Image super-resolution edge reconstruction method and system based on morphology and coverage rate perception

Abstract

The invention discloses an image super-resolution edge reconstruction method based on morphology and coverage rate perception. The method comprises the following steps: dividing an image into an edge region and a non-edge region based on depth and brightness information; anti-aliasing edge detail recovery is realized through coverage rate perceived depth mode classification and a table look-up filter; generating direction features by using the brightness gradient and keeping details through a table look-up filter; and combining the reconstruction results of the current frame and the historical frame to ensure the image stability in a dynamic scene. The invention also provides an image super-resolution edge reconstruction system. According to the method provided by the invention, the quality of an object edge and a newly appearing pixel region can be remarkably improved, sawtooth artifacts are effectively inhibited, and relatively high time stability is kept in a dynamic scene, so that the requirements of real-time rendering and visual enhancement application are met.

Description

Technical Field

[0001] This invention belongs to the fields of computer graphics and real-time rendering technology, and particularly relates to an image super-resolution edge reconstruction method and system based on morphology and coverage awareness. Background Technology

[0002] In recent years, temporal accumulation-based anti-aliasing and super-resolution methods have gradually developed into commercial products. These methods utilize inter-frame temporal consistency to introduce information from historical frames into the current frame, thereby approximating a higher sampling rate within a limited sampling budget. Representative solutions include AMD's FSR3 / FSR4 and NVIDIA's DLSS3 / DLSS4, both based on temporal anti-aliasing (TAA), generating high-quality results by accumulating and reusing pixel information across multiple frames. AMD's FSR3 combines frame interpolation with spatial super-resolution, significantly improving the output frame rate while maintaining low computational overhead; the latest FSR4 further improves edge processing and temporal stability, reducing the probability of ghosting and motion blur artifacts. NVIDIA's DLSS3, leveraging dedicated optical flow estimation hardware and deep learning models, introduces frame generation technology while achieving super-resolution, significantly improving image smoothness; DLSS4 further expands the training data and model structure, further improving edge sharpness and overall stability in dynamic scenes. Although the above methods achieve a good balance between image quality and performance on high-performance hardware platforms, their computational complexity and hardware dependence are high. They usually require the support of a dedicated graphics card and AI acceleration unit, making it difficult to run in real time on mobile devices or low-power devices.

[0003] Patent document CN117437110A discloses a hardware acceleration method and system for image super-resolution reconstruction based on in-memory indexing. It replaces the linear pixel block index with cross-shaped pixel blocks and uses a high-resolution pixel block arrangement of r×r with rotation in four directions of 0°, 90°, 180° and 270° to replace the rotation of the entire image. This simplifies the rotation operation of image adaptive enhancement, which is difficult to implement in hardware, and greatly saves the amount of data read and written, making it easier to allocate hardware resources more fully and rationally.

[0004] Patent document CN118505505A discloses an image super-resolution reconstruction method, apparatus, electronic device, and storage medium. The method includes: acquiring an image to be super-resolution reconstructed; inputting the image to be super-resolution reconstructed into a pre-trained image super-resolution reconstruction model to obtain a super-resolution reconstructed image of the image to be super-resolution reconstructed; the image super-resolution reconstruction model includes a convolutional neural network using a target convolution mode, wherein the target convolution mode is used to rotate a first-size rotational convolution kernel a preset number of times to simulate a second-size receptive field, the second size being larger than the first size, and the number of pixels processed by the rotational convolution kernel in a single operation being N, where N is less than 4. Summary of the Invention

[0005] The present invention aims to provide an image super-resolution edge reconstruction method and system based on morphology and coverage awareness, which can achieve efficient image super-resolution reconstruction on mobile devices and low-power devices.

