Satellite image tile map rapid rendering method and system based on artificial intelligence

Abstract

The invention relates to the technical field of tile map rendering, and discloses a satellite image tile map rapid rendering method and system based on artificial intelligence, and the system comprises a cutting module which is configured to judge whether a suspected key region is a key region or not according to the importance characteristics of the region; when it is judged that the suspected key area is the key area, non-uniform tile cutting is conducted on the key area and the background area, and a plurality of tiles are obtained; the target resolution determination module is configured to extract the regional feature parameters of each tile, calculate a regional importance score and determine the target resolution of the tile according to the regional importance score; the processing and storage module is configured to process and store the tiles according to the target resolution; and the splicing module is configured to adjust the image color after the tiles are spliced through the dynamic color balancing module to obtain a satellite image tile map. According to the invention, the efficiency and quality of large-scale high-resolution satellite image online browsing can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of tile rendering technology, and more specifically, to a method and system for rapid rendering of satellite image tile maps based on artificial intelligence. Background Technology

[0002] Traditional online satellite imagery tile rendering technology, through multi-level pyramid tiling and caching mechanisms, can accelerate initial loading speed to some extent. However, when interactively browsing high-resolution, large-scale imagery, this technology reveals multiple bottlenecks.

[0003] First, traditional pyramid slicing uses a fixed resolution and uniform size to segment images, making it impossible to perform differentiated processing for target areas or scenes with rich details. This not only wastes storage and bandwidth resources but also causes key areas to become blurry when zoomed in due to insufficient resolution.

[0004] Secondly, static caching and pre-rendering methods based on simple eviction policies (such as LRU) struggle to accurately predict user browsing patterns. During rapid panning or zooming, "blank windows" or "loading gaps" can easily occur, negatively impacting the user experience.

[0005] Furthermore, to improve detail, applying super-resolution models (such as ESRGAN and EDSR) to low-resolution tiles during offline batch processing can generate high-resolution tiles, but this requires a huge amount of computation and is time-consuming. Moreover, multiple resolution versions need to be stored simultaneously, resulting in poor flexibility and making it difficult to meet real-time interactive requirements.

[0006] In addition, the geometric misalignment, brightness and color inconsistencies that are common when multi-tile tiles are stitched together are difficult to completely eliminate with simple linear blending or filtering methods, resulting in obvious stitching gaps, color blocks jumping around, and disrupting the visual coherence of the overall map.

[0007] Finally, although modern browsers and mobile hardware have strong GPU computing capabilities, existing systems still mainly rely on server-side centralized rendering and super-resolution, resulting in high server load and poor scalability, while client-side computing power is not fully utilized, leading to both idle resources and response latency.

[0008] Therefore, it is necessary to design a rapid rendering method and system for satellite image tile maps based on artificial intelligence to solve the problems existing in the current technology. Summary of the Invention

[0009] In view of this, the present invention proposes a method and system for fast rendering of satellite image tile maps based on artificial intelligence, aiming to improve the efficiency and quality of online browsing of large-scale high-resolution satellite images.

[0010] In one aspect, this invention proposes a rapid rendering system for satellite image tile maps based on artificial intelligence, comprising: The cutting module is configured to acquire satellite images to be processed, perform region identification on the satellite images based on an improved U-Net network to obtain suspected key regions and background regions; extract the regional importance features of the suspected key regions based on the ECA attention mechanism, and determine whether the suspected key regions are key regions based on the regional importance features; when the suspected key regions are determined to be key regions, non-uniform tile cutting is performed on the key regions and background regions to obtain several tiles; The target resolution determination module is configured to extract regional feature parameters for each of the tiles, calculate a regional importance score, and determine the target resolution of the tile based on the regional importance score. The processing and storage module is configured to process and store the tiles according to the target resolution; The stitching module is configured to eliminate seams in the boundary regions of adjacent tiles using an edge fusion module based on a generative adversarial network, and to adjust the color of the stitched image using a dynamic color equalization module to obtain a satellite image tile map.

[0011] Furthermore, when extracting the regional importance features of the suspected key regions based on the ECA attention mechanism, the regional importance features are obtained by the following formula:

[0012] Where Resi represents the regional importance feature of the i-th suspected key region; Ri represents the feature map of the i-th suspected key region; and f represents the mapping function based on the ECA attention mechanism.

[0013] Furthermore, when determining whether the suspected critical region is a critical region based on the regional importance characteristics, the process includes: Obtain the threshold value of the region importance feature corresponding to the region importance feature; The region importance feature is compared with the region importance feature threshold, and the suspected critical region is determined to be a critical region based on the comparison result. If the regional importance feature is greater than the regional importance feature threshold, then the suspected critical region is determined to be a critical region. If the regional importance feature is less than or equal to the regional importance feature threshold, then the suspected key region is determined not to be the key region, and the suspected key region is marked as a background region.

[0014] Furthermore, when performing non-uniform tile cutting on the key area and background area to obtain several tiles, the process includes: When cutting the background area, the cutting is performed with a first cutting precision; When segmenting the key region, the ratio of the region importance feature corresponding to the key region to the region importance feature threshold is obtained and recorded as the importance feature ratio. The importance feature ratio is compared with the first importance feature ratio and the second importance feature ratio, and the cutting accuracy corresponding to the key region is determined based on the comparison result; wherein, the first importance feature ratio is less than the second importance feature ratio; When the ratio of the importance features is less than or equal to the ratio of the first importance features, the cutting accuracy is determined to be the second cutting accuracy. When the ratio of the importance features is greater than the first ratio of the importance features and less than or equal to the second ratio of the importance features, the cutting accuracy is determined to be the third cutting accuracy. When the ratio of the importance features is greater than the ratio of the second importance features, the cutting accuracy is determined to be the fourth cutting accuracy.

[0015] Further, when extracting the regional feature parameters of each tile, calculating the regional importance score, and determining the target resolution of the tile based on the regional importance score, the process includes: The region feature parameters are analyzed to obtain the total number of pixels, texture complexity, edge density, and color richness corresponding to each tile; The importance score of the region is calculated based on the total number of pixels, texture complexity, edge density, and color richness. The regional importance score is compared with the historical resolution group, and the target resolution of the tile is determined based on the comparison result; When there is a historical regional importance score in the historical resolution group that is the same as the regional importance score, the historical target resolution corresponding to the historical regional importance score shall be used as the target resolution. When there is no historical regional importance score in the historical resolution group that is the same as the regional importance score, calculate the absolute value of the difference between the regional importance score and each historical regional importance score, record it as the absolute value of the difference, and extract the minimum absolute value of the difference. The target resolution is determined based on the absolute value of the minimum difference.

