a nonlinear function.

[0124] Scaling α i2 This can be determined based on the texture complexity and the size of the structuring element: in It is a set of structural elements The average size information is given by f2, which is a non-linear function.

[0125] Preferably, in one embodiment, during the data augmentation operation, the matched data augmentation parameter α is used... i For training sample X i Perform data augmentation operations. Define a composite data augmentation function E(X). i ,α i It will train sample X i and data augmentation parameter α i As input, the output is the enhanced training sample X. i′ Composite data augmentation functions can be composed of multiple basic data augmentation operations, for example:

[0126] X i ′ =E(X) i ,α i )=Contrast(Brightness(Rotate(Scale(X i ,α i2 ),α i1 ),α i3 ),α i4 )

[0127] Where Scale is the scaling operation, Rotate is the rotation operation, Brightness is the brightness adjustment operation, Contrast is the contrast adjustment operation, and α is the scaling operation. i3 and α i4 These are the brightness adjustment factor and the contrast adjustment factor, respectively.

[0128] In the preferred embodiment above, the actual physical meanings of each parameter are as follows: B: Number of training samples acquired in batches. H, W, C: Height, width, and number of channels of the training sample image. h, w: Height and width of the local region. N: Number of local regions into which the training sample is divided. L: Number of attention layers and data augmentation parameters. K: Number of structuring elements at different scales. M: Number of structuring elements. ω lβ1 and β2: Fusion weights of attention maps at each layer. The parameters of the attention function in the l-th layer. θ P : Parameters of the parameter matching model. The attention function for layer l calculates the attention map for layer l. Upsample: An upsampling function used to adjust the size of the attention map. F(L) ij G ij ): Feature fusion function that fuses LBP features and GLCM features. ij Local region R ij Texture complexity. T i Training sample X i Texture complexity. Structural element extraction function to determine training sample X i The set of structural elements. The data augmentation parameter matching function matches data augmentation parameters based on texture complexity and structuring set. E(X) i ,α i ): Composite data augmentation function, applied to training samples X i Perform data augmentation operations. f1, f2: Non-linear functions used to determine the rotation angle and scaling ratio.

[0129] Preferably, in a specific application scenario, the training sample set acquired in batches is read. For each training sample X i Calculate the attention map M of each layer in sequence. i,l And merge them to obtain the final attention map M. i Then, the training samples are divided into local regions, and the LBP feature L for each local region is calculated. ij and GLCM features G ij And fused to obtain the fusion feature F ij Then, the texture complexity T of each local region is calculated. ij The texture complexity T of the entire training sample i Next, the structuring element extraction function S based on multi-scale morphology is used to determine the structuring element set of the training samples. Then use the parameter matching model P according to T i and Matching data augmentation parameter α i Finally, the composite data augmentation function E is used to augment the training samples X. i Perform data augmentation to obtain augmented training samples X. i′ After processing, the enhanced training sample set is output.

[0130] Preferably, in a specific application scenario, the above-mentioned preferred technical solution has the following technical advantages:

[0131] 1. Traditional data augmentation methods typically perform a global, uniform operation on the image, failing to effectively distinguish the importance of different regions within the image. In map images, some key regions (such as specific features and boundaries) are crucial for model learning, but traditional methods cannot specifically augment these regions, resulting in augmented samples that insufficiently supplement the model's effective information. This application's layered attention mechanism uses an L-layer attention function to progressively refine the attention given to features at different levels of the image. Attention maps M for each layer are shown. i,l It can capture the importance of image regions at different scales, from macro to micro. The final fused attention map M i This allows for more precise identification of key areas in map images. For example, in map images containing complex features, it can highlight key areas such as road intersections and city centers, making subsequent data augmentation operations more targeted. The model can better learn the features of key areas, improving recognition accuracy. Each attention map layer performs a weighted operation (F) on the feature map layer below. i,l =F i,l-1 Upsample(M) i,l This approach integrates features from different levels, enabling the model to learn richer and more representative features and understand map image content at different scales. Compared to traditional single-level analysis, this significantly enhances the model's ability to understand complex map scenes.

[0132] 2. Traditional texture complexity calculation methods often rely on only a single feature, such as using only LBP to describe texture, which cannot fully reflect the rich texture information in map images. The texture features of different features in a map are complex and diverse; a single feature cannot cover all feature differences, leading to inaccurate texture complexity assessments and affecting the targeted nature of data augmentation. This application integrates LBP and GLCM features, using the feature fusion function F(L... ij G ij )=β1L ij +β2G ij It integrates local texture patterns and spatial correlation information of textures. This makes the local texture complexity T... ijThe calculation is more comprehensive and accurate. For example, when distinguishing between forests and grasslands, GLCM features can reflect the spatial differences in texture distribution between the two. Combining LBP features can more accurately assess texture complexity, providing a more accurate basis for subsequent data augmentation. This allows the augmented samples to better distinguish different land cover textures and improves the model's ability to learn texture features. By adjusting the weights of β1 and β2, the contributions of LBP and GLCM features in texture complexity calculation can be adaptively adjusted according to the characteristics of different map scenes and land cover types. This flexibility allows the method to better adapt to various complex map images, significantly improving the adaptability and accuracy of texture complexity calculation compared to fixed feature calculation methods.

