Image super-resolution reconstruction method and system

Abstract

The invention discloses an image super-resolution reconstruction method and system, and belongs to the field of digital cultural heritage protection, and the method comprises the steps: constructing a multi-resolution image pyramid, carrying out the semantic analysis based on a low-resolution layer, recognizing a key structure and a uniform region, and generating a self-adaptive dicing strategy; dynamically selecting a super-resolution model in combination with texture complexity grading through streaming loading of the blocks and the overlapping bands to a GPU (Graphic Processing Unit); and executing global color micro-correction in a Lab color space, and generating a seamless high-definition image by applying Poisson fusion in an overlapping zone region. According to the method, boundary artifacts, texture fractures and color deviations caused by blocking processing can be eliminated, video memory occupation is reduced, and calculation efficiency is optimized.

Description

Technical Field

[0001] This application relates to the field of digital cultural heritage protection, specifically to an image super-resolution reconstruction method and system. Background Technology

[0002] In the field of digital cultural heritage preservation, Super-Resolution (SR) of high-resolution murals (such as 50,000×50,000 pixels) is a key technology for achieving digital archiving and restoration.

[0003] Current mainstream super-resolution models (such as ESRGAN and RCAN) require loading full-size images directly for computation. Taking a 50,000×50,000 pixel mural as an example, 4× super-resolution would require approximately 447GB of video memory (calculation formula: 50,000×50,000×4×4 bytes / 1024³≈447GB), which far exceeds the video memory capacity of consumer-grade GPUs (typical value 12-24GB). This means that existing solutions can only be implemented using professional-grade GPUs, resulting in high equipment costs and making it difficult to popularize them in small and medium-sized cultural heritage institutions.

[0004] To overcome hardware limitations, existing technologies generally employ image segmentation processing strategies (such as CR-SRGAN and RTMISR). However, these methods have significant drawbacks: independent super-resolution at the segment boundaries creates seams (such as broken figure outlines and abrupt color changes), damaging the integrity of the mural art; local color statistical differences in different segments lead to overall color distortion (such as local oversaturation / grayish dark areas); and obvious visual breaks occur in continuous decorative areas (such as cloud patterns and wavy lines), affecting the accuracy of restoration. Actual test data shows that after segmentation processing, the PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) of the mural both decrease, failing to meet the acceptable threshold for cultural relic restoration (generally PSNR>35dB, SSIM>0.95).

[0005] Meanwhile, existing block-segmentation schemes employ static model strategies (such as using a strong model for the entire image), which still call heavy models for low-texture areas such as flat backgrounds, resulting in computational redundancy (such as processing time of more than 12 hours for a single image). Furthermore, critical areas (such as human faces and fine patterns) suffer from blurred details (such as the loss of eyelashes and hair strands) due to insufficient model capabilities. Industry trends show that manual intervention in block segmentation and post-processing accounts for more than 60% of the total time required for cultural relic digitization projects.

[0006] In summary, the existing technical approaches have three major limitations: direct processing with professional-grade GPUs is only suitable for large institutions due to the excessive memory requirements; static segmentation combined with a unified model results in boundary artifacts, color breaks, and low efficiency; manual intervention in segmentation relies on expert experience, is time-consuming, and has poor consistency. None of these can meet the urgent needs of cultural relic restoration for high fidelity, low cost, and automation. Summary of the Invention

[0007] This application provides an image super-resolution reconstruction method and system, which can solve the technical problems existing in the prior art, such as memory bottleneck, block boundary artifacts, and imbalance between model efficiency and fidelity in ultra-large image super-resolution reconstruction.

[0008] In a first aspect, embodiments of this application provide an image super-resolution reconstruction method, the image super-resolution reconstruction method comprising: Construct a multi-resolution image pyramid, which includes multiple levels of low-resolution layers of the original mural image; Semantic analysis is performed based on the low-resolution layer of the multi-resolution image pyramid to identify key structural regions and uniform regions in the image. An adaptive slicing strategy is generated based on the semantic analysis results. The adaptive slicing strategy includes dividing the image into multiple slices and setting an overlap band between adjacent slices. The current slice and the overlapping band are streamed to the processing unit, the texture complexity of each slice is calculated, and a super-resolution model is dynamically selected to perform super-resolution processing on the slice. Perform global color micro-correction on the super-resolution slices; Gradient fusion is performed in the overlapping area to generate a seamlessly stitched high-definition image.