[0016] Further, determining the target resolution based on the absolute value of the minimum difference includes: When the absolute value of the minimum difference is unique, obtain the historical region importance score corresponding to the absolute value of the minimum difference, and use the historical target resolution corresponding to the historical region importance score as the target resolution; When the absolute value of the minimum difference is not unique, obtain all historical region importance scores corresponding to the absolute value of the minimum difference, calculate the average value of all historical region importance scores, and record it as the average historical region importance score. The target resolution is determined based on the average historical regional importance score.

[0017] Further, when determining the target resolution based on the average historical region importance score, the following steps are included: The average historical region importance score is compared with the first average historical region importance score and the second average historical region importance score, and the target resolution is determined based on the comparison result; wherein the first average historical region importance score is less than the second average historical region importance score. When the average historical region importance score is less than or equal to the first average historical region importance score, the target resolution is determined as the first target resolution; When the average historical region importance score is greater than the first average historical region importance score and less than or equal to the second average historical region importance score, the target resolution is determined to be the second target resolution. When the average historical region importance score is greater than the second average historical region importance score, the target resolution is determined to be the third target resolution.

[0018] Furthermore, when processing and storing the tiles according to the target resolution, the process includes: Obtain the resolution threshold corresponding to the target resolution; When the target resolution corresponding to the tile is less than or equal to the resolution threshold, the tile is determined to be a low-resolution tile. When the target resolution corresponding to the tile is greater than the resolution threshold, the tile is determined to be a high-resolution tile; When the tile is a low-resolution tile, the tile is compressed using the JPEG2000 compression format, and the compressed low-resolution tile is enhanced by the ESRGAN super-resolution model. When the tile is a high-resolution tile, the tile is compressed using a lossless compression format, and the compressed high-resolution tile is stored in a high-performance storage device.

[0019] Furthermore, when obtaining a satellite image tile map, the edge fusion algorithm based on a generative adversarial network eliminates seams in the boundary regions of adjacent tiles and adjusts the colors of the tile-stitched image using a dynamic color equalization module, the following steps are included: Obtain the boundary regions of adjacent tiles and use these boundary regions as input to the generative adversarial network; Use a generator to generate an image of the seam in the boundary region, and obtain the generated image; The generated image is distinguished from the real image by a discriminator, and the discrimination result is used to calculate the adversarial loss to update the parameters of the generator. Obtain the seam area image output by the generator, and merge the seam area image with the original tile image to replace the original seam area, thereby obtaining the stitched image; Perform color analysis on the stitched image, calculate the color difference ΔE of the stitched area in the stitched image, and determine whether color balance adjustment is needed based on the color difference ΔE of the stitched area. If ΔE > 3, then the color of the splicing area is dynamically adjusted until the color difference ΔE ≤ 3, completing the seam elimination and color balancing process, and outputting the satellite image tile map.

[0020] Compared with existing technologies, the beneficial effects of this invention are as follows: The AI-based satellite image tile map rapid rendering system provided by this invention can effectively improve the efficiency and quality of online browsing of large-scale high-resolution satellite images. Firstly, in terms of image processing, region identification and key region judgment are performed through an improved U-Net network and ECA attention mechanism, followed by non-uniform tile cutting. This method allows for targeted processing based on the importance of each region. For key regions, cutting can be performed with higher precision, preserving more detailed information; while for background regions, cutting is done with relatively lower precision, reducing unnecessary processing and thus improving overall processing efficiency. Regarding resolution determination, the target resolution is determined by calculating the region importance score based on the tile's regional feature parameters, ensuring that tiles of different importance are matched with appropriate resolutions. For low-resolution tiles, JPEG2000 compression format is used with super-resolution enhancement, saving storage space while maintaining a certain level of image quality; for high-resolution tiles, lossless compression format is used and stored on high-performance storage devices, ensuring that high-resolution data of important regions is properly preserved. During the stitching process, a generative adversarial network (GAN) edge blender is used to eliminate seams at the boundaries of adjacent tiles, and a dynamic color equalization module adjusts the image colors, resulting in a more natural and clearer stitched satellite image tile map. This reduces visual discontinuities and color differences, providing users with a high-quality browsing experience. Compared with traditional satellite image tile map rendering methods, the system and method of this invention fully utilize artificial intelligence technology, achieving significant improvements in both efficiency and quality. It better meets the needs of large-scale, high-resolution online browsing of satellite imagery, providing strong support for research and applications in related fields.

[0021] In another aspect, this invention also proposes a method for fast rendering of satellite image tile maps based on artificial intelligence, comprising the following steps: The satellite images to be processed are acquired, and the regions to be processed are identified based on the improved U-Net network to obtain suspected key regions and background regions. Based on the ECA attention mechanism, the regional importance features of the suspected key regions are extracted, and the region importance features are used to determine whether the suspected key regions are key regions. When the suspected critical area is determined to be a critical area, non-uniform tile cutting is performed on the critical area and the background area to obtain several tiles. Extract the regional feature parameters of each tile, calculate the regional importance score, and determine the target resolution of the tile based on the regional importance score; The tiles are processed and stored according to the target resolution; An edge fusion algorithm based on generative adversarial networks eliminates seams in the boundary regions of adjacent tiles, and a dynamic color equalization module adjusts the colors of the spliced ​​tile image to obtain a satellite image tile map.

[0022] It is understandable that the aforementioned AI-based satellite image tile rendering method and system have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A structural block diagram of a satellite image tile map fast rendering system based on artificial intelligence provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a method for rapid rendering of satellite image tile maps based on artificial intelligence, provided in an embodiment of the present invention. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] See Figure 1 As shown in some embodiments of this application, this embodiment provides a rapid rendering system for satellite image tile maps based on artificial intelligence, including: The cutting module is configured to acquire satellite images to be processed, perform region identification on the satellite images based on an improved U-Net network to obtain suspected key regions and background regions; extract the regional importance features of the suspected key regions based on the ECA attention mechanism, and determine whether the suspected key regions are key regions based on the regional importance features; when the suspected key regions are determined to be key regions, non-uniform tile cutting is performed on the key regions and background regions to obtain several tiles; The target resolution determination module is configured to extract regional feature parameters for each of the tiles, calculate a regional importance score, and determine the target resolution of the tile based on the regional importance score. The processing and storage module is configured to process and store the tiles according to the target resolution; The stitching module is configured to eliminate seams in the boundary regions of adjacent tiles using an edge fusion module based on a generative adversarial network, and to adjust the color of the stitched image using a dynamic color equalization module to obtain a satellite image tile map.