[0133] 3. Traditional methods for determining structural elements typically perform morphological operations at a single scale, making it difficult to adapt to the structural characteristics of features of varying sizes in map images. For maps containing both large urban areas and small buildings at different scales, single-scale operations cannot comprehensively capture their structural information, resulting in unrepresentative structural elements and impacting data augmentation effectiveness. This application utilizes sets of structural elements at different scales. Performing multi-scale morphological operations can comprehensively capture the structural features of features of different sizes in map images. For example, large-scale structural elements can capture the outlines of large features such as cities and lakes, while small-scale structural elements can refine the structure of small features such as buildings and road signs. The set of structural elements is determined by combining attention maps. This allows the determination of structuring elements to better match the actual structural distribution of map images, providing richer and more accurate structural information for data augmentation. It helps the model learn the structural features of map features at different scales and improves the model's ability to recognize features of different sizes.

[0134] Structure Extraction Function Attention map M i By incorporating attention into the data, priority is given to structural features in important regions of the image. This makes the identified structural elements more representative and better reflects the structural information of key regions. Compared to methods that do not consider attention, this improves the accuracy and effectiveness of structural element extraction, thereby enhancing the strengthening effect of data augmentation on key structural features.

[0135] 4. Traditional data augmentation parameter settings are often based on experience or simple rules, such as fixed rotation angles and scaling ratios, without fully considering the texture complexity and structural features of the image itself. This one-size-fits-all approach cannot personalize augmentation for the characteristics of different map images, resulting in poor augmentation effects and making it difficult for the model to learn more effective information from the augmented samples. This application implements a data augmentation parameter matching model through a neural network. Able to determine the texture complexity T of each map image i and struct set Precise matching data augmentation parameter α i For example, for regions with high texture complexity and intricate structural elements, larger rotation angles and more detailed scaling ratios can be applied to highlight their complex features; for regions with simple textures and uniform structures, relatively conservative enhancement parameters are used to avoid over-enhancement that could lead to information loss. This personalized parameter adaptation greatly improves the targeting of data augmentation, enabling the model to learn information that better matches the actual features of the map image, thereby enhancing the model's generalization ability and recognition performance.

[0136] Parameter matching models can learn the complex nonlinear relationships between texture complexity, structuring information, and data augmentation parameters. Compared to parameter settings based on simple rules, they can more accurately determine appropriate augmentation parameters based on image features. For example, by learning from a large number of map image samples, the model can discover the potential relationship between specific structuring element orientations and rotation angles, thereby providing more optimized data augmentation schemes for different map images and further improving data augmentation effects and model training quality.

[0137] Optionally, the attention-based region augmentation model calculates the texture complexity of the training samples and determines the structuring elements on the training samples, and matches data augmentation parameters based on the texture complexity and structuring elements to perform data augmentation on the batch-acquired training samples, including:

[0138] Based on the input layer, structuring elements are extracted from the training samples and encoded to generate structuring element feature vectors;

[0139] Based on the encoder, multi-scale texture feature analysis is performed on the training sample vector to obtain a texture feature map, and Fourier transform is performed on the texture feature map to obtain quantization energy. Texture complexity is calculated based on the quantization energy.

[0140] A saliency heatmap is generated for the training samples based on the pre-trained MGAN-UNet.

[0141] Based on the saliency heatmap and the texture complexity, the structuring element feature vector is enhanced to augment the training samples acquired in batches.

[0142] Preferably, in one embodiment, when generating the structuring element feature vector, the training sample set acquired in batches is set as follows: in Let H be the i-th training sample image, where H is the image height, W is the image width, and C is the number of image channels.

[0143] Using a structuring element extraction function based on multi-scale morphology and attention mechanisms From training sample Xi Extracting the structural element set Where M i It is an attention map generated through a multi-layered attention mechanism. It is a collection of structural elements at different scales.

[0144] Multi-layer attention mechanism: Define an L-layer attention function Each attention function will input the feature map F i,l-1 (When l = 1, F) i,0 =X i Mapped to an attention graph

[0145]

[0146] F i,l =F i,l-1 Upsample(M) i,l )

[0147] Final attention map in

[0148] For each structuring element S ij Encoding is performed using a deep autoencoder. Generate its feature vector v ij A depth autoencoder consists of an encoder and a decoder. The encoder part is as follows:

[0149] in These are the parameters for the encoder and decoder, respectively.

[0150] Preferably, in one embodiment, when generating the structuring element feature vector, a weighted pooling method is used to generate sample X. i Structural element feature vector V i :

[0151]

[0152] Where w ij It is the structural element S ij The weights can be determined based on their importance in the attention map, for example...