[0006] To achieve the first objective of this invention, the following technical solution is provided, comprising the following steps: The low-resolution image input at the current time is divided into edge regions and non-edge regions to obtain the corresponding edge region pixel set and non-edge region pixel set; A depth pattern classification is performed on the target pixel block selected from the edge region pixel set and the neighboring pixel blocks of the target pixel block to derive the edge direction vector of the target pixel block and generate the corresponding direction code based on the edge direction vector; The lookup table pre-built using the directional encoding index is used to output the first index result, and the coverage sensing filter corresponding to the edge is constructed using the first index result. For the brightness difference of each pixel block in the non-edge region pixel set, a corresponding gradient vector is constructed and a corresponding feature code is generated based on the gradient vector; The lookup table is indexed using the feature encoding to output a second index result, and a reconstruction filter for the corresponding feature is constructed using the second index result; The coverage-aware filter and the reconstruction filter are used to reconstruct high-resolution pixels in the corresponding region of the low-resolution image to generate an anti-aliased high-resolution image. The initial high-resolution image at the current moment is accumulated and fused with the high-resolution image at the previous moment to update each pixel with weight, thereby outputting a high-resolution reconstructed image.

[0007] This invention divides an image into edge and non-edge regions based on depth and brightness information; secondly, it performs edge reconstruction by restoring anti-aliased edge details through coverage-aware depth pattern classification and lookup table filters; thirdly, it performs non-edge reconstruction by generating directional features using brightness gradients and preserving details through lookup table filters; and finally, it performs temporal accumulation fusion by combining the reconstruction results of the current frame with those of historical frames to ensure image stability in dynamic scenes.

[0008] Specifically, the process of segmenting the low-resolution image is as follows: The depth values ​​of four pixels within a target 2×2 pixel block are compared. When the depth difference between pixels exceeds a preset threshold, the target pixel block is determined to be an edge region. When the depth difference does not exceed the threshold, the target pixel block is determined to be a non-edge region.

[0009] Specifically, the coverage-aware filter is constructed using a coverage-aware weighted fusion strategy.

[0010] Specifically, the expression for the coverage-aware filter is as follows:

[0011] in, Represents the reconstructed high-resolution pixels The color, Represents the normalization factor. Represents low-resolution neighborhood samples The set, Sampling points representing reconstruction The color, For space core, The amplitude kernel is modulated by the coverage rate.

[0012] Specifically, when deriving the edge direction vector of the target pixel block, the receptive field is set to... Low-resolution regions are extracted, and nine overlapping regions are generated within the receptive field. Depth block orientation encoding enhances the robustness of edge orientation estimation; Based on the orthogonality and saliency criteria, two neighboring edge direction vectors orthogonal to the edge direction vector of the target pixel block are selected, and the three edge direction vectors are used as a triple to represent the local edge.

[0013] Specifically, for the set of pixels in non-edge regions, an upsampling strategy is used to obtain the feature encoding of each pixel block. The upsampling strategy includes: Calculate the brightness map of a low-resolution image; The brightness difference between the horizontal and vertical directions of each pixel block is calculated based on the brightness map to output the corresponding gradient vector; Without explicitly calculating the angle, the gradient direction and intensity level are obtained from the gradient vector using the ratio relationship, and the non-edge labels, gradient directions and intensity levels are combined into triples to represent the corresponding feature codes.

[0014] Specifically, the time accumulation strategy includes: The super-resolution image from the previous moment Based on rendering motion vectors Reverse deformation to the current viewpoint yields the deformed image from the previous moment. And on Perform bounding box trimming; The pixel-wise fusion weights are calculated based on three types of confidence cues, including motion amplitude, occlusion mask, and distance confidence. Based on the cropped The high-resolution reconstructed image at that moment is calculated by combining pixel-by-pixel fusion weights.

[0015] Specifically, the expression for the high-resolution reconstructed image is as follows: ; ; in, Reconstruct a high-resolution image for the current frame. Indicates mixed weights, Indicates the amplitude of motion. Indicates the masking code. Indicates the distance confidence level. This represents a constant term.