[0026] It is understandable that the AI-based satellite image tile rendering system provided in this embodiment can effectively improve the efficiency and quality of online browsing of large-scale high-resolution satellite imagery. Firstly, in terms of image processing, an improved U-Net network and ECA attention mechanism are used for region identification and key region judgment, followed by non-uniform tile cutting. This approach allows for targeted processing based on the importance of each region. For key regions, cutting can be performed with higher precision, preserving more detailed information; while for background regions, cutting is done with relatively lower precision, reducing unnecessary processing and thus improving overall processing efficiency. Regarding resolution determination, a region importance score is calculated based on the tile's regional feature parameters to determine the target resolution, ensuring that tiles of different importance are matched with appropriate resolutions. For low-resolution tiles, JPEG2000 compression format is used with super-resolution enhancement, saving storage space while maintaining a certain level of image quality; for high-resolution tiles, lossless compression format is used and stored on high-performance storage devices, ensuring that high-resolution data of important regions is properly preserved. During the stitching process, a generative adversarial network (GAN) edge blender is used to eliminate seams at the boundaries of adjacent tiles, and a dynamic color equalization module adjusts the image colors, resulting in a more natural and clearer stitched satellite image tile map. This reduces visual discontinuities and color differences, providing users with a high-quality browsing experience. Compared with traditional satellite image tile map rendering methods, the system and method of this invention fully utilize artificial intelligence technology, achieving significant improvements in both efficiency and quality. It better meets the needs of large-scale, high-resolution online browsing of satellite imagery, providing strong support for research and applications in related fields.

[0027] In this embodiment, the satellite image to be processed needs to undergo preprocessing. The preprocessing steps include: (1) Coordinate transformation and projection: The satellite image to be processed usually contains geographic coordinate information, so it is necessary to convert the satellite image to be processed into a unified geographic coordinate system for subsequent processing. This transformation is performed using a common geographic information system library, such as GDAL (Geospatial Data Abstraction Library). (2) Cloud removal: The satellite image to be processed is often affected by clouds, which can obscure surface information and affect subsequent image analysis. Cloud removal algorithms (such as cloud detection methods based on deep learning) are used to detect and remove clouds from the image to ensure that the key information in the satellite image to be processed is clearly visible. (3) Color correction: The satellite image to be processed may have color differences due to different sensors or acquisition conditions. In order to ensure the standardization of the image, color correction is required. Image enhancement techniques (such as histogram equalization) are used to adjust the color tone of the image to make it suitable for subsequent processing.

[0028] In this embodiment, the improved U-Net network is specifically as follows: 1. Using a pre-trained backbone network as the encoder: The standard U-Net encoder consists of simple repetitive convolutional blocks. We improve it by replacing the native encoder with a ResNet network pre-trained on ImageNet, removing its top-level classifier, and retaining the feature extraction part.

[0029] 2. Introducing an attention mechanism: An attention gate is embedded in the skip connection path. The encoder feature map passes through this module before being passed to the decoder. The encoder features are weighted by the high-level semantic features of the decoder as a "gating" signal, which automatically highlights the target area and suppresses the background area, enhances the target feature response, achieves end-to-end fine segmentation, and reduces post-processing dependence.

[0030] 3. Optimize skip connections: Drawing inspiration from the FPN concept, a top-down and horizontal connection approach is adopted to selectively and adaptively fuse features from different levels, so that the fused feature map contains high-quality semantic and positional information.

[0031] 4. Utilizing advanced convolution and loss functions: Partially employing dilated convolution to expand the receptive field without reducing spatial resolution; replacing some standard convolutions with depthwise separable convolutions to achieve model lightweighting. A composite loss function is used: DiceLoss optimizes the intersection-over-union (IoU) ratio between the segmentation results and the ground truth labels, alleviating the background pixel dominance problem; FocalLoss reduces the weight of easily classified samples, focusing on difficult-to-classify samples. In practice, BCE, DiceLoss, and FocalLoss are often combined with weights to guide training, improving the learning performance of minority classes.

[0032] In other words, the improved U-Net network in this embodiment is a powerful segmentation model that integrates a pre-trained ResNet encoder, skip connections with embedded attention gates, and adopts the Dice-Focal composite loss function.

[0033] In this embodiment, the steps for region identification based on the improved U-Net network to obtain suspected key regions and background regions in the satellite image to be processed are as follows: Step 1: Data Preparation and Preprocessing Data collection: Input: High-resolution satellite imagery (e.g., from Google Earth, Sentinel-2, etc.), typically processed as RGB three-channel images or multi-channel images including near-infrared and other features.

[0034] Label: A mask image registered with the input image. This is a single-channel image where the value of each pixel represents its category (e.g., 0 for background, 1 for building), and is usually generated manually by labeling in GIS software.

[0035] Data preprocessing: Image cropping: Satellite images are usually large in size and need to be cropped into smaller tiles suitable for network input (such as 256x256 or 512x512 pixels).

[0036] Data standardization: Normalize the image pixel values ​​from the range [0,255] to [0,1], or perform standardization with a mean of 0 and a variance of 1 to accelerate training convergence.

[0037] Data augmentation: This is a crucial step used to increase data diversity and prevent overfitting. Common operations include random rotation, flipping, brightness / contrast adjustment, adding noise, and elastic deformation. These can be easily implemented using libraries such as ImageDataGenerator or alumnimentations in TensorFlow / Keras.

[0038] Step 2: Model Building Define the network structure: The improved U-Net described above is built using the deep learning framework (TensorFlow).

[0039] The encoder uses a pre-trained ResNet34 (with its top classification layer removed), and its feature outputs at each level serve as the input to the decoder and the source of skip connections. The decoder consists of a series of upsampling and convolutional layers. Attention gates are added to the skip connection paths, ensuring that the encoder features are first filtered by attention weights before being concatenated with the decoder features.

[0040] Compilation model: Choose an optimizer: Adam or SGDwithMomentum are commonly used, as they can adaptively adjust the learning rate.