[0153] Preferably, in one embodiment, an encoder with residual connections and attention mechanisms is used when calculating texture complexity. For the training sample vector X i Multi-scale texture feature analysis was performed to obtain texture feature maps at different scales.

[0154]

[0155] encoder In this process, the feature map update of each layer uses residual connections:

[0156] Attention is the attention module, and Conv is the convolution operation.

[0157] Preferably, in one embodiment, during Fourier transform and quantization energy calculation, the texture feature map at each scale is... Perform a two-dimensional discrete cosine transform (DCT). Obtain the frequency domain representation

[0158] Calculate the energy spectrum in the frequency domain.

[0159] Where ρ(u,v) is a frequency weighting function used to highlight certain frequency components, for example...

[0160] Adaptive quantization of the energy spectrum yields the quantized energy.

[0161] in It is an adaptive threshold, which can be calculated using the Otsu algorithm.

[0162] Preferably, in one embodiment, when calculating texture complexity, the weighted sum of quantization energies at all scales is calculated as the texture complexity T. i :

[0163] Where γ s The weight of scale s is determined based on the importance of the feature at that scale.

[0164] Preferably, in one embodiment, a pre-trained improved MGAN-UNet model is used when generating the saliency heatmap. For training sample X i Generate a saliency heatmap H i The improved MGAN-UNet model introduces the idea of ​​Generative Adversarial Networks (GANs), using a generator... and discriminator Conduct competitive training.

[0165]

[0166] During training, the discriminator The goal is to distinguish between true saliency heatmaps and heatmaps generated by the generator. The goal is to generate heatmaps that can fool the discriminator. The loss function is:

[0167]

[0168] At the same time, pixel-level loss is introduced. and perceived loss To improve the quality of heatmaps: Where λ1 and λ2 are weighting coefficients.

[0169] Preferably, in one embodiment, during structuring feature vector enhancement and data augmentation, the saliency heatmap H is used as the basis. i Texture complexity T i For the eigenvector V of the structural element i Perform enhancements. Define the enhancement function. for: Where σ(T) i ) is based on texture complexity T i A defined scaling factor, such as σ(T) i ) = tanh(T i / T max ), T max It is the maximum value of texture complexity. It is a weighted adjustment of the structuring element feature vector based on the attention mechanism of saliency heatmap.

[0170] Preferably, in one embodiment, the enhanced structuring element feature vector V is used. i′ For training sample X i Perform data augmentation. Define the data augmentation function. for: Where Affine is the affine transformation function, T(V) i′ T(V) is the affine transformation matrix generated from the enhanced structuring element eigenvectors, which can be implemented using a multilayer perceptron (MLP): T(V) i′ ) = MLP(V i′ ;θ T )

[0171] In the above preferred scheme, the physical meaning of each parameter is explained as follows: B: Number of training samples acquired in batches. H, W, C: Height, width, and number of channels of the training sample image. N i Sample X i The number of structuring elements. S: Number of scales in multi-scale texture feature analysis. L: Number of layers in the multi-layer attention mechanism. K: Number of structuring elements at different scales. ω l : Fusion weights of multi-layer attention maps. The parameters of the attention function in the l-th layer. θ C : Parameters of the depth autoencoder. θ F θ: Parameters of an encoder with residual connections and attention mechanisms. M : Parameters of the improved MGAN-UNet model. θ T γ: Parameters for generating the affine transformation matrix using a multilayer perceptron. s λs: The weight of scale s in texture complexity calculation. λ1, λ2: Weight coefficients of the loss function in the improved MGAN-UNet model. T max : The maximum value of texture complexity. Structural element extraction function based on multi-scale morphology and attention mechanism. A deep autoencoder encodes structuring elements. An encoder with residual connections and attention mechanisms performs multi-scale texture feature analysis. Two-dimensional discrete cosine transform function. A pre-trained improved MGAN-UNet model generates saliency heatmaps. Structural element eigenvector enhancement function. Data augmentation functions that use affine transformations to augment data.

[0172] Preferably, based on the above-mentioned preferred scheme, the execution process in a scenario is as follows:

[0173] Read the training sample set acquired in batches For each training sample X i Perform the following operations in sequence:

[0174] 1. Generate multi-layer attention map M i Call the structure extraction function ε to extract the structure set.

[0175] 2. Use a depth autoencoder Each structuring element is encoded, and weighted pooling is used to generate the structuring element feature vector V. i .

[0176] 3. Invoke the encoder with residual connections and attention mechanisms. Multi-scale texture feature analysis is performed, and DCT transformation and adaptive quantization are applied to the texture feature map to calculate the texture complexity T. i .

[0177] 4. Use a pre-trained improved MGAN-UNet model Generate a saliency heatmap H i .

[0178] 5. Call the enhancement function For the eigenvector V of the structural element i Enhancement is performed to obtain V i′ .

[0179] 6. Call the data augmentation function For training sample X i Data augmentation is performed to obtain the augmented sample X. i′ .