[0009] In conjunction with the first aspect, in one implementation, the semantic analysis includes: identifying key structural regions based on edge detection; and identifying uniform regions based on region segmentation.

[0010] In conjunction with the first aspect, in one embodiment, the streaming loading step includes: when processing the current slice, loading only the current slice and the overlapping strip to the processing unit; and immediately releasing the memory occupied by the current slice and the overlapping strip after processing is completed.

[0011] In conjunction with the first aspect, in one implementation, the texture complexity calculation includes: calculating the image entropy of the slice; and determining the texture complexity level based on the image entropy.

[0012] In conjunction with the first aspect, in one implementation, the dynamic selection of the super-resolution model includes: selecting a lightweight super-resolution model when the texture complexity level is low; and selecting a powerful super-resolution model when the texture complexity level is high.

[0013] In conjunction with the first aspect, in one implementation, the global color micro-correction includes: calculating global statistics for the a-channel and b-channel of an image in the Lab color space; and fine-tuning the a-channel and b-channel of a slice based on the global statistics.

[0014] In conjunction with the first aspect, in one implementation, the gradient fusion includes: applying the Poisson equation to solve for gradient continuity in the overlapping region.

[0015] In conjunction with the first aspect, in one implementation, the adaptive slicing strategy further includes: avoiding setting slicing boundaries in critical structural regions.

[0016] In conjunction with the first aspect, in one embodiment, the method further includes: pre-scaling the original image before constructing the multi-resolution image pyramid.

[0017] Secondly, embodiments of this application provide an image super-resolution reconstruction system, the image super-resolution reconstruction system comprising: Multi-resolution image building unit, used to construct a multi-resolution image pyramid, which includes multiple levels of low-resolution layers of the original mural image; The semantic analysis unit is used to perform semantic analysis based on the low-resolution layer of the multi-resolution image pyramid to identify key structural regions and uniform regions in the image. An adaptive slicing generation unit is used to generate an adaptive slicing strategy based on the semantic analysis results. The adaptive slicing strategy includes dividing the image into multiple slices and setting an overlap band between adjacent slices. A streaming processing unit is used to stream the current slice and the overlapping band to the processing unit, calculate the texture complexity of each slice, and dynamically select a super-resolution model to perform super-resolution processing on the slice. The image correction unit is used to perform global color micro-correction on the super-resolution slices; An image fusion unit is used to perform gradient fusion in the overlapping area to generate a seamlessly stitched high-definition image.

[0018] The beneficial effects of the technical solutions provided in this application include: Semantic analysis is performed on the low-resolution layer using a multi-resolution image pyramid to identify key structural regions and uniform regions, generating an adaptive slicing strategy. This strategy includes avoiding structural edges and prioritizing passage through uniform texture regions, thereby eliminating boundary artifacts and texture breaks caused by slicing. A streaming loading mechanism dynamically loads only the current slice and overlapping bands to the processing unit, and combines texture complexity hierarchy to achieve dynamic model selection, reducing memory usage and optimizing computational efficiency. Global color micro-correction performs local fine-tuning of the a and b channels in Lab space based on global statistics, and ensures gradient continuity in overlapping areas through gradient domain fusion, eliminating slicing color deviation. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of an embodiment of the image super-resolution reconstruction method of this application; Figure 2 This is a schematic diagram of the functional modules of an embodiment of the image super-resolution reconstruction system of this application. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0022] In a first aspect, embodiments of this application provide an image super-resolution reconstruction method.

[0023] In one embodiment, reference is made to Figure 1 , Figure 1 This is a schematic flowchart of an embodiment of the image super-resolution reconstruction method of this application. Figure 1 As shown, image super-resolution reconstruction methods include: Step S1: Construct a multi-resolution image pyramid, which includes multiple levels of low-resolution layers of the original mural image.

[0024] Step S2: Perform semantic analysis based on the low-resolution layer of the multi-resolution image pyramid to identify key structural regions and uniform regions in the image.

[0025] Step S3: Generate an adaptive slicing strategy based on the semantic analysis results. The adaptive slicing strategy includes dividing the image into multiple slices and setting an overlap band between adjacent slices.

[0026] Step S4: Stream the current slice and overlapping band to the processing unit, calculate the texture complexity of each slice, and dynamically select a super-resolution model to perform super-resolution processing on the slice.