[0041] Choose a loss function: Select either Binary Crosssentropy (binary classification) or Categorical Crosssentropy (multi-class classification), and usually use it in combination with DiceLoss: TotalLoss=BCELoss+DiceLoss.

[0042] Choose evaluation metrics: monitor Accuracy, IoU (Intersection over Union, the core metric), Precision, and Recall.

[0043] Step 3: Model Training Configure training parameters: Batch size: Select according to GPU memory, such as 8, 16, or 32.

[0044] Number of iterations: Usually set to a large value, and used in conjunction with an early stopping strategy.

[0045] Learning rate: An initial value (such as 0.001) can be set, and a learning rate decay strategy can be used to reduce the learning rate when training stagnates for fine-tuning.

[0046] Perform training: Divide the prepared dataset into a training set, a validation set (and a test set).

[0047] The training data is input into the model, and the weights are iteratively optimized.

[0048] After each epoch, the performance is evaluated on the validation set, and the weights of the best-performing model on the validation set are saved.

[0049] Early stopping is used to automatically terminate training if the performance on the validation set no longer improves after several consecutive epochs, thus preventing overfitting.

[0050] Step 4: Model Evaluation and Inference Model evaluation: The final trained model is fully evaluated using a reserved test set (data that the model has never seen before).

[0051] Analyze the confusion matrix and calculate metrics such as IoU and F1-Score to comprehensively measure model performance. IoU is a key metric for measuring the overlap of segmented regions.

[0052] Model inference: The trained model is then deployed and applied.

[0053] Input new satellite imagery (which may require the same cropping and standardization preprocessing first).

[0054] The model outputs a probability map (the probability that each pixel belongs to a "building").

[0055] The final segmentation result is obtained through post-processing (such as setting a threshold, such as 0.5, to convert the probability map into a binary mask; using morphological operations (such as opening and closing operations) to remove small noise points or fill holes).

[0056] The predicted mask is overlaid on the original image for visualization, revealing potential key and background regions.

[0057] Specifically, when extracting the regional importance features of the suspected key regions based on the ECA attention mechanism, the regional importance features are obtained by the following formula:

[0058] Where Resi represents the regional importance feature of the i-th suspected key region; Ri represents the feature map of the i-th suspected key region; and f represents the mapping function based on the ECA attention mechanism.

[0059] Understandably, the mapping function f, based on the ECA attention mechanism, adaptively learns the channel importance of each suspected key region feature map. The ECA attention mechanism uses one-dimensional convolution to perform local interactions along the channel dimension, avoiding the complex matrix operations of traditional attention mechanisms, thus significantly improving computational efficiency. It utilizes the local correlation between channels to assign different weights to each channel, highlighting important channel features. In practical applications, the region importance feature Resi obtained through this formula can more accurately reflect the importance of suspected key regions. Based on this feature, the system can more accurately determine whether a suspected key region is indeed a key region. For example, when the Resi value is higher than a certain preset threshold, the suspected key region can be determined to be a key region; otherwise, it is considered not to be a key region. This feature extraction method based on the ECA attention mechanism makes the system more intelligent and efficient in identifying key regions. Compared to traditional region importance judgment methods, it can better capture regions with key information in satellite imagery, avoiding over-processing of irrelevant regions, and further improving the performance of the entire satellite image tile map fast rendering system. Moreover, this mechanism, combined with the improved U-Net network, forms an organic whole, laying a solid foundation for subsequent steps such as non-uniform tile cutting and resolution determination, from initial region identification to precise judgment of key areas. Through this series of operations, the system can perform differentiated processing based on the actual conditions of different regions in satellite imagery, thereby greatly improving processing efficiency while ensuring high-quality rendering results and meeting the needs of large-scale online browsing of high-resolution satellite imagery.

[0060] Specifically, determining whether a suspected critical region is a critical region based on the aforementioned regional importance characteristics includes: Obtain the threshold value of the region importance feature corresponding to the region importance feature; The region importance feature is compared with the region importance feature threshold, and the suspected critical region is determined to be a critical region based on the comparison result. If the regional importance feature is greater than the regional importance feature threshold, then the suspected critical region is determined to be a critical region. If the regional importance feature is less than or equal to the regional importance feature threshold, then the suspected key region is determined not to be the key region, and the suspected key region is marked as a background region.

[0061] Understandably, by using this method of comparison and judgment based on regional importance feature thresholds, the system can accurately distinguish between key regions and background regions using quantitative standards. This quantitative judgment avoids interference from subjective factors, making the identification of key regions more objective and reliable. In practical applications, the setting of regional importance feature thresholds can be flexibly adjusted according to different application scenarios and needs. For example, in situations where the requirements for key information are high, the threshold can be appropriately increased to ensure that only those regions with extremely high importance are identified as key regions, thereby concentrating resources on high-precision processing of these regions; while in scenarios where processing efficiency is high, the threshold can be appropriately decreased to expand the scope of key regions and improve processing speed to a certain extent. In addition, this judgment method, combined with the aforementioned improved U-Net network and ECA attention mechanism, forms a complete and efficient key region identification system. The improved U-Net network is responsible for the initial region identification, acquiring suspected key regions and background regions; the ECA attention mechanism extracts features from suspected key regions to obtain regional importance features; finally, by comparing with the threshold, the key regions are accurately determined.

[0062] Specifically, when performing non-uniform tile cutting on the key area and background area to obtain several tiles, the process includes: When cutting the background area, the cutting is performed with a first cutting precision; When segmenting the key region, the ratio of the region importance feature corresponding to the key region to the region importance feature threshold is obtained and recorded as the importance feature ratio. The importance feature ratio is compared with the first importance feature ratio and the second importance feature ratio, and the cutting accuracy corresponding to the key region is determined based on the comparison result; wherein, the first importance feature ratio is less than the second importance feature ratio; When the ratio of the importance features is less than or equal to the ratio of the first importance features, the cutting accuracy is determined to be the second cutting accuracy. When the ratio of the importance features is greater than the first ratio of the importance features and less than or equal to the second ratio of the importance features, the cutting accuracy is determined to be the third cutting accuracy. When the ratio of the importance features is greater than the ratio of the second importance features, the cutting accuracy is determined to be the fourth cutting accuracy.