[0180] After processing, the enhanced training sample set is output.

[0181] The technical solutions described above, in data augmentation scenarios based on map resource identification, have the following technical advantages over traditional technologies in several aspects, including structural element extraction, texture complexity calculation, saliency heatmap generation, and data augmentation:

[0182] 1. Traditional structuring element extraction methods are typically based on fixed morphological operations or simple edge detection algorithms, lacking attention to important regions in the image and struggling to adapt to map features of varying scales and complexities. Furthermore, in terms of structuring element encoding, traditional methods often employ simple feature descriptors, failing to fully capture the complex features of structuring elements, resulting in limited model understanding and learning capabilities regarding map structure. This application employs a structuring element extraction function based on multi-scale morphology and attention mechanisms. Attention map M generated by a multi-layer attention mechanism i Able to focus on key areas in a map, a collection of structural elements at different scales. This allows for the capture of structural features of varying sizes. For example, when recognizing city maps, large-scale structural elements can extract the overall outline of the city, while small-scale structural elements can focus on details such as streets and buildings. At the same time, the attention mechanism ensures priority processing of important areas, improving the accuracy and comprehensiveness of structural element extraction.

[0183] This application uses a depth autoencoder. Encoding structuring elements allows for the learning of their deep feature representations. Compared to traditional simple feature descriptors, deep autoencoders can automatically mine complex features from structuring elements, improving feature expressiveness. For example, when dealing with map structures with complex textures and shapes, deep autoencoders can better capture their unique features, providing more valuable information for subsequent model training.

[0184] 2. Traditional methods for calculating texture complexity typically consider only a single texture feature, such as Local Binary Pattern (LBP) or Gray-Level Co-occurrence Matrix (GLCM), which cannot fully reflect the diversity and complexity of textures in map images. Moreover, in terms of frequency domain analysis, traditional methods use simple Fourier transforms, lacking effective weighting and adaptive quantization of different frequency components, resulting in inaccurate assessments of texture complexity.

[0185] This application features an encoder with residual connectivity and attention mechanisms. It can extract texture features at multiple scales and enhance the expressive power of these features through residual connections and attention mechanisms. Residual connections can alleviate the vanishing gradient problem, enabling the model to learn deeper-level features; attention mechanisms can focus on important parts of the texture, improving the efficiency of feature extraction. For example, when processing map images containing different terrains and features, multi-scale feature extraction can capture texture changes from macro to micro levels, and attention mechanisms can highlight key texture regions, thus more accurately assessing texture complexity. Two-dimensional discrete cosine transform (DCT) is employed. Instead of the traditional Fourier transform, DCT is better suited for processing image signals, as it can better concentrate the image's energy. Simultaneously, it uses an adaptive quantization method, with an adaptive threshold calculated based on the Otsu algorithm. Quantizing the energy spectrum can more accurately reflect the true complexity of the texture. For example, when processing map images with different lighting conditions and texture densities, adaptive quantization can adjust the quantization threshold according to the actual situation of the image, avoiding the limitations of traditional fixed threshold quantization.

[0186] 3. Traditional salient heatmap generation methods are typically based on handcrafted features or simple machine learning models, making it difficult to accurately capture salient regions in map images. Furthermore, these methods lack effective evaluation and optimization of the generated heatmap quality, resulting in blurry and inaccurate heatmaps that cannot provide reliable guidance for subsequent data augmentation. This application uses a pre-trained improved MGAN-UNet model. The generation of saliency heatmaps incorporates the concept of Generative Adversarial Networks (GANs), using a generator... and discriminator Adversarial training can generate more realistic and accurate saliency heatmaps. Simultaneously, pixel-level loss is introduced. and perceived loss This improves the quality of heatmaps, making them not only similar to real heatmaps at the pixel level but also possessing good visual effects at the perceptual level. For example, when dealing with complex map scenes, the improved MGAN-UNet model can more accurately identify important map elements, such as city centers and transportation hubs, and the generated salient heatmaps can provide more targeted guidance for subsequent data augmentation.

[0187] 4. Traditional data augmentation methods typically employ random transformations such as rotation, scaling, and flipping, lacking consideration for the image's inherent features. This can lead to excessive differences between the augmented and original samples, or failure to highlight key image features, thus impacting model training performance. This application's structuring feature vector augmentation function... According to the significance heatmap H i Texture complexity T i For the eigenvector V of the structural element i Enhancement is performed to make the enhanced feature vectors better reflect the key features of the map. For example, for regions with high texture complexity, the enhancement function can appropriately increase the feature weights of the region to highlight its texture information; for regions with high saliency, the enhancement function can focus on preserving the structural features of the region.

[0188] Data augmentation functions Using the affine transformation matrix T(V) generated based on a multilayer perceptron (MLP) i′ Affine transformations applied to training samples can adaptively adjust transformation parameters based on the enhanced structuring element feature vectors. Compared to traditional random transformations, adaptive affine transformations can more specifically enhance key map features, improving the model's ability to recognize different map features. For example, when processing map images containing irregularly shaped features, adaptive affine transformations can adjust according to the structural features of the features, making the enhanced samples more consistent with reality.