[0027] Step S5: Perform global color micro-correction on the super-resolution slices.

[0028] Step S6: Perform gradient fusion in the overlapping area to generate a seamlessly stitched high-definition image.

[0029] In this embodiment, a multi-resolution image pyramid is constructed, and semantic analysis based on the low-resolution layer is performed to generate an adaptive slicing strategy. The boundary avoids key structural regions and prioritizes passing through uniform texture areas. Combined with a streaming loading mechanism and dynamic model selection based on texture complexity hierarchy, as well as global color micro-correction and gradient domain fusion in Lab color space, the memory bottleneck, block boundary artifacts, color deviation, and imbalance between model efficiency and fidelity in image super-resolution reconstruction in related technologies are solved. The semantic analysis accurately plans the slicing boundary to eliminate structural fractures. The streaming loading mechanism avoids full-image memory explosion, achieves consumer-grade GPU adaptation, texture hierarchy, dynamic allocation of model resources to ensure details in key areas, and global color correction and gradient fusion to ensure color consistency and seamless stitching.

[0030] Furthermore, in one embodiment, the semantic analysis includes: identifying key structural regions based on edge detection; and identifying uniform regions based on region segmentation.

[0031] In this embodiment, the edge detection algorithm can accurately identify key structural areas such as the outlines of figures and boundaries in the mural, and the region segmentation algorithm can effectively divide uniform texture areas such as the background of the mural. By planning the block boundaries in uniform areas and avoiding key structural areas, the texture continuity and color consistency at the boundaries are ensured during block processing, thereby eliminating structural breaks and boundary artifacts.

[0032] By identifying key structural regions based on edge detection and identifying uniform regions based on region segmentation, the block boundary artifact problem in image super-resolution reconstruction in related technologies is solved, and block boundaries are avoided in key structural regions to eliminate structural breaks.

[0033] Furthermore, in one embodiment, the above-described streaming loading step includes: when processing the current slice, only the current slice and the overlapping band are loaded into the processing unit. The memory occupied by the current slice and the overlapping band is released immediately after processing is completed.

[0034] In this embodiment, by loading only the current slice and the overlapping band into the processing unit and releasing the memory immediately after processing, the memory bottleneck problem of image super-resolution reconstruction in related technologies is solved, and the cost of the processing unit is reduced.

[0035] Furthermore, in one embodiment, the texture complexity calculation includes: calculating the image entropy of the slice; and determining the texture complexity level based on the image entropy.

[0036] In this embodiment, image entropy is a standard metric in information theory used to quantify the complexity of image texture (higher entropy values ​​indicate more complex textures, while lower entropy values ​​indicate simpler textures). Image entropy thresholds are used to classify complexity (e.g., entropy < 0.5 indicates low complexity, ≥ 0.5 indicates high complexity), which is directly mapped to the model selection strategy. Specifically, low-entropy regions (such as flat backgrounds) do not require high-precision models, so lightweight models with low computational cost are selected. High-entropy regions (such as faces or patterns) require high-precision models to preserve details; therefore, powerful yet lightweight models are selected.

[0037] By calculating the image entropy of the segments and determining the texture complexity level based on the image entropy, the problem of the imbalance between model efficiency and fidelity in image super-resolution reconstruction in related technologies is solved. The model resources are dynamically allocated through texture hierarchy to ensure details in key areas and optimize the processing speed of non-key areas.

[0038] Furthermore, in one embodiment, the aforementioned global color micro-correction includes: calculating global statistics for the a-channel and b-channel of the image in the Lab color space; and fine-tuning the a-channel and b-channel of the slice based on the aforementioned global statistics.

[0039] In this embodiment, the color perception characteristics of the Lab color space are utilized to quantify the overall color distribution of the mural through global statistics (such as mean and standard deviation), so that the a and b channels of each segment are matched with the global color features, thereby achieving gradient continuity in the overlapping area and completely eliminating the color deviation caused by segmentation processing.

[0040] By calculating the global statistics of the a and b channels in the Lab color space and fine-tuning the a and b channels of the segmented mural based on the global statistics, the color inconsistency problem in image super-resolution reconstruction in related technologies is solved, the local oversaturation and graying of dark areas caused by segmentation processing are eliminated, and the overall color consistency of the mural is ensured.