[0016] To achieve the second objective of this invention, the following technical solution is provided: an image super-resolution edge reconstruction system, used to perform the steps of the above-described morphology- and coverage-aware image super-resolution edge reconstruction method, comprising: Edge region reconstruction unit, used for anti-aliasing reconstruction of edge regions; Non-edge region upsampling unit, used for high-resolution reconstruction of non-edge regions; The temporal accumulation unit outputs the reconstruction result of the current frame based on the edge region reconstruction unit and the non-edge region upsampling unit. It performs motion vector alignment and pixel-by-pixel weighted fusion with historical frames to output the final high-resolution reconstructed image.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: An edge reconstruction method is proposed, which can accurately recover the sub-pixel coverage information of object boundaries under low-resolution input conditions, thereby effectively eliminating jagged artifacts and improving the clarity and visual quality of edge regions. By using Directional Encoding and Lookup Table (LUT), the filtering weights of edge and non-edge regions are learned and stored offline. During the inference stage, high-resolution reconstruction can be completed simply by looking up the table, which greatly reduces the computational overhead. It has good real-time performance and lightweight features, making it suitable for mobile and low-power devices. Attached Figure Description

[0018] Figure 1 A flowchart of the image super-resolution edge reconstruction method based on morphology and coverage awareness provided in this embodiment; Figure 2 This is a schematic diagram of edge and non-edge region feature encoding provided in this embodiment; Figure 3 This is a schematic diagram of super-resolution image reconstruction provided in this embodiment; Figure 4This is a comparison chart of the results of the present invention solution and the FSR1 method provided in this embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown in this embodiment, an image super-resolution edge reconstruction method based on morphology and coverage awareness is provided, which can achieve efficient image super-resolution reconstruction on mobile devices and low-power devices. While ensuring low computational overhead, this invention can significantly improve the quality of object edges and newly emerging pixel regions, effectively suppress jagged artifacts, and maintain high temporal stability in dynamic scenes, thereby meeting the needs of real-time rendering and visual enhancement applications.

[0021] The input low-resolution image is divided into regions.

[0022] In this step, a low-resolution image generated by the rendering pipeline is first input, and then, combined with the corresponding depth and brightness maps, each pixel is classified into regions. Specifically, when local... If a pixel block contains a significant depth difference, the pixel is determined to be an edge region; if no significant depth difference exists, the pixel is determined to be a non-edge region. The determination result is represented in encoded form, where... Indicates non-edge regions. Indicates the edge region.

[0023] By performing the above region segmentation operation, we can distinguish edge pixels that require fine anti-aliasing processing from non-edge pixels that can be processed by efficient upsampling, thus laying the foundation for differential filtering in subsequent steps.

[0024] High-resolution reconstruction of the edge regions.

[0025] After region segmentation, for pixels identified as edges, their colors need to be estimated at the sub-pixel level to restore the smoothness and accuracy of object boundaries. The basic idea is to determine the relative proportion of the target pixel to its adjacent surfaces based on coverage awareness, and then use this proportion to perform weighted fusion of pixel colors.

[0026] like Figure 2 As shown, after edge detection is completed, the pixel closest to the target pixel is selected. The depth sampling block determines the directional pattern of the boundary crossing within the block by comparing the relative depth relationships of four sampling points. To standardize encoding and facilitate subsequent table lookups, this embodiment categorizes local boundary patterns into 12 directional types: four orthogonal directions (up, down, left, and right) and two coverage patterns (e.g., "shallow coverage / deep coverage") corresponding to each of the four diagonal directions (upper left, upper right, lower left, and lower right), totaling 12 categories. Each category corresponds to a discrete index, denoted as... In terms of implementation, we can first... A three-valued relation matrix (near / far / equal) is constructed based on deep relationships. Then, a predefined mapping table normalizes all combinations to the aforementioned 12 directional types. This mapping eliminates redundancy caused by specific numerical differences, ensuring that different combinations fall into the same directional category. Finally, the target pixel obtains its directional index. .

[0027] However, based on only one The depth relationships of small blocks are often insufficient to stably describe edge structures. Therefore, this embodiment further expands the receptive field to... A region is defined, in which multiple overlapping small patch patterns are generated, and the two most representative orthogonal neighbors are selected based on their directional consistency and gradient magnitude, along with the center block index. Together they form a triplet This triplet can more comprehensively reflect the local geometric features of the edge.