[0189] Optionally, training the constructed deep learning model based on the batch-wise enhanced training samples until the map resource recognition model is obtained includes:

[0190] Based on the five-level downsampling structure layer in the backbone encoder module, multi-scale feature extraction is performed on the batches of enhanced training samples to obtain five-level multi-scale feature maps.

[0191] The spatial and channel attention of the five-level multi-scale feature map are calculated based on the spatial-channel dual-dimensional attention gating module, so as to obtain the weighted feature map by weighting the five-level multi-scale feature map.

[0192] Based on the bilinear interpolation decoder module, the weighted feature map is upsampled using bilinear interpolation to obtain an upsampled feature map, and attention modulation is applied to the upsampled feature map to generate a difference feature map.

[0193] Based on the output layer, the difference feature map is mapped to the target category to obtain the category probability feature map. The deep learning model is then trained based on the category probability feature map until the training is completed to obtain the map resource recognition model.

[0194] Preferably, in one embodiment, during the five-level multi-scale feature extraction, the batch-wise augmented training sample set is set as follows: in Let H represent the i-th augmented training sample image, where H is the image height, W is the image width, and C is the number of image channels. The backbone encoder module contains five levels of downsampling structure layers, each of which can be implemented using convolution and pooling operations. Let E be the operation function of the l-th downsampling structure layer. l (·), where l=1,2,…,5.

[0195] For the i-th augmented training sample X i The feature map obtained after passing through the l-th level downsampling structure layer is: The calculation formula is as follows:

[0196] Where, when l = 1,

[0197] The operation function E for each level of the downsampling structure layer l This can be specifically expressed as:

[0198] in, This represents the convolution operation. These are the parameters of the l-th level convolutional layer; This indicates a max pooling operation. These are the parameters of the l-th pooling layer.

[0199] After passing through a five-level downsampling structure layer, a five-level multi-scale feature map is obtained.

[0200] Preferably, in one embodiment, during the spatial-channel dual-dimensional attention gating module processing, the spatial-channel dual-dimensional attention gating module is used to calculate spatial attention and channel attention on the five-level multi-scale feature map, and to perform weighted processing on the feature map. For the l-th level feature map... Spatial attention map The calculation is as follows:

[0201] Where AvgPool(·) and MaxPool(·) represent average pooling and max pooling operations, respectively. This represents the convolution operation. These are the parameters of the spatial attention convolutional layer, and Sigmoid(·) is the activation function.

[0202] Preferably, in one embodiment, during channel attention calculation, the channel attention map... The calculation is as follows:

[0203] Where GlobalAvgPool(·) represents the global average pooling operation, and FC(·,θ) represents the fully connected layer operation. and These are the parameters of the channel attention fully connected layer, and ReLU(·) is the activation function.

[0204] The spatial attention map and channel attention map are used to weight the feature map to obtain the weighted feature map.

[0205] Here, ⊙ represents element-wise multiplication.

[0206] Preferably, in one embodiment, during processing by the bilinear interpolation decoder module, the bilinear interpolation decoder module performs bilinear interpolation upsampling on the weighted feature map. Let the upsampling operation function be U(·), for the l-th level weighted feature map... The feature map obtained after upsampling is Where, size l+1 It is the size of the (l+1)th level feature map.

[0207] Attention modulation is applied to the upsampled feature map to generate the difference feature map I. i l Let the attention modulation function be A(·):

[0208] in, This is the feature map obtained from the previous downsampling stage, used to provide contextual information. The attention modulation function A is expressed as: Wherein, Concat(·) represents the concatenation operation. This represents the convolution operation. These are the parameters of the attention-modulated convolutional layer.

[0209] Preferably, in one embodiment, during target category mapping, the difference feature map is processed based on the output layer. Perform target category mapping to obtain the category probability feature map P. i Let the target category mapping function be O(·):

[0210] Where Conv(·,θ) out ) represents the convolution operation, θ outThese are the parameters of the output convolutional layer, and Softmax(·) is the activation function used to convert the output into a probability distribution.

[0211] Preferably, in one embodiment, during model training, the constructed deep learning model is denoted as M, and its parameters are θ. For each augmented training sample X... i Its corresponding real label is Y i The cross-entropy loss function L is used to measure the predicted class probability feature map P. i Compared with the real label Y i Differences between them: Among them, L CE (·,·) represents the cross-entropy loss function.

[0212] The model's parameters θ are updated using optimization algorithms (such as stochastic gradient descent) until training is complete, resulting in a map resource recognition model.

[0213] Where α is the learning rate. It is the gradient of the loss function with respect to the parameter θ at the t-th iteration.