[0041] Furthermore, in one embodiment, the gradient fusion includes: applying the Poisson equation to solve for gradient continuity in the overlapping region.

[0042] In this embodiment, by solving the Poisson equation, a smooth gradient transition in the overlapping areas is ensured, allowing the super-resolution blocks to seamlessly merge at the boundaries and guaranteeing the overall visual consistency of the mural. This solves the boundary artifact problem in image super-resolution reconstruction in related technologies and eliminates seams and color abrupt changes caused by block processing.

[0043] Furthermore, in one embodiment, the above adaptive slicing strategy further includes: avoiding setting slicing boundaries in critical structural regions.

[0044] In this embodiment, based on the key structural regions (such as human contours and facial features) identified through semantic analysis, the block boundaries are dynamically planned and set only in areas with uniform texture (such as the background of a mural). This ensures the continuity of key structures during super-resolution processing, thereby eliminating boundary artifacts such as broken human contours and abrupt color changes, and protecting the integrity of the mural art. This solves the problem of block boundary artifacts in image super-resolution reconstruction in related technologies and eliminates structural breaks.

[0045] Furthermore, in one embodiment, the above method further includes: pre-scaling the original image before constructing the multi-resolution image pyramid.

[0046] In this embodiment, by scaling the original image to a power of 2 (e.g., 50000×50000 → 51200×51200), resolution mismatch during pyramid construction is avoided. This allows lower-resolution layers to more accurately preserve key structural features, providing high-quality input for semantic analysis and ultimately achieving precise alignment between the block boundaries and the artistic structure. This addresses the insufficient semantic analysis accuracy in image super-resolution reconstruction in related technologies, ensuring a reasonable resolution distribution across pyramid levels, improving the accuracy of key structural region recognition, thereby optimizing the adaptive slicing strategy and eliminating block boundary artifacts.

[0047] In one specific embodiment, by constructing a globally perceptive-guided adaptive streaming super-resolution reconstruction framework, an intelligent processing paradigm that progresses from coarse to fine and is content-aware is realized. This deeply integrates the global structural information of the mural into the entire process of segmentation, reconstruction, and fusion, effectively solving the problems of memory bottlenecks, boundary artifacts, and efficiency imbalances in the super-resolution reconstruction of ultra-high resolution murals.

[0048] First, the original ultra-high resolution mural image is downsampled to generate multi-level low-resolution preview images, constructing a multi-resolution image pyramid. Based on semantic analysis of the pyramid, a segmentation plan is generated, employing a streaming data loading mechanism: only the current segment and its overlapping portion are dynamically loaded into GPU memory, and immediately released and the next segment loaded after processing. This mechanism transforms fixed memory requirements into controllable, flowing small-block occupancy, completely overcoming hardware memory limitations.

[0049] In the low-resolution layer, traditional algorithms are used to identify key structural regions (such as human outlines) and uniform texture regions (such as background color). When generating the slicing strategy, the cutting lines are ensured to avoid key structural regions and preferentially pass through uniform texture regions. Each slice contains overlapping bands with adjacent slices, providing necessary data support for subsequent elimination of boundary seams.

[0050] For each slice, the image entropy is calculated to quantize the texture complexity: regions with high entropy values ​​(such as faces and patterns) are classified as high-complexity regions, while regions with low entropy values ​​(such as uniform backgrounds) are classified as low-complexity regions. Lightweight models are used to accelerate the processing of low-complexity regions, while powerful lightweight models are used for high-complexity regions, achieving a dynamic balance between efficiency and fidelity.

[0051] The low-resolution image is converted to the Lab color space, and global statistics (such as mean and standard deviation) for the a and b channels are calculated. After super-resolution, the image is segmented and local color fine-tuned based on the global statistics to eliminate color deviations in the segments. A Poisson fusion algorithm is applied to the overlapping areas, and gradient continuity constraints ensure seamless stitching, completely eliminating abrupt changes in brightness and color, and preserving the integrity of the mural art.

[0052] In summary, this invention enables super-resolution of ultra-large mural images with hundreds of millions of pixels under limited video memory (such as a single consumer-grade GPU) through a streaming block processing mechanism based on a multi-resolution pyramid, thus breaking through the hardware limitations on the application scope of the algorithm.