[0063] Understandably, the preferred cutting accuracy is a gradual increase from the first to the fourth cutting accuracy. This non-uniform tile cutting method allows for differentiated processing based on the importance of key areas. For background areas, a relatively low first cutting accuracy is used because background areas typically contain less critical information; a lower cutting accuracy reduces unnecessary processing and improves efficiency. For key areas, different cutting accuracies are determined based on their importance feature ratios. When the importance feature ratio is less than or equal to the first importance feature ratio, the key area is relatively less important, and a second cutting accuracy is used, ensuring sufficient processing accuracy without excessive resource consumption. When the importance feature ratio is between the first and second importance feature ratios, the key area has some importance, and a third cutting accuracy better preserves its detailed information. When the importance feature ratio is greater than the second importance feature ratio, the key area is very important, and the highest fourth cutting accuracy maximizes the preservation of its details, providing more accurate data for subsequent processing and analysis. This method of determining cutting accuracy based on importance feature ratios makes the system more flexible and efficient in processing satellite imagery. It can rationally allocate processing resources based on the actual conditions of different regions in the image, improving overall processing efficiency while ensuring the processing quality of key areas. Moreover, this cutting method is closely integrated with the preceding region identification and key region judgment steps, forming a complete processing flow. From the initial analysis of the image to the final tile cutting, it fully considers the importance of different regions in the satellite image, laying a solid foundation for subsequent steps such as target resolution determination, processing and storage, and stitching.

[0064] In this embodiment, the preferred value for the first cutting precision is 256×256 pixels. This preferred value is derived from the analysis and experiments of a large amount of satellite imagery data. For the background area, its main function is to provide overall environmental information, with relatively little key information. Using a cutting precision of 256×256 pixels can satisfy the acquisition of basic information of the background area while effectively reducing the amount of data processed. The preferred value for the second cutting precision is 512×512 pixels. When the importance feature ratio of the key area is less than or equal to the first importance feature ratio, it indicates that the importance of the key area is relatively low, but a certain level of processing precision is still required. A cutting precision of 512×512 pixels ensures a certain level of processing precision without excessive resource consumption, achieving a good balance between processing efficiency and the preservation of key area information. The preferred value for the third cutting precision is 1024×1024 pixels. When the importance feature ratio is between the first and second importance feature ratios, the key area has a certain level of importance. A cutting precision of 1024×1024 pixels better preserves the detailed information of the area, providing richer data for subsequent analysis and processing, and helping to more accurately identify and analyze features and targets within the key area. The preferred cutting precision is 2048×2048 pixels. When the ratio of importance features is greater than the ratio of the second most important features, it indicates that the key area is very important. Using the highest cutting precision of 2048×2048 pixels can preserve the details of the key area to the greatest extent, ensuring that even minute features and targets within the key area can be clearly captured, providing the most accurate data support for subsequent high-precision processing and analysis.

[0065] Specifically, when extracting the regional feature parameters of each tile, calculating the regional importance score, and determining the target resolution of the tile based on the regional importance score, the process includes: The region feature parameters are analyzed to obtain the total number of pixels, texture complexity, edge density, and color richness corresponding to each tile; The importance score of the region is calculated based on the total number of pixels, texture complexity, edge density, and color richness. The regional importance score is compared with the historical resolution group, and the target resolution of the tile is determined based on the comparison result; When there is a historical regional importance score in the historical resolution group that is the same as the regional importance score, the historical target resolution corresponding to the historical regional importance score shall be used as the target resolution. When there is no historical regional importance score in the historical resolution group that is the same as the regional importance score, calculate the absolute value of the difference between the regional importance score and each historical regional importance score, record it as the absolute value of the difference, and extract the minimum absolute value of the difference. The target resolution is determined based on the absolute value of the minimum difference.

[0066] In this embodiment, firstly, when calculating the region importance score, it is necessary to comprehensively consider the total number of pixels, texture complexity, edge density, and color richness, and normalize them all, mapping them to the interval [0,1]. This eliminates the differences in the units and numerical ranges of different factors, allowing each factor to be compared and calculated on the same scale. The specific calculation formula can be expressed as: Region Importance Score = Total Pixel Weight × Total Pixels + Texture Complexity Weight × Texture Complexity Quantitative Value + Edge Density Weight × Edge Density Calculated Value + Color Richness Weight × Color Richness Evaluation Value.

[0067] In this embodiment, the preferred weight for the total number of pixels is 0.3. This value is determined after comprehensively considering the influence of the total number of pixels in satellite image tile maps on the importance of a region. In satellite imagery, the total number of pixels reflects, to some extent, the size and information capacity of the region. A larger total number of pixels means that the region contains more details and information, and may occupy a more important position in the overall image. However, the importance of a region cannot be judged solely by the total number of pixels, hence the weight of 0.3. The preferred weight for texture complexity is 0.25. Texture complexity reflects the complexity of the texture within a region. Complex textures often contain more features and information, which is important for determining the type and nature of a region. The preferred weight for edge density is 0.2. Edge density reflects the clarity and complexity of the region's boundaries. Higher edge density may indicate that the region differs significantly from its surrounding environment, making it a region with distinct characteristics. The preferred weight for color richness is 0.25. Color richness reflects the diversity of colors within a region. Rich colors can reflect the material composition and environmental characteristics of the region. By rationally allocating the weights of each factor, the importance score of the region can be calculated more accurately, thus providing a reliable basis for subsequently determining the target resolution of the tiles.

[0068] Understandably, texture complexity is measured by calculating the variations in grayscale values ​​within a tile image. For example, the gray-level co-occurrence matrix (GLCM) method can be used. This method analyzes the spatial relationships between different grayscale levels in an image, calculating features such as contrast, correlation, energy, and homogeneity. These features, combined, reflect the complexity of the texture. Higher contrast indicates a clearer and more complex texture; higher correlation indicates stronger regularity; higher energy indicates a more concentrated texture distribution; and higher homogeneity means smaller local variations in the texture. Edge density is obtained through edge detection algorithms, such as the Canny edge detection algorithm. This algorithm first Gaussian smooths the image to reduce noise, then calculates the gradient magnitude and direction, performs non-maximum suppression to eliminate non-edge pixels, and finally uses double thresholding to determine the true edges. The edge density is obtained by calculating the ratio of the number of detected edge pixels to the total number of pixels in the tile. A higher edge density indicates richer edge features in the tile image. Color richness can be determined by calculating the quantity and distribution of different colors in the image. Images can be converted to the HSV color space, which more intuitively represents the hue, saturation, and brightness of colors. The richness of the colors is assessed by counting the number of different combinations of hue, saturation, and brightness, as well as the percentage of pixels for each combination. The more color combinations there are, and the more evenly they are distributed, the higher the color richness.