[0214] The physical meanings of the parameters in the above preferred scheme are explained as follows: N: Number of augmented training samples in batches. H, W, C: Height, width, and number of channels of the augmented training sample image. The parameters of the l-th level convolutional layer. The parameters of the l-th level pooling layer. Parameters of spatial attention convolutional layers. Parameters of the channel attention fully connected layer. Attention modulates the parameters of the convolutional layer. θ out θ: Parameters of the output convolutional layer. α: Parameters of the deep learning model. E: Learning rate. l (·): Operation function of the l-th downsampling structure layer. Conv(·,θ): Convolution operation function. MaxPool(·,θ): Max pooling operation function. AvgPool(·): Average pooling operation function. GlobalAvgPool(·): Global average pooling operation function. FC(·,θ): Fully connected layer operation function. Sigmoid(·): Sigmoid activation function. ReLU(·): ReLU activation function. U(·,size): Bilinear interpolation upsampling operation function. A(·,·): Attention modulation function. O(·): Target class mapping function. L CE (·,·): Cross-entropy loss function.

[0215] In a specific application scenario, the execution process of the above solution is as follows:

[0216] Read the batch of augmented training sample sets For each augmented training sample X i Perform the following operations in sequence:

[0217] 1. Through the five-level downsampling structure layer of the backbone encoder module, X... i Multi-scale feature extraction was performed to obtain a five-level multi-scale feature map.

[0218] 2. Using a spatial-channel dual-dimensional attention gating module, spatial and channel attention are calculated on the five-level multi-scale feature map, and weighted processing is performed to obtain the weighted feature map.

[0219] 3. The weighted feature map is upsampled using bilinear interpolation and attention modulation is applied through the bilinear interpolation decoder module to generate the difference feature map.

[0220] 4. Based on the output layer, analyze the difference feature map. Perform target category mapping to obtain the category probability feature map P. i .

[0221] 5. Calculate the category probability feature map P i Compared with the real label Y i The cross-entropy loss between the two is used to update the model's parameters θ using an optimization algorithm.

[0222] Repeat the above steps until training is complete, and you will obtain the map resource recognition model.

[0223] In a specific application scenario, the above-mentioned technical solution has the following technical advantages over traditional technologies in terms of multi-scale feature extraction, attention mechanism application, feature decoding and modulation, and model training.

[0224] 1. Traditional feature extraction methods often operate only at a single scale, making it difficult to capture feature information of different sizes and levels in map images. For example, when identifying elements such as urban buildings, road networks, and natural terrain in a map, single-scale feature extraction may overlook important detailed features or macroscopic overall features, resulting in an incomplete understanding of the map by the model. This application employs a hierarchical downsampling structure layer design to achieve multi-scale feature extraction. Each level of the downsampling structure layer gradually reduces the size of the feature map through convolution and pooling operations, while increasing the level of feature abstraction. Feature maps at different levels can reflect the feature information of the map image at different scales, effectively capturing everything from microscopic details to macroscopic overall contour features. For example, shallower feature maps can capture detailed information such as road textures and building edges, while deeper feature maps can reflect macroscopic features such as the overall layout of the city and the direction of mountains. This multi-scale feature extraction approach enables the model to understand map images more comprehensively and improves its ability to recognize map elements at different scales.

[0225] 2. Traditional models typically treat all features equally when processing feature maps, lacking attention to and enhancement of important features. In map images, the importance of features in different regions and channels varies for map resource identification, but traditional techniques cannot automatically distinguish the degree of importance of these features, causing the model to be interfered with by irrelevant features and reducing recognition accuracy. This application's spatial-channel dual-dimensional attention gating module performs weighted processing on feature maps by calculating spatial attention maps and channel attention maps. The spatial attention map can focus on important regions in the map image, enhancing the feature representation of these regions; the channel attention map can highlight important feature channels and suppress information from irrelevant channels. For example, when identifying lakes in a map, the spatial attention map can focus attention on the area where the lake is located, while the channel attention map can highlight feature channels related to the lake's color and texture, thereby improving the model's accuracy in identifying lakes. This attention mechanism allows the model to pay more attention to important feature information, reduce interference from irrelevant information, and improve the model's feature selection ability and recognition performance.

[0226] 3. Traditional decoder modules often employ simple interpolation methods, such as nearest neighbor interpolation, during upsampling. This method leads to the loss of image detail, resulting in unclear and inaccurate recovered feature maps. Furthermore, traditional decoders lack effective utilization of contextual information when processing feature maps, making it difficult to fuse feature information at different scales. This application's bilinear interpolation decoder module uses bilinear interpolation for upsampling, which can more smoothly recover the size of the feature map and reduce information loss. Simultaneously, through an attention modulation function, the upsampled feature map is concatenated and convolved with the feature map obtained from the previous downsampling stage, effectively utilizing contextual information and achieving the fusion of feature information at different scales. For example, when recovering detailed information in a map image, the feature map obtained from the previous downsampling stage can provide richer contextual information, helping the decoder better recover detailed elements such as roads and buildings in the map. This combination of bilinear interpolation and attention modulation improves the quality of feature map recovery and feature fusion capabilities, enabling the model to more accurately identify various resources in the map.