[0028] After obtaining the directional encoding, this embodiment does not directly invoke a complex neural network. Instead, it pre-stores the filter weights obtained during offline training in a lookup table (LUT). During the inference phase, the system only needs to look up the corresponding weights in the LUT based on the directional triples to quickly obtain them and apply them to local low-resolution sampling points, thereby generating high-resolution edge pixels with anti-aliasing properties. This approach significantly reduces the computational cost of real-time inference while ensuring edge quality.

[0029] More specifically, the edge region color reconstruction employs a coverage-aware weighted fusion strategy, which can be equivalently described as a type of bilateral filtering:

[0030] in, Represents the reconstructed high-resolution pixels The color, Represents the normalization factor. Represents low-resolution neighborhood samples The set, Sampling points representing reconstruction The color, For space core, The amplitude kernel is modulated by the coverage rate.

[0031] in Take the Gaussian kernel:

[0032] in Sampling points To reconstruct pixels The Euclidean distance at the center For spatial kernel scale High-resolution reconstruction is performed on non-edge regions.

[0033] Amplitude kernel The coverage of each surface by the target pixel is determined: if the sampling points Sampling on the surface Above, that is ,but If sampling points Sampling on the surface Above, that is ,but ,in and Derived from local depth patterns to reflect and For reconstructed pixels The relative coverage ratio.

[0034] This coverage ratio is obtained by classifying pixel blocks by depth pattern: within edge regions, the nearest pixel is selected. of Four sampling points of low-resolution depth sampling block The depth relationships are ternary encoded, and all combinations are categorized into 12 typical edge direction types (including top, bottom, left, right, and four diagonal directions and their shallow / deep cover variants) according to a pre-built mapping table. The direction category corresponding to the center block is denoted as... The mapping table is stored in RAM for fast lookup.

[0035] When the result of the region division is When the value is 0, it indicates that the current pixel is in a non-edge region.

[0036] In this case, deep pattern classification is not required; instead, feature encoding and filtering weight lookup are performed based on brightness gradients.

[0037] Specifically, in this embodiment, the brightness map of the input low-resolution image is first calculated, and the brightness difference in the horizontal and vertical directions is calculated in each pixel block to obtain the gradient vector. To meet the requirements of real-time operation, this invention employs a lightweight direction encoding method, which determines the gradient direction using ratio relationships without explicitly calculating angles. With gradient intensity level .

[0038] More specifically, an upsampling strategy based on brightness gradients is applied to non-edge regions, including: (1) Calculate the brightness map of the input low-resolution image; (2) Calculate the brightness difference in the horizontal and vertical directions for each pixel block to obtain the gradient vector. ; (3) Without explicitly calculating the angle, the gradient direction can be obtained using the ratio relationship. With strength level Forming a pattern descriptor .

[0039] According to the method, in step 4, the pattern descriptor Used for indexing the third multilayer perceptron The prediction is used to forecast the filtering weights for non-edge regions; the prediction results are pre-calculated and stored as a lookup table (LUT) in the offline stage, and directly retrieved in the inference stage using a lookup table; the LUTs for edge and non-edge regions are designed with a unified structure, and the index range is [insert range here]. With an output dimension of 64, a single LUT can efficiently process high-resolution images, including both edge and non-edge components. reconstruction.

[0040] Through the above steps, the feature encoding of non-edge regions is unified into triples. .

[0041] The triplet and Figure 2 The edge regions are encoded with a consistent structure, allowing edge and non-edge regions to share the same lookup table (LUT) structure.

[0042] The entire frame of high-resolution image is reconstructed based on a lookup table.

[0043] To enhance the robustness of edge orientation estimation, the receptive field is extended to Low-resolution region, and generate 9 overlapping regions within that region. Depth block orientation encoding; based on this, adaptive selection and... Two orthogonal neighborhood directions and Forming directional triplet encoding To characterize local edge structures.

