Super-resolution high-precision map construction method and device, storage medium and program product

Abstract

The application relates to the technical field of high-definition map construction, in particular to a super-resolution high-precision map construction method, equipment, a storage medium and a program product. The super-resolution high-precision map construction method comprises the following steps: (1) inputting a low-resolution image, extracting multi-scale features and calculating an entropy value, dividing the image into regions with different complexities according to the entropy value, optimizing the number of iterations by using a dynamic time step sampling strategy, iteratively refining the image to a target resolution, balancing fidelity and realism, and outputting a high-resolution image; (2) synthesizing a series of map prior scenes for training a model, introducing a shielding enhancement technology to improve the robustness of the model, fusing prior information and real-time perception data, and generating a super-resolution high-precision map. The application significantly improves the image super-resolution reconstruction quality and efficiency, naturalness and fidelity, and improves the performance and robustness of the model under various prior scenes and shielding conditions.

Description

Technical Field

[0001] This application relates to the field of high-definition map construction technology, and in particular to a method, device, storage medium and program product for constructing super-resolution high-precision maps using a change map prior. Background Technology

[0002] With the rapid development of technologies such as autonomous driving and intelligent navigation, high-definition maps, as one of the core supports for these technologies, are facing increasingly higher demands for accuracy and real-time performance. High-definition maps not only provide autonomous vehicles with precise road information, traffic signs, and obstacle locations, but also form the foundation for intelligent navigation systems to achieve accurate path planning and real-time navigation. However, in real-world applications, the construction of high-definition maps faces a series of complex and challenging issues.

[0003] Real-world maps change frequently and are highly complex, including road widening, building construction or demolition, and traffic sign updates. These changes require high-resolution maps to reflect the latest state of the Earth's surface in real time. However, high-precision map annotation is an extremely time-consuming and labor-intensive task, requiring significant human and material resources. Therefore, the scarcity of high-quality training data is a major factor limiting the accuracy and generalization ability of high-resolution map construction.

[0004] During map building, environmental factors such as occlusion have a significant impact on the accuracy and real-time performance of high-definition maps. For example, due to the obstruction of trees, buildings, and other objects, autonomous vehicles may not be able to directly obtain map information of the obscured areas.

[0005] Therefore, how to enrich data resources, adapt to complex and ever-changing environmental conditions, and enhance the generalization ability of models have become key issues that urgently need to be addressed in the field. Summary of the Invention

[0006] To address the above issues, this application provides a method, device, storage medium, and program product for constructing super-resolution high-precision maps using a changing map prior. This high-precision map construction method can achieve efficient super-resolution reconstruction, accurately identify and optimize the number of iterations to improve image quality. At the same time, it innovates a synthesis method to generate map priors to enrich the training set, improve the model's generalization ability, and introduces occlusion enhancement technology to enhance the model's robustness. By improving the map prior query design and fusing prior information with real-time perception data, it generates high-precision, real-time updated high-definition maps, meeting the real-time requirements of high-definition maps for technologies such as autonomous driving.

[0007] A method for constructing a super-resolution high-precision map includes the following steps:

[0008] (1) Super-resolution reconstruction of low-resolution images: Input a low-resolution image, extract multi-scale features and calculate entropy value, divide the image into regions of different complexities according to the entropy value, optimize the number of iterations using dynamic time step sampling strategy, iterate and refine the image to the target resolution, balance fidelity and realism, adjust model parameters, and output a high-resolution image.

[0009] (2) Constructing real-time high-definition maps based on variable map priors: A series of map prior scenes are generated based on the synthesis method to train the model. At the same time, occlusion enhancement technology is introduced to improve the robustness of the model. Finally, the map prior query design is improved, and prior information and real-time perception data are integrated to generate super-resolution high-precision maps.