[0227] 4. Traditional models, when mapping target categories, employ simple linear mapping or fixed classifiers, lacking the ability to perform non-linear transformations and adaptive adjustments to features. In terms of model training, traditional methods do not fully consider the diversity and complexity of data, resulting in poor generalization ability. This application's target category mapping function uses convolution and softmax activation functions, enabling non-linear transformations of the difference feature map, mapping it to different target categories, and outputting a category probability feature map. This non-linear mapping method better captures the complex relationship between features and categories, improving classification accuracy. Regarding model training, the cross-entropy loss function and optimization algorithms are used to update model parameters, adaptively adjusting the model parameters according to the actual data conditions. This allows the model to better adapt to different map data, improving its generalization ability. For example, when processing map data from different regions and styles, the model can learn the features of this data by continuously adjusting parameters, thereby more accurately identifying resources in the map.

[0228] This application also provides a spatiotemporal intelligent data training and inference method based on deep learning, which includes:

[0229] Collect spatiotemporal intelligent data samples and label the spatiotemporal intelligent data samples to form a spatiotemporal intelligent dataset;

[0230] Based on the set model training parameters, training samples are obtained in batches from the spatiotemporal intelligent dataset;

[0231] Data augmentation is performed on the training samples acquired in batches to generate augmented training samples in batches.

[0232] Based on the batch-wise enhanced training samples, the constructed deep learning model is trained until the training is completed and a spatiotemporal intelligent data training inference model is obtained.

[0233] Preferably, in one embodiment, when collecting and labeling spatiotemporal intelligent data samples, the set of collected spatiotemporal intelligent data samples is set as follows: in Let N represent the i-th original spatiotemporal intelligent data sample, where N is the total number of samples and D is the dimension of the data.

[0234] For each sample x i Perform annotation, and use y as the annotation result. i This indicates that for classification tasks, y i It can be a one-hot encoded vector; for regression tasks, y i It can be a real-valued vector. The labeled spatiotemporal intelligent dataset is...

[0235] Preferably, in one embodiment, when acquiring training samples in batches, model training parameters are set, including batch size B, number of training epochs E, etc. In each epoch e (e = 1, 2, ..., E), samples are acquired from the dataset... B samples are randomly selected to form a batch. Where b represents the b-th batch in this round.

[0236] The sample set for the b-th batch in round e is:

[0237] Where i1, i2, ..., i B It is an index randomly selected from 1 to N.

[0238] Preferably, in one embodiment, during data augmentation, the training samples acquired in batches are augmented to increase data diversity and the model's generalization ability. Let the data augmentation function be... For batch Each sample (x) i ,y i After data augmentation, the augmented sample (x) is obtained. i′ ,y i ),in

[0239] Data augmentation functions It can be a combination of multiple operations, such as: random rotation: x i′ =Rotate(x i ,θ), where θ is a random angle. Random scaling: x i′ =Scale(x)i ,s), where s is a random scaling factor. Random translation: x i′ =Translate(x i ,(t x ,t y )), where (t x ,t y ) is a random translation vector.

[0240] After data augmentation, a batch of augmented training sample sets is obtained.

[0241] Preferably, in one embodiment, during model training, a deep learning model M(x; θ) is constructed, where x is the input data and θ is the model parameters.

[0242] For enhancing the training sample set Each sample (x) i′ ,y i ), x i′ The input is fed into model M to obtain the model's output.

[0243] Using loss function To measure the model output The difference between the actual label y and the true label y. For batches. The average loss is:

[0244] Common loss functions include cross-entropy loss (for classification tasks) and mean squared error loss (for regression tasks).

[0245] During backpropagation and parameter updates, optimization algorithms (such as stochastic gradient descent, SGD) are used to update the model's parameters θ. In the e-th round and b-th batch, the parameter update formula is:

[0246] Where α is the learning rate. It is a loss function The gradient with respect to the parameter θ.

[0247] Repeat the forward propagation, loss calculation, backpropagation, and parameter update process described above until the training termination condition is met, such as reaching the preset training epoch E. Finally, the trained spatiotemporal intelligent data training inference model M(x; θ) is obtained. * ), where θ * These are the model parameters at the end of training.

[0248] In the above preferred scheme, the physical meanings of each parameter are explained as follows: N: Total number of original spatiotemporal intelligent data samples. D: Data dimension. B: Batch size. E: Number of training epochs. θ: Parameters of the deep learning model. α: Learning rate. Data augmentation functions are used to perform data augmentation operations on input samples. `Rotate(·,θ)`: A random rotation function that rotates the input sample around a specified angle θ. `Scale(·,s)`: A random scaling function that scales the input sample by a specified scaling factor s. `Translate(·,(t)` x ,t y ): Random translation function, which shifts the input sample by a specified translation vector (t). x ,t y The input data x is translated. M(x; θ): a deep learning model that maps the input data x to the output. Loss function, used to measure model output The difference between the actual label y and the real label y. loss function The gradient with respect to the parameter θ.