[0053] This invention avoids cutting at the edges of important structures by using semantically guided adaptive slicing, and by combining gradient domain fusion technology, it alleviates the boundary artifacts and texture breaks caused by traditional slicing methods, thus ensuring the overall artistic integrity of the mural after super-resolution.

[0054] This invention employs an image hierarchical processing strategy based on texture complexity recognition to intelligently classify image regions and assign models of varying complexity, achieving more efficient super-resolution. This mechanism significantly reduces the overall computational load while ensuring high-quality reconstruction of critical details, making rapid processing of images with hundreds of millions of pixels feasible.

[0055] This invention effectively corrects local color deviations that may be caused by independent processing of blocks by using local color correction based on global color statistics, thus ensuring the uniformity and authenticity of the colors of the entire mural.

[0056] This invention integrates the entire process from global analysis, intelligent segmentation, hierarchical processing to seamless fusion, and can adaptively complete the super-resolution task of ultra-large mural images without manual intervention, providing a reliable tool for the batch digitization super-resolution of cultural mural heritage.

[0057] Secondly, embodiments of this application also provide an image super-resolution reconstruction system.

[0058] In one embodiment, reference is made to Figure 2 , Figure 2 This is a functional module diagram of an embodiment of the image super-resolution reconstruction system of this application. Figure 2 As shown, the image super-resolution reconstruction system includes: Multi-resolution image building unit 1 is used to build a multi-resolution image pyramid, which includes multiple levels of low-resolution layers of the original mural image.

[0059] Semantic analysis unit 2 is used to perform semantic analysis based on the low-resolution layer of the multi-resolution image pyramid mentioned above, and to identify key structural regions and uniform regions in the image.

[0060] The adaptive segmentation generation unit 3 is used to generate an adaptive segmentation strategy based on the semantic analysis results. The adaptive segmentation strategy includes segmenting the image into multiple segments and setting an overlap band between adjacent segments.

[0061] The streaming processing unit 4 is used to stream the current slice and the aforementioned overlapping band to the processing unit, calculate the texture complexity of each slice, and dynamically select a super-resolution model to perform super-resolution processing on the slice.

[0062] Image correction unit 5 is used to perform global color micro-correction on the super-resolution slices.

[0063] Image fusion unit 6 is used to perform gradient fusion in the aforementioned overlapping area to generate a seamlessly stitched high-definition image.

[0064] In this embodiment, a multi-resolution image pyramid is constructed, and semantic analysis based on the low-resolution layer is performed to generate an adaptive slicing strategy. The boundary avoids key structural regions and prioritizes passing through uniform texture areas. Combined with a streaming loading mechanism and dynamic model selection based on texture complexity hierarchy, as well as global color micro-correction and gradient domain fusion in Lab color space, the memory bottleneck, block boundary artifacts, color deviation, and imbalance between model efficiency and fidelity in image super-resolution reconstruction in related technologies are solved. The semantic analysis accurately plans the slicing boundary to eliminate structural fractures. The streaming loading mechanism avoids full-image memory explosion, achieves consumer-grade GPU adaptation, texture hierarchy, dynamic allocation of model resources to ensure details in key areas, and global color correction and gradient fusion to ensure color consistency and seamless stitching.

[0065] In one specific embodiment, the multi-resolution image construction unit 1 receives an ultra-high-resolution original image of a mural with hundreds of millions of pixels as input. First, through a series of downsampling operations, it generates an image pyramid consisting of the original image and its multiple low-resolution versions. These preview images at different scales provide a global view for subsequent semantic analysis and processing planning.

[0066] The semantic analysis unit 2 receives the low-resolution layer of the multi-resolution image pyramid as input. At the lower resolution pyramid level, it uses an efficient edge detection algorithm to quickly identify the key structural contours (such as the outlines of figures and objects) and uniform texture areas (such as the wall background) in the mural, providing a basis for subsequent segmentation strategies.

[0067] The adaptive segmentation generation unit 3 receives the semantic analysis results as input and generates an adaptive segmentation strategy based on the recognition results of key structural regions and uniform texture regions. This strategy ensures that the segmentation boundaries avoid key structural regions and preferentially pass through uniform texture regions, while also generating a segmentation plan that includes overlapping bands.