[0010] By employing the above technical solutions, in image super-resolution reconstruction, this application accurately identifies the complexity and reconstruction difficulty of different regions in the image through a multi-index latent entropy module, and optimizes the number of iterations by combining a dynamic time-step sampling strategy, making the reconstruction process more efficient and targeted. Simultaneously, the trade-off mechanism between fidelity and realism ensures that the output image presents a more natural and realistic visual effect while retaining the original information. The output high-resolution image provides high-resolution image input for map construction.

[0011] In terms of real-time high-definition map construction, this application utilizes a synthetic method to generate rich prior map data, effectively addressing the problem of scarce real-world map change annotations. Furthermore, map occlusion enhancement technology further improves the model's map construction capabilities under occlusion conditions. By fusing prior information with real-time perception data, high-precision, real-time updated high-definition maps are generated, meeting the real-time requirements of technologies such as autonomous driving for high-definition maps.

[0012] One preferred approach to the super-resolution high-precision map construction method is as follows: In step (1), the input low-resolution image, multi-scale features are extracted and entropy values ​​are calculated. Based on the entropy values, the image is divided into regions of different complexities. The iteration number is optimized using a dynamic time step sampling strategy. The image is iteratively refined to the target resolution, which specifically includes:

[0013] Input a low-resolution image of a real-world scene, use a deep learning network to extract multi-scale features from the image, and calculate the weighted average entropy of the multi-scale features; convert the weighted average entropy into a complexity index; divide the image into regions of different complexities according to the complexity index; use a dynamic time step sampling strategy to allocate an appropriate number of iterations to each region, and the model iteratively refines each region to generate a target resolution image as output; where the number of iterations and the complexity of each region are linearly positively correlated.

[0014] By employing the above technical solution, this approach quantifies the complexity of regions by calculating the entropy of image data in the latent space, thereby identifying which regions require more iterations for fine reconstruction and which regions can achieve satisfactory reconstruction results with fewer steps. A dynamic time-step sampling strategy is adopted, adaptively adjusting the number of iterations for different image regions based on the output of the multi-index latent entropy module, thus optimizing the allocation of computational resources, reducing unnecessary iterations, and accelerating the overall processing flow.

[0015] One preferred approach to this super-resolution high-precision map construction method is that, in step (1), balancing fidelity and realism, adjusting model parameters, and outputting a high-resolution image specifically includes:

[0016] Design a tradeoff function that takes fidelity score and realism score as input and outputs a tradeoff score. The tradeoff is calculated using a linear relationship, resulting in the formula: S = λ·P + (1-λ)·R. In this formula, S represents the tradeoff score, P represents the fidelity score of the target resolution image, R represents the realism score of the target resolution image, and λ is a weighting coefficient between 0 and 1.

[0017] Adjust the model parameters so that the realism score reaches its maximum when the fidelity score is fixed at a certain value; repeatedly adjust the model parameters so that the fidelity score and realism score are adjusted synchronously, so that the balanced score reaches its maximum value, and output the final high-resolution image.

[0018] By adopting the above technical solution, and by introducing a trade-off mechanism between fidelity (i.e., restoring the similarity between the image and the original high-resolution image) and realism (i.e. restoring the naturalness and credibility of the image), it is ensured that the naturalness and credibility of the image are maintained while improving the image resolution.

[0019] One preferred approach to this super-resolution high-precision map construction method is that, in step (2), the generation of a series of prior map scenes based on the synthesis method for training the model specifically includes:

[0020] Define a synthesis function that receives diverse scene parameters and synthesizes realistic map data based on these parameters. Combine the set of realistic map data with the set of real data to obtain an enhanced training set, which contains a series of prior map scenes. Input the series of prior map scenes from the enhanced training set into the model to train the model.

[0021] By adopting the above technical solution, this method uses a synthesis method to generate a series of map prior knowledge, simulates various possible map change scenarios, and enriches the content of the training set.

[0022] One preferred approach to this super-resolution high-precision map construction method is that, in step (2), the introduction of occlusion enhancement technology to improve model robustness specifically includes:

[0023] During training: First, an occlusion mask is applied to the input data to occlude part of the map, thus simulating occlusion in the real world; then, occlusion-enhanced data is added to train the model, so that the model can still use prior information to construct maps even when occluded.