[0249] Preferably, based on the above scheme, during execution, the collected and labeled spatiotemporal intelligent dataset is read. Then, in each round of training, a batch of samples is randomly selected. Data augmentation was performed to obtain Next, the enhanced samples are input into the deep learning model M for forward propagation, and the loss is calculated. Then, the gradient of the loss with respect to the model parameters is calculated through backpropagation. The model parameters θ are then updated using an optimization algorithm. This process is repeated until the preset number of training rounds E is reached, ultimately outputting the trained spatiotemporal intelligent data training inference model M(x; θ). * ).

[0250] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A map resource identification method based on deep learning, characterized in that, include: Collect map resource image samples and annotate the map resource image samples to form a map image dataset; Based on the set model training parameters, training samples are obtained in batches from the map image dataset; Data augmentation is performed on the training samples acquired in batches to generate augmented training samples in batches. Based on the five-level downsampling structure layer in the backbone encoder module, multi-scale feature extraction is performed on the batches of enhanced training samples to obtain five-level multi-scale feature maps. The spatial and channel attention of the five-level multi-scale feature map are calculated based on the spatial-channel dual-dimensional attention gating module, so as to obtain the weighted feature map by weighting the five-level multi-scale feature map. Based on the bilinear interpolation decoder module, the weighted feature map is upsampled using bilinear interpolation to obtain an upsampled feature map, and attention modulation is applied to the upsampled feature map to generate a difference feature map. Based on the output layer, the difference feature map is mapped to the target category to obtain the category probability feature map. The deep learning model is then trained based on the category probability feature map until the training is completed to obtain the map resource recognition model. The map resource image to be identified is acquired and input into the map resource identification model for map resource detection and identification.

2. The method according to claim 1, characterized in that, Collect map resource image samples, including: Based on the established spatial distribution strategy and shooting specifications, map resource image samples are collected from the map resource image sample library. The spatial distribution strategy includes: uniform distribution rules of sample points for different land cover regions. The shooting specifications include: the number of sample points on a single map resource image sample.

3. The method according to claim 1, characterized in that, The map resource image samples are labeled, including: The labeled patches on the map resource image sample are identified, and their geographic coordinates are associated to determine the latitude and longitude and land use code corresponding to the patches, and a patch attribute table record is generated.

4. The method according to claim 1, characterized in that, The method further includes: estimating high-frequency and low-frequency textures based on multi-scale target and pixel segmentation boundaries identified by map resources; and determining the model training parameters based on the high-frequency and low-frequency textures.

5. The method according to claim 1, characterized in that, The data augmentation of the training samples acquired in batches includes: The attention-based region augmentation model calculates the texture complexity of the training samples and determines the structuring elements on the training samples. Based on the texture complexity and structuring elements, data augmentation parameters are matched to perform data augmentation on the batch-acquired training samples.

6. The method according to claim 5, characterized in that, The attention-based region augmentation model calculates the texture complexity of the training samples and determines the structuring elements on the training samples. Based on the texture complexity and structuring elements, it matches data augmentation parameters to perform data augmentation on the batch-acquired training samples, including: Based on the input layer, structuring elements are extracted from the training samples and encoded to generate structuring element feature vectors; Based on the encoder, multi-scale texture feature analysis is performed on the training sample vector to obtain a texture feature map, and Fourier transform is performed on the texture feature map to obtain quantization energy. Texture complexity is calculated based on the quantization energy. A saliency heatmap is generated for the training samples based on the pre-trained MGAN-UNet. Based on the saliency heatmap and the texture complexity, the structuring element feature vector is enhanced to augment the training samples acquired in batches.

7. A spatiotemporal intelligent data training and inference method based on deep learning, characterized in that, include: Collect spatiotemporal intelligent data samples and label the spatiotemporal intelligent data samples to form a spatiotemporal intelligent dataset; Based on the set model training parameters, training samples are obtained in batches from the spatiotemporal intelligent dataset; Data augmentation is performed on the training samples acquired in batches to generate augmented training samples in batches. Based on the five-level downsampling structure layer in the backbone encoder module, multi-scale feature extraction is performed on the batches of enhanced training samples to obtain a five-level multi-scale feature map. The spatial and channel attention of the five-level multi-scale feature map are calculated based on the spatial-channel dual-dimensional attention gating module, so as to obtain the weighted feature map by weighting the five-level multi-scale feature map. Based on the bilinear interpolation decoder module, the weighted feature map is upsampled using bilinear interpolation to obtain an upsampled feature map, and attention modulation is applied to the upsampled feature map to generate a difference feature map. Based on the output layer, the difference feature map is mapped to the target category to obtain the category probability feature map. The constructed deep learning model is then trained based on the category probability feature map until the training is completed, resulting in a spatiotemporal intelligent data training inference model.

Images