[0069] Understandably, by accurately acquiring and analyzing these regional feature parameters, the regional importance score can be calculated more precisely. The regional importance score comprehensively considers multiple factors such as the total number of pixels in the tile, texture complexity, edge density, and color richness, thus reflecting the importance of the tile more comprehensively. When comparing with historical resolution groups, the system can more accurately determine the appropriate target resolution for the tile based on the regional importance score. In this way, higher resolutions can be assigned to tiles containing key information, complex textures, rich edges, or diverse colors to ensure image quality; while lower resolutions can be assigned to tiles with relatively less information and simple textures, thereby saving storage space and processing resources, and further improving the efficiency and performance of the entire satellite image tile map fast rendering system.

[0070] Specifically, determining the target resolution based on the absolute value of the minimum difference includes: When the absolute value of the minimum difference is unique, obtain the historical region importance score corresponding to the absolute value of the minimum difference, and use the historical target resolution corresponding to the historical region importance score as the target resolution; When the absolute value of the minimum difference is not unique, obtain all historical region importance scores corresponding to the absolute value of the minimum difference, calculate the average value of all historical region importance scores, and record it as the average historical region importance score. The target resolution is determined based on the average historical regional importance score.

[0071] Understandably, this method of determining the target resolution based on the absolute value of the minimum difference ensures that a suitable resolution can be found for the tiles in various situations. When the absolute value of the minimum difference is unique, the corresponding historical target resolution is directly adopted, which is simple to operate and can quickly determine the result, making the system processing more efficient. However, when the absolute value of the minimum difference is not unique, the target resolution is determined by calculating the average historical region importance score. This comprehensively considers multiple similar historical cases, avoids the bias that may be caused by a single historical score, and makes the determination of the target resolution more scientific and reasonable.

[0072] Specifically, determining the target resolution based on the average historical region importance score includes: The average historical region importance score is compared with the first average historical region importance score and the second average historical region importance score, and the target resolution is determined based on the comparison result; wherein the first average historical region importance score is less than the second average historical region importance score. When the average historical region importance score is less than or equal to the first average historical region importance score, the target resolution is determined as the first target resolution; When the average historical region importance score is greater than the first average historical region importance score and less than or equal to the second average historical region importance score, the target resolution is determined to be the second target resolution. When the average historical region importance score is greater than the second average historical region importance score, the target resolution is determined to be the third target resolution.

[0073] In this embodiment, the first target resolution is preferably 0.5 meters per pixel, the second target resolution is preferably 2.0 meters per pixel, and the third target resolution is preferably 10.0 meters per pixel; the first target resolution is smaller than the second target resolution, and the second target resolution is smaller than the third target resolution.

[0074] Understandably, this method of determining the target resolution based on the average historical region importance score compared to different thresholds further refines the resolution allocation. When the average historical region importance score is low, less than or equal to the first average historical region importance score, it indicates that the tile is relatively low in importance. Using the first target resolution conserves resources while meeting basic information display requirements. When the average historical region importance score is between the first and second average historical region importance scores, it indicates that the tile has some importance. Using the second target resolution achieves a balance between resource consumption and information preservation. When the average historical region importance score is greater than the second average historical region importance score, it means that the tile is very important. Using the third target resolution maximizes the preservation of tile details, ensures image quality, and provides high-quality data for subsequent analysis and processing.

[0075] Specifically, processing and storing the tiles according to the target resolution includes: Obtain the resolution threshold corresponding to the target resolution; When the target resolution corresponding to the tile is less than or equal to the resolution threshold, the tile is determined to be a low-resolution tile. When the target resolution corresponding to the tile is greater than the resolution threshold, the tile is determined to be a high-resolution tile; When the tile is a low-resolution tile, the tile is compressed using the JPEG2000 compression format, and the compressed low-resolution tile is enhanced by the ESRGAN super-resolution model. When the tile is a high-resolution tile, the tile is compressed using a lossless compression format, and the compressed high-resolution tile is stored in a high-performance storage device.

[0076] In this embodiment, when using the ESRGAN super-resolution model for super-resolution enhancement, given low-resolution tiles, high-resolution tiles are obtained through the ESRGAN model:

[0077] In this model, Thigh represents high-resolution tiles, and Tlow represents low-resolution tiles. The ESRGAN super-resolution model is a super-resolution method based on Generative Adversarial Networks (GANs). It combines a generator and a discriminator, improving image resolution and quality through adversarial training. The generator's main task is to generate high-resolution images from low-resolution images, while the discriminator is responsible for distinguishing between the generated high-resolution images and the real high-resolution images. During training, the generator continuously learns to generate results that are closer to real high-resolution images, while the discriminator continuously improves its discrimination ability. This adversarial training process enables the generator to produce high-quality super-resolution images.

[0078] It's understandable that JPEG2000 compression is used for low-resolution tiles because it offers excellent compression performance and quality. Employing advanced techniques like wavelet transform, it achieves a high compression ratio while maintaining image quality, reducing data storage and transmission burden. The compressed low-resolution tiles are then enhanced using the ESRGAN super-resolution model, effectively improving their resolution and clarity, making them more suitable for subsequent analysis and processing. For high-resolution tiles, lossless compression is used to ensure no image information is lost during compression, preserving the original tile quality. Storing the compressed high-resolution tiles on high-performance storage devices meets the need for fast read / write operations of high-resolution data, facilitating efficient processing and analysis. After processing and storing the tiles, tile stitching can be performed. By stitching the processed tiles according to their positions in the original satellite image, a complete satellite image can be reconstructed. During stitching, issues such as overlapping areas and color matching between tiles need to be considered to ensure a seamless and color-consistent stitched image. Image fusion algorithms, such as multi-resolution fusion algorithms, can be used to process overlapping areas, resulting in a natural transition in the stitched image. Simultaneously, color correction methods can be employed to adjust color differences between different tiles, improving the overall quality of the stitched image. This AI-based method and system for rapid rendering of satellite image tile maps forms a complete and efficient processing flow, from satellite image region identification, key region judgment, tile cutting, target resolution determination, processing and storage to final stitching. By employing differentiated processing methods for different regions and rationally allocating processing resources, the overall processing efficiency and performance are improved while ensuring the processing quality of key areas, providing strong support for the application of satellite imagery.