[0024] By adopting the above technical solution, adding occlusion enhancement data during the training process enables the model to still use prior information to build maps even when occluded, and allows autonomous vehicles to maintain high navigation accuracy when facing occlusions such as trees and buildings.

[0025] One preferred approach to this super-resolution high-precision map construction method is that, in step (2), the improved map query design, which integrates prior information and real-time sensing data to generate a super-resolution high-precision map, specifically includes:

[0026] Prior information is added to the original query to optimize the query. The optimized query locates key areas on the map. The system prioritizes collecting real-time perception data of key areas and integrates this real-time perception data with the map's prior scene to generate a fused super-resolution high-definition map.

[0027] By employing the aforementioned technical solution, this method generates high-precision, real-time updated high-definition maps by fusing prior information with real-time perception data, providing reliable navigation support for autonomous vehicles. This not only improves the driving safety and comfort of autonomous vehicles but also provides strong assurance for their autonomous navigation in complex and ever-changing road environments.

[0028] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the super-resolution high-precision map construction method as described above.

[0029] A computer-readable storage medium having a computer program stored thereon, which, when executed, implements the super-resolution high-precision map construction method as described above.

[0030] A computer program product, when run on a computer, enables the computer to execute the super-resolution high-precision map construction method as described above.

[0031] In summary, the super-resolution high-precision map construction method, device, storage medium, and program product of this application have the following beneficial effects:

[0032] A. Significantly improves the reconstruction quality of images and maps.

[0033] In image super-resolution reconstruction, this application accurately identifies the complexity and reconstruction difficulty of different regions in an image through a multi-index latent entropy module, and optimizes the number of iterations by combining a dynamic time-step sampling strategy, making the reconstruction process more efficient and targeted. Meanwhile, the trade-off mechanism between fidelity and realism ensures that the output image presents a more natural and realistic visual effect while preserving the original information.

[0034] In real-time high-definition map construction, this application utilizes synthetic methods to generate rich prior map data, effectively addressing the problem of scarce real-world map change annotations. Through comparative experiments and optimized training processes, the general-purpose model demonstrates performance comparable to or even superior to expert models across multiple prior scenarios. Furthermore, map occlusion enhancement techniques further improve the model's map construction capabilities under occlusion conditions.

[0035] B. Improve processing efficiency and robustness

[0036] In the process of image super-resolution reconstruction, the dynamic time step sampling strategy adaptively adjusts the number of iterations according to the complexity of the image region, effectively reducing unnecessary iterations and thus accelerating the overall processing flow.

[0037] In terms of real-time high-definition map construction, the improved map prior query design more fully incorporates prior map information and, combined with real-time sensing data, achieves high-precision, real-time updated high-definition map generation. Meanwhile, map occlusion enhancement technology enables the model to operate stably even when facing complex situations such as occlusion, improving the system's robustness. Attached Figure Description

[0038] Figure 1 This is a flowchart of the super-resolution high-precision map construction method of this application. Detailed Implementation

[0039] The technical solutions in the embodiments are described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the following embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0040] A method for constructing super-resolution high-precision maps using a changed map prior, referencing Figure 1 The method includes the following steps in sequence:

[0041] (1) Super-resolution reconstruction of low-resolution images: Input a low-resolution image, extract multi-scale features and calculate entropy values, identify regions that need fine reconstruction, and then use a dynamic strategy to optimize the number of iterations to refine the image to near the target high resolution. At the same time, a tradeoff function is used to balance fidelity and realism, and the parameters are adjusted until optimal. Finally, a natural and realistic high-resolution image is output, achieving both quality and efficiency.