[0068] The streaming processing unit 4 receives the adaptive slicing strategy as input, performs streaming data loading, dynamically loads only the current slice and overlapping bands to the GPU processing unit, and releases the memory immediately after processing. At the same time, it calculates the image entropy of each slice and dynamically selects a lightweight model (low complexity region) or a powerful lightweight model (high complexity region) for super-resolution processing based on the texture complexity level.

[0069] Image correction unit 5 receives the super-resolution slice as input, calculates global a / b channel statistics in Lab color space, and performs local fine-tuning of the a and b channels of the slice based on the above statistics to eliminate color deviation caused by slice processing.

[0070] Image fusion unit 6 receives super-resolution and corrected segments as input, applies the Poisson equation to solve for gradient continuity in the overlapping region, and achieves pixel-level seamless stitching by maintaining the continuity of image gradient, thus completely eliminating abrupt changes in brightness and color.

[0071] The functions of each module in the above-mentioned image super-resolution reconstruction system correspond to the steps in the above-mentioned image super-resolution reconstruction method embodiment, and their functions and implementation processes will not be described in detail here.

[0072] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0073] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0074] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0075] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0076] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0077] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0078] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. An image super-resolution reconstruction method, characterized in that, The image super-resolution reconstruction method includes: Construct a multi-resolution image pyramid, which includes multiple levels of low-resolution layers of the original mural image; Semantic analysis is performed based on the low-resolution layer of the multi-resolution image pyramid to identify key structural regions and uniform regions in the image. An adaptive slicing strategy is generated based on the semantic analysis results. The adaptive slicing strategy includes dividing the image into multiple slices and setting an overlap band between adjacent slices. The current slice and the overlapping band are streamed to the processing unit, the texture complexity of each slice is calculated, and a super-resolution model is dynamically selected to perform super-resolution processing on the slice. Perform global color micro-correction on the super-resolution slices; Gradient fusion is performed in the overlapping area to generate a seamlessly stitched high-definition image.

2. The image super-resolution reconstruction method as described in claim 1, characterized in that, The semantic analysis includes: identifying key structural regions based on edge detection; and identifying uniform regions based on region segmentation.

3. The image super-resolution reconstruction method as described in claim 1, characterized in that, The streaming loading step includes: when processing the current slice, only the current slice and the overlapping band are loaded into the processing unit; and the memory occupied by the current slice and the overlapping band is released immediately after processing is completed.

4. The image super-resolution reconstruction method as described in claim 1, characterized in that, The texture complexity calculation includes: calculating the image entropy of the slice; and determining the texture complexity level based on the image entropy.

5. The image super-resolution reconstruction method as described in claim 1, characterized in that, The dynamic selection of the super-resolution model includes: selecting a lightweight super-resolution model when the texture complexity level is low; and selecting a powerful super-resolution model when the texture complexity level is high.

6. The image super-resolution reconstruction method as described in claim 1, characterized in that, The global color micro-correction includes: calculating global statistics for the a-channel and b-channel of the image in the Lab color space; and fine-tuning the a-channel and b-channel of the slice based on the global statistics.

7. The image super-resolution reconstruction method as described in claim 1, characterized in that, The gradient fusion includes applying the Poisson equation to solve for gradient continuity in the overlapping region.

8. The image super-resolution reconstruction method as described in claim 1, characterized in that, The adaptive slicing strategy further includes: avoiding setting slicing boundaries in critical structural regions.

9. The image super-resolution reconstruction method as described in claim 1, characterized in that, The method further includes pre-scaling the original image before constructing the multi-resolution image pyramid.

10. An image super-resolution reconstruction system, characterized in that, The image super-resolution reconstruction system includes: Multi-resolution image building unit, used to construct a multi-resolution image pyramid, which includes multiple levels of low-resolution layers of the original mural image; The semantic analysis unit is used to perform semantic analysis based on the low-resolution layer of the multi-resolution image pyramid to identify key structural regions and uniform regions in the image. An adaptive slicing generation unit is used to generate an adaptive slicing strategy based on the semantic analysis results. The adaptive slicing strategy includes dividing the image into multiple slices and setting an overlap band between adjacent slices. A streaming processing unit is used to stream the current slice and the overlapping band to the processing unit, calculate the texture complexity of each slice, and dynamically select a super-resolution model to perform super-resolution processing on the slice. The image correction unit is used to perform global color micro-correction on the super-resolution slices; An image fusion unit is used to perform gradient fusion in the overlapping area to generate a seamlessly stitched high-definition image.

Images