[0079] Specifically, when obtaining a satellite image tile map, the edge fusion algorithm based on a generative adversarial network eliminates seams in the boundary regions of adjacent tiles and adjusts the colors of the stitched image using a dynamic color equalization module. Obtain the boundary regions of adjacent tiles and use these boundary regions as input to the generative adversarial network; Use a generator to generate an image of the seam in the boundary region, and obtain the generated image; The generated image is distinguished from the real image by a discriminator, and the discrimination result is used to calculate the adversarial loss to update the parameters of the generator. Obtain the seam area image output by the generator, and merge the seam area image with the original tile image to replace the original seam area, thereby obtaining the stitched image; Perform color analysis on the stitched image, calculate the color difference ΔE of the stitched area in the stitched image, and determine whether color balance adjustment is needed based on the color difference ΔE of the stitched area. If ΔE > 3, then the color of the splicing area is dynamically adjusted until the color difference ΔE ≤ 3, completing the seam elimination and color balancing process, and outputting the satellite image tile map.

[0080] Understandably, Generative Adversarial Networks (GANs) are used to eliminate seams at tile stitching. By training the generator to compete against the discriminator, the transition effect in the seam region is optimized, ultimately resulting in a smooth transition at the seams in the stitched image. Formula: Generative adversarial loss can be expressed as:

[0081] Where LGAN represents generative adversarial loss, D represents the output probability of the discriminator, G represents the output of the generator, input represents the input boundary region, and target represents the ground truth boundary region image.

[0082] The color difference ΔE in the splicing area is expressed as:

[0083] Here, ΔE represents the color difference between the spliced ​​areas, and L1, a1, b1 and L2, a2, b2 are the coordinate values ​​of the colors on both sides of the spliced ​​area in the CIELAB color space, respectively. By calculating this color difference, the degree of color difference between the two sides of the spliced ​​area can be quantified.

[0084] Understandably, during training, the generator aims to minimize the LGAN, making it difficult for the discriminator to distinguish between generated and real seam region images, thus making the generated seam region images more realistic. The discriminator, on the other hand, aims to maximize the LGAN, accurately identifying which images are generated and which are real. Through this adversarial training method, the generator continuously optimizes its parameters, gradually generating high-quality seam region images and effectively eliminating seams at the boundaries of adjacent tiles. The dynamic color equalization module adjusts the colors of the stitched image to ensure color consistency in the stitched satellite image tile map. In practical applications, different tiles may have color differences due to factors such as shooting time and lighting conditions, and these differences are particularly noticeable at the stitching points. By calculating the color difference ΔE in the stitched area and using a set threshold (here, 3), it is determined whether color equalization adjustment is needed. When the color difference ΔE is greater than 3, it indicates that the color difference in the stitched area is large, requiring dynamic adjustment. Methods such as histogram matching and color mapping can be used to adjust the colors of the stitched area, making the stitched image have a natural and consistent color transition. This approach, combining an edge fusion processor based on generative adversarial networks (GANs) with a dynamic color equalization module, effectively solves the problems of seams and color inconsistencies in the satellite image tile stitching process. By optimizing the transition effects in seam areas through adversarial training and dynamically adjusting colors based on color differences, the overall quality and visual effect of the satellite image tile are improved, providing more accurate and clearer foundational data for subsequent applications such as geographic information analysis, environmental monitoring, and urban planning. Simultaneously, this method fully leverages the advantages of artificial intelligence technology, achieving automated and intelligent processing, reducing human intervention, and improving processing efficiency, thus better meeting the needs of large-scale satellite image data processing.

[0085] See Figure 2 As shown in some embodiments of this application, this embodiment provides a method for fast rendering of satellite image tile maps based on artificial intelligence, including the following steps: S100: Acquire satellite images to be processed, and perform region identification on the satellite images to be processed based on the improved U-Net network to obtain suspected key regions and background regions; S200: Extract the regional importance features of the suspected key regions based on the ECA attention mechanism, and determine whether the suspected key regions are key regions based on the regional importance features; S300: When the suspected critical area is determined to be a critical area, non-uniform tile cutting is performed on the critical area and the background area to obtain several tiles. S400: Extract the regional feature parameters of each tile, calculate the regional importance score, and determine the target resolution of the tile based on the regional importance score; S500: Process and store the tiles according to the target resolution; S600: An edge fusion processor based on a generative adversarial network eliminates seams in the boundary regions of adjacent tiles and adjusts the color of the image after tile splicing through a dynamic color equalization module to obtain a satellite image tile map.

[0086] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0087] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0088] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0089] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A rapid rendering system for satellite image tile maps based on artificial intelligence, characterized in that, include: The segmentation module is configured to acquire satellite images to be processed, and perform region identification on the satellite images to be processed based on the improved U-Net network to obtain suspected key regions and background regions; Based on the ECA attention mechanism, the regional importance features of the suspected key regions are extracted, and the region importance features are used to determine whether the suspected key regions are key regions. When the suspected critical area is determined to be a critical area, non-uniform tile cutting is performed on the critical area and the background area to obtain several tiles. The target resolution determination module is configured to extract regional feature parameters for each of the tiles, calculate a regional importance score, and determine the target resolution of the tile based on the regional importance score. The processing and storage module is configured to process and store the tiles according to the target resolution; The stitching module is configured to eliminate seams in the boundary regions of adjacent tiles using an edge fusion module based on a generative adversarial network, and to adjust the color of the stitched image using a dynamic color equalization module to obtain a satellite image tile map.

2. The rapid rendering system for satellite image tile maps based on artificial intelligence according to claim 1, characterized in that, When extracting the regional importance features of the suspected key regions based on the ECA attention mechanism, the regional importance features are obtained by the following formula: Where Resi represents the regional importance feature of the i-th suspected key region; Ri represents the feature map of the i-th suspected key region; and f represents the mapping function based on the ECA attention mechanism.

3. The rapid rendering system for satellite image tile maps based on artificial intelligence according to claim 2, characterized in that, When determining whether a suspected critical region is a critical region based on the aforementioned regional importance characteristics, the following methods are included: Obtain the threshold value of the region importance feature corresponding to the region importance feature; The region importance feature is compared with the region importance feature threshold, and the suspected critical region is determined to be a critical region based on the comparison result. If the regional importance feature is greater than the regional importance feature threshold, then the suspected critical region is determined to be a critical region. If the regional importance feature is less than or equal to the regional importance feature threshold, then the suspected key region is determined not to be the key region, and the suspected key region is marked as a background region.