[0042] (2) Constructing Real-Time High-Definition Maps Based on Variable Map Priors: An innovative synthesis method is adopted to generate map priors. By defining a synthesis function S to receive diverse parameters and generate realistic map data, the training set is enriched. Then, a high-definition map task is defined, and the performance evaluation function E is used to guide model learning. The effectiveness of the synthetic data augmentation method is demonstrated, and occlusion enhancement technology is introduced to improve the robustness of the model. Finally, the map prior query design is improved, and prior information and real-time perceived data are integrated to generate high-precision, real-time updated high-definition maps.

[0043] Step (1) specifically includes the following AC steps.

[0044] A. Multi-index Latent Entropy Module: This module quantifies the complexity of regions by calculating the entropy of image data in the latent space, thereby identifying which regions require more iterations for fine reconstruction and which regions can achieve satisfactory reconstruction results with fewer steps. This step is implemented through the following ad sub-steps.

[0045] a. Input Image: Input a low-resolution image of a real-world scene. LR The quality of the input image (such as noise level and blur) will affect the subsequent processing results and computational complexity.

[0046] b. Feature extraction: Utilizing a deep learning network (CNN) to extract features from low-resolution images. LR The multi-scale features include various information such as texture, edge, and color, which have different distributions and entropy values ​​in the latent space. The extracted feature set is represented as:

[0047] F ext =F(I LR )

[0048] Where F() is the feature extraction function, F ext Represents the extracted feature set, containing features from I LR Multi-scale features extracted from it.

[0049] c. Entropy Calculation: Calculate the entropy values ​​of these features in the latent space to obtain the complexity index and reconstruction difficulty of each image region. Define the feature set F. ext Let f be a certain feature vector. i Its probability distribution is P(fi If ), then the entropy value of this feature can be expressed as:

[0050]

[0051] Where j represents the feature vector f i The possible values ​​of ...

[0052] By considering the entropy values ​​of multiple features, we can calculate the average entropy or the weighted average entropy. Define the feature set F. ext If there are N features, then the average entropy It can be represented as:

[0053]

[0054] If each feature has a different weight, then a weighted average entropy is used. :

[0055]

[0056] Among them, w i Representing feature f i The weights, H(f) i ) represents feature f i The entropy value.

[0057] Based on entropy, a complexity index is defined to measure the complexity and reconstruction difficulty of different regions of an image. The complexity index C can be expressed as:

[0058]

[0059] If the contributions of multiple features are considered, the complexity index C can be expressed as:

[0060]

[0061] Here, f is a mapping function used to convert entropy values ​​into complexity metrics.

[0062] d. Region segmentation: Based on the complexity index calculated from the entropy value, the image is divided into regions of different complexities. These regions include smooth regions (such as the sky and water surface), textured regions (such as leaves and grass), and edge regions (such as building outlines).

[0063] B. Dynamic Time Step Sampling Strategy: Based on the output of the multi-index latent entropy module, the iteration number (i.e., time step) for different image regions is adaptively adjusted, thereby optimizing the allocation of computational resources, reducing unnecessary iterations, and accelerating the overall processing flow. This step is implemented in the following AC sub-steps.

[0064] a. Receive complexity metrics: First, receive complexity metrics from the multi-metric latent entropy module. These metrics are typically represented numerically or as vectors and are used to measure the complexity and reconstruction difficulty of different regions of the image.

[0065] b. Determine the number of iterations: Based on the received complexity index C, the dynamic time-step sampling strategy allocates an appropriate number of iterations for each image region. Here, a linear relationship is used for allocation, i.e.:

[0066]

[0067] Among them, C i T is the complexity index of the i-th region. i This represents the number of iterations for the i-th region. α and β are model parameters, where α is a positive number used to adjust the relationship between the number of iterations and the complexity.

[0068] c. Iterative refinement: After determining the number of iterations for each region, the model begins iterative refinement of each region.

[0069] Iterative thinning typically involves a series of mathematical operations and image processing techniques, such as gradient descent, convolution operations, and nonlinear activation. These operations aim to progressively optimize image quality, gradually bringing it closer to the target high-resolution image.

[0070] In each iteration, the model adjusts its optimization strategy based on the current image quality and complexity metrics to ensure optimal reconstruction results within a finite number of iterations. Let I be the image after the k-th iteration. k The iterative refinement process can then be expressed as:

[0071]

[0072] Where f represents the iterative refinement function, θ represents the model parameters, and T i This represents the number of iterations for the i-th region.

[0073] Through multiple iterations of refinement, the model generates an image with a resolution close to the target as output.

[0074] C. Trade-off between fidelity and realism: By introducing a trade-off mechanism between fidelity (i.e., restoring the similarity between the image and the original high-resolution image) and realism (i.e., restoring the naturalness and credibility of the image), we ensure that the naturalness and credibility of the image are maintained while improving the image resolution. This step is implemented through the following A / E sub-steps.

[0075] a. Fidelity Assessment and Realism Optimization: Peak Signal-to-Noise Ratio (PSNR) or Structural Similarity Index (SSIM) are used to measure the similarity between the restored image and the original high-resolution image. While maintaining a certain level of fidelity, the realism of the image is optimized by adjusting model parameters or introducing additional regularization terms (such as Perceptual Quality Index (PIQ) or Natural Image Quality Evaluation (NIQE)). This typically involves adjusting details such as image texture, color, and contrast.

[0076] b. Trade-off Mechanism Design: Design a trade-off function that takes fidelity and realism as input and outputs a balanced score. Here, a linear relationship is used for the trade-off:

[0077] S = λ·P + (1-λ)·R

[0078] Where S represents the weighted score, P represents the fidelity score, R represents the realism score, and λ is a weighting coefficient between 0 and 1, used to adjust the relative importance between fidelity and realism.

[0079] c. Parameter adjustment and optimization: Based on the results of the trade-off function, adjust the model parameters (such as the number of iterations, the weight of the regularization term, etc.) to improve the realism of the image while maintaining a certain level of fidelity.

[0080] d. Iterative optimization: Repeat the above process until the optimal trade-off point is found, that is, the score after the trade-off is maximized.

[0081] e. Final Output: After balancing fidelity and realism, the final high-resolution image is output. This image achieves a relative balance between fidelity and realism, preserving the important information of the original image while presenting a natural and lifelike visual effect.

[0082] Step (2) specifically includes the following DH steps.

[0083] D. Using synthetic methods to generate map priors: Real-world map changes are not adequately labeled in the dataset, and their frequency of occurrence is extremely low, approaching 0.

[0084] P(map change) ≈ 0 (training set)

[0085] This scarcity severely limits the model's ability to learn map change patterns (such as road expansion, changes in building height, etc.).

[0086] Therefore, this application utilizes a synthetic method to generate map priors. The core of this method lies in simulating various possible map change scenarios, thereby enriching the content of the training dataset. A synthetic function S is defined, which can receive a series of changing parameters (such as road width, building height, etc.) and generate realistic synthetic map data based on these parameters:

[0087]

[0088] A single composite map data M is generated by calling a single composite function. synth Numerous calls to the synthesis function S (parameters) generate a synthetic map dataset D. synth It contains a series of M synth .

[0089] A typical training set is a collection of real data. By adding more synthetic map data to the typical training set, an enhanced training set is obtained. This method serves as an enhancement technique for building general-purpose high-definition maps. The enhanced training set is represented as follows:

[0090]

[0091] Among them, D real D represents the set of real data. synth This represents the set of synthetic map data. The augmented training set contains a series of prior map scenarios.

[0092] E. High-resolution map completion task definition: Input a series of map prior scenes into the model to learn to process different map priors, where P is the set of prior scenes. scenes Represented as:

[0093]

[0094] Among them, S n This represents the nth prior scenario.

[0095] Define a performance evaluation function E that accepts the model output M. output and real label M true As input, and return a performance score:

[0096]

[0097] F. Comparison of General Model and Expert Model: Constructing expert models and general models respectively:

[0098]

[0099]

[0100] Among them, M expertFor expert models, M generalist For a general model, F represents the model function, and P represents the prior. specific It is P scenes A specific scenario example. P specific Representing the expert model M expert The single scenario prior upon which it relies. P diverse Representing the general model M generalist The diverse priors upon which P relies scenes Is it to construct P diverse The basic unit set, namely P diverse From P scenes The model extracts joint priors from multiple scenarios to train a general model that can adapt to cross-scenario tasks.