4. The rapid rendering system for satellite image tile maps based on artificial intelligence according to claim 3, characterized in that, When performing non-uniform tile cutting on the key area and background area to obtain several tiles, the process includes: When cutting the background area, the cutting is performed with a first cutting precision; When segmenting the key region, the ratio of the region importance feature corresponding to the key region to the region importance feature threshold is obtained and recorded as the importance feature ratio. The importance feature ratio is compared with the first importance feature ratio and the second importance feature ratio, and the cutting accuracy corresponding to the key region is determined based on the comparison result; wherein, the first importance feature ratio is less than the second importance feature ratio; When the ratio of the importance features is less than or equal to the ratio of the first importance features, the cutting accuracy is determined to be the second cutting accuracy. When the ratio of the importance features is greater than the first ratio of the importance features and less than or equal to the second ratio of the importance features, the cutting accuracy is determined to be the third cutting accuracy. When the ratio of the importance features is greater than the ratio of the second importance features, the cutting accuracy is determined to be the fourth cutting accuracy.

5. The rapid rendering system for satellite image tile maps based on artificial intelligence according to claim 4, characterized in that, When extracting the regional feature parameters of each tile, calculating the regional importance score, and determining the target resolution of the tile based on the regional importance score, the process includes: The region feature parameters are analyzed to obtain the total number of pixels, texture complexity, edge density, and color richness corresponding to each tile; The importance score of the region is calculated based on the total number of pixels, texture complexity, edge density, and color richness. The regional importance score is compared with the historical resolution group, and the target resolution of the tile is determined based on the comparison result; When there is a historical regional importance score in the historical resolution group that is the same as the regional importance score, the historical target resolution corresponding to the historical regional importance score shall be used as the target resolution. When there is no historical regional importance score in the historical resolution group that is the same as the regional importance score, calculate the absolute value of the difference between the regional importance score and each historical regional importance score, record it as the absolute value of the difference, and extract the minimum absolute value of the difference. The target resolution is determined based on the absolute value of the minimum difference.

6. The rapid rendering system for satellite image tile maps based on artificial intelligence according to claim 5, characterized in that, Determining the target resolution based on the absolute value of the minimum difference includes: When the absolute value of the minimum difference is unique, obtain the historical region importance score corresponding to the absolute value of the minimum difference, and use the historical target resolution corresponding to the historical region importance score as the target resolution; When the absolute value of the minimum difference is not unique, obtain all historical region importance scores corresponding to the absolute value of the minimum difference, calculate the average value of all historical region importance scores, and record it as the average historical region importance score. The target resolution is determined based on the average historical regional importance score.

7. The rapid rendering system for satellite image tile maps based on artificial intelligence according to claim 6, characterized in that, When determining the target resolution based on the average historical region importance score, the following are included: The average historical region importance score is compared with the first average historical region importance score and the second average historical region importance score, and the target resolution is determined based on the comparison result; wherein the first average historical region importance score is less than the second average historical region importance score. When the average historical region importance score is less than or equal to the first average historical region importance score, the target resolution is determined as the first target resolution; When the average historical region importance score is greater than the first average historical region importance score and less than or equal to the second average historical region importance score, the target resolution is determined to be the second target resolution. When the average historical region importance score is greater than the second average historical region importance score, the target resolution is determined to be the third target resolution.

8. The rapid rendering system for satellite image tile maps based on artificial intelligence according to claim 7, characterized in that, When processing and storing the tiles according to the target resolution, the process includes: Obtain the resolution threshold corresponding to the target resolution; When the target resolution corresponding to the tile is less than or equal to the resolution threshold, the tile is determined to be a low-resolution tile. When the target resolution corresponding to the tile is greater than the resolution threshold, the tile is determined to be a high-resolution tile; When the tile is a low-resolution tile, the tile is compressed using the JPEG2000 compression format, and the compressed low-resolution tile is enhanced by the ESRGAN super-resolution model. When the tile is a high-resolution tile, the tile is compressed using a lossless compression format, and the compressed high-resolution tile is stored in a high-performance storage device.

9. The rapid rendering system for satellite image tile maps based on artificial intelligence according to claim 8, characterized in that, When obtaining a satellite image tile map, the edge fusion algorithm based on a generative adversarial network eliminates seams in the boundary regions of adjacent tiles and adjusts the colors of the stitched image using a dynamic color equalization module, the following steps are included: Obtain the boundary regions of adjacent tiles and use these boundary regions as input to the generative adversarial network; Use a generator to generate an image of the seam in the boundary region, and obtain the generated image; The generated image is distinguished from the real image by a discriminator, and the discrimination result is used to calculate the adversarial loss to update the parameters of the generator. Obtain the seam area image output by the generator, and merge the seam area image with the original tile image to replace the original seam area, thereby obtaining the stitched image; Perform color analysis on the stitched image, calculate the color difference ΔE of the stitched area in the stitched image, and determine whether color balance adjustment is needed based on the color difference ΔE of the stitched area. If ΔE > 3, then the color of the splicing area is dynamically adjusted until the color difference ΔE ≤ 3, completing the seam elimination and color balancing process, and outputting the satellite image tile map.

10. A method for rapid rendering of satellite image tile maps based on artificial intelligence, applied to the rapid rendering system of satellite image tile maps based on artificial intelligence as described in any one of claims 1-9, characterized in that, include: The satellite images to be processed are acquired, and the regions to be processed are identified based on the improved U-Net network to obtain suspected key regions and background regions. Based on the ECA attention mechanism, the regional importance features of the suspected key regions are extracted, and the region importance features are used to determine whether the suspected key regions are key regions. When the suspected critical area is determined to be a critical area, non-uniform tile cutting is performed on the critical area and the background area to obtain several tiles. Extract the regional feature parameters of each tile, calculate the regional importance score, and determine the target resolution of the tile based on the regional importance score; The tiles are processed and stored according to the target resolution; An edge fusion algorithm based on generative adversarial networks eliminates seams in the boundary regions of adjacent tiles, and a dynamic color equalization module adjusts the colors of the spliced ​​tile image to obtain a satellite image tile map.

Images