[0101] Comparative experiments revealed that the general model exhibits performance comparable to, and even superior to, expert models across multiple prior scenarios, namely:

[0102]

[0103] S perf,expert This is the performance score of the expert model. S perf,generalist It is the performance score of the general model, which is achieved through M generalist Performance calculations were performed across various scenarios. The performance of the general model demonstrates the effectiveness of the synthetic data augmentation method.

[0104] G. Map Occlusion Enhancement: During training, occlusion masks can be applied to the input data to simulate real-world occlusion by partially obscuring the map. The enhanced training set D augmented Includes the original data X and its occluded version X occluded The occluded input data X occluded Represented as:

[0105]

[0106] Where X represents the original input data, M occluded ⨀ represents the occlusion mask, and ⨀ represents element-wise multiplication.

[0107] The process of training a model can be represented as follows:

[0108]

[0109] Among them, D augmented Let M represent the augmented training set, M represent the model, and θ represent the model. trained This represents the parameters of the trained model.

[0110]

[0111] Moutput For the model to test the occlusion data X occluded,test The prediction results.

[0112] Adding occlusion augmentation data during training allows the model to still construct maps using prior information even when occluded.

[0113] H. Map Prior Query Design: Improve the query design to more fully incorporate prior map information. The query function is represented as:

[0114]

[0115] Among them, Q optimized Q represents the query result. original This represents the original query, P represents prior information, and f optimize This represents the optimization function.

[0116] By optimizing query Q optimized After locating the key area, the system prioritizes collecting real-time data from that area. real-time Through advanced algorithms and technologies, real-time data D real-time With prior map HDM prior Fusion generates high-precision, real-time updated high-definition (HDM) maps. fused The merged high-definition map (HDM) fused Represented as:

[0117]

[0118] HDM prior D represents a priori high-definition maps. real-time f represents real-time sensing data. fuse Q represents the fusion function. optimized As a weight matrix, for real-time data D real-time Perform weighted fusion.

[0119] A computer program generated based on the above-mentioned super-resolution high-precision map construction method is stored on an electronic device, which can be a computer, mobile phone, microcontroller, etc. It includes a memory and a processor. The computer program is stored in the memory and can be run by the processor. When the processor executes the computer program, it implements the super-resolution high-precision map construction method.

[0120] This embodiment also describes a computer-readable storage medium on which the above-described computer program is stored. The computer-readable storage medium can be a USB flash drive, hard disk, optical disk, etc. When the computer program is executed, it implements the super-resolution high-precision map construction method described above.

[0121] This embodiment also introduces a computer program product that, when run on a computer, executes the super-resolution high-precision map construction method described above.

[0122] The super-resolution high-precision map construction method of this application significantly improves the quality and efficiency, naturalness and realism of image super-resolution reconstruction by integrating multiple advanced technologies such as multi-index entropy modules, dynamic sampling, fidelity trade-off, synthetic data and occlusion enhancement. It also improves the performance and robustness of the model across multiple prior scenarios and occlusion conditions. By improving the design of map prior queries and fusing prior information with real-time perception data, it generates high-precision, real-time updated high-definition maps, meeting the real-time requirements of high-definition maps for technologies such as autonomous driving.

[0123] Taking a practical application scenario as an example, in complex and ever-changing road environments, including road widening, new building construction, and traffic sign updates, the proposed super-resolution high-precision map construction method based on changing map priors allows autonomous vehicles to acquire low-resolution images in real time via onboard cameras and other sensors, and then perform super-resolution reconstruction to generate high-resolution road images. Simultaneously, the map prior data generated through innovative synthesis methods enriches the training set of autonomous vehicles, improving their adaptability to different road environments. Furthermore, the introduction of occlusion enhancement technology enables autonomous vehicles to maintain high navigation accuracy even when facing obstructions such as trees and buildings. This application provides reliable navigation support for autonomous vehicles, which not only improves their driving safety and comfort but also provides strong assurance for autonomous navigation in complex and ever-changing road environments.

[0124] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for constructing a super-resolution high-precision map, characterized in that, The method comprises the following steps: (1) super-resolution reconstruction of a low-resolution image: input a low-resolution image, extract multi-scale features and calculate an entropy value, divide the image into regions of different complexities according to the entropy value, optimize the number of iterations using a dynamic time step sampling strategy, iteratively refine the image to the target resolution, balance fidelity and realism, adjust the model parameters, and output a high-resolution image; (2) constructing a real-time high-definition map based on a variable map prior: a series of map prior scenes are generated based on a synthesis method for training the model, and a shielding enhancement technique is introduced to improve the robustness of the model; finally, the prior information and real-time perception data are fused to generate a super-resolution high-precision map by improving the design of the map prior query; In step (1), the input low-resolution image, the extraction of multi-scale features and the calculation of the entropy value, the division of the image into regions of different complexities according to the entropy value, the optimization of the number of iterations using a dynamic time step sampling strategy, and the iterative refinement of the image to the target resolution specifically include: input a low-resolution image of a real-world scene, extract multi-scale features of the image using a deep learning network, calculate the weighted average entropy of the multi-scale features, convert the weighted average entropy into a complexity index, divide the image into regions of different complexities according to the complexity index, assign appropriate iteration numbers to each region using a dynamic time step sampling strategy, and generate a target resolution image as output by iteratively refining each region with the model; wherein the iteration number of each region is linearly positively correlated with the complexity; In step (1), the balancing of fidelity and realism, the adjustment of model parameters, and the output of a high-resolution image specifically include: design a trade-off function that takes fidelity score and realism score as input and outputs a trade-off score, and use a linear relationship to trade off to get the formula: S = λ·P + (1-λ)·R, where S represents the trade-off score, P represents the fidelity score of the target resolution image, R represents the realism score of the target resolution image, and λ is a weight coefficient between 0 and 1; adjust the model parameters so that the fidelity score is fixed at a value and the realism score reaches a maximum; repeatedly adjust the model parameters so that the fidelity score and the realism score are adjusted simultaneously, so that the trade-off score reaches a maximum value, and output the final high-resolution image. 2.The super-resolution high-definition map construction method of claim 1, wherein, In step (2), the generation of a series of map prior scenes based on a synthesis method for training the model specifically includes: define a synthesis function that receives diversified scene parameters and synthesizes realistic map data based on these scene parameters, combine the set of realistic map data with the set of real data to obtain an enhanced training set, and the enhanced training set contains a series of prior map scenes; input the series of map prior scenes of the enhanced training set into the model to train the model. 3.The super-resolution high-definition map construction method of claim 1, wherein, In step (2), the introduction of a shielding enhancement technique to improve the robustness of the model specifically includes: In the training process: firstly, the occlusion mask is applied to the input data to occlude part of the map to simulate the occlusion in the real world; then the occlusion enhanced data is added to train the model, so that the model can still use prior information to construct the map in the occluded case. 4.The method of claim 1, wherein, In step (2), the improved map query design fuses prior information and real-time perception data to generate a super-resolution high-precision map, specifically including: The prior information is added to the original query to optimize the query, the key area of the map is located through the optimized query, the system preferentially collects real-time perception data of the key area, and the real-time perception data and the map prior scene are fused to generate a fused super-resolution high-definition map.

5. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the super-resolution high-precision map construction method according to any one of claims 1 to 4 when executing the computer program.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed to implement the super-resolution high-precision map construction method according to any one of claims 1 to 4.

7. A computer program product, characterised in that, When the computer program product runs on the computer, the computer executes the super-resolution high-precision map construction method according to any one of claims 1 to 4.

Images