[0044] In actual implementation, all The combinations of values ​​are mapped to filter weights generated during offline training and stored in a three-dimensional lookup table. During inference, the corresponding filter weights can be obtained simply by looking up the table, thus enabling efficient high-resolution reconstruction of non-edge regions. Through this design, the present invention significantly reduces the computational overhead of non-edge regions while ensuring detail fidelity, providing support for overall lightweight design and real-time performance.

[0045] like Figure 3 As shown, feature encodings of edge regions and non-edge regions are obtained. Subsequently, in this embodiment, it is used as an index to access the uniformly constructed 3D lookup table (LUT) to obtain the corresponding filter weights. This LUT is generated offline by training a multilayer perceptron and stores the filters for edge and non-edge regions as a unified structure, with an index range of... The output dimension is 64. During the inference phase, the reconstruction of each high-resolution pixel requires only one table lookup operation to obtain the weight vector. Pixel color estimation is completed by combining adjacent low-resolution sampling points:

[0046] By following the steps above, high-resolution images can be quickly reconstructed across the entire frame, achieving anti-aliasing and detail-preserving results.

[0047] To reduce redundant queries and improve efficiency, Output of a single LUT access at upsampling rate To simultaneously reconstruct four adjacent high-resolution pixels, the weights are pre-calculated values ​​related to the location.

[0048] In practical use, the criteria for determining non-edge regions are: when the local area is... There is no significant depth difference in the depth sampling block, and the direction encoding is correct. Record it as 0, and mark the region as a non-edge region.

[0049] Improve stability in dynamic scenarios by accumulating time.

[0050] like Figure 3Furthermore, to mitigate flickering and artifacts that may occur in dynamic scenes, this embodiment introduces a temporal accumulation module based on spatial reconstruction. Specifically, the super-resolution image of the previous frame is used to accumulate the temporal accumulation module. Based on motion vectors By changing the perspective to the current frame, we can obtain... To avoid noise introduced by misalignment, the result was cropped using an all-boundary box (AABB). Subsequently, pixel-by-pixel fusion weights were calculated based on three types of confidence cues. Firstly, the amplitude of motion. Firstly, smaller motion indicates more reliable historical information; secondly, occlusion masking. The new display pixel is set to 1, and the others are set to 0. Distance confidence level. , defined as the distance from the high-resolution pixel to the nearest low-resolution sampling point; the smaller the distance, the higher the reliability. The fusion weight is defined as:

[0051] in, Reconstruct a high-resolution image for the current frame. Indicates mixed weights, Indicates the amplitude of motion. Indicates the masking code. Indicates the distance confidence level.

[0052] The final result is an image reconstructed from the current frame. Image aligned with history The final result is obtained by weighted fusion:

[0053] By accumulating over time, this invention can improve the temporal consistency of video sequences while maintaining spatial sharpness, thereby maintaining stable, high-quality super-resolution effects in dynamic scenes. like Figure 4 The image shown is a comparison chart between the solution provided in the above embodiments and the conventional FSR1 method. Figure 4 It can be seen that the method of the present invention is more effective than the conventional FSR1 method in edge reconstruction, that is, the jaggedness of the edges is weaker.

[0054] This embodiment also provides an image super-resolution edge reconstruction system for performing the steps of the morphology- and coverage-aware image super-resolution edge reconstruction method provided in the above embodiments, including: (1) Edge region reconstruction unit, used to generate directional coding based on pixel block depth pattern The coverage-aware filter is then obtained by looking up a table to achieve anti-aliasing reconstruction of edge regions. (2) Non-edge region upsampling unit, used to generate mode descriptors based on brightness gradient and obtain filters by looking up tables, so as to realize high-resolution reconstruction of non-edge regions; (3) Time accumulation unit, used to align the current frame reconstruction result with the historical frame with motion vectors and perform pixel-by-pixel weighted fusion to improve the spatial consistency and temporal stability of the output result.

[0055] Furthermore, the terms "upper," "lower," "inner," "outer," "front," and "rear" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0056] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention should be included in the scope of the claims of the present invention.

[0057] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for image super-resolution edge reconstruction based on morphology and coverage awareness, characterized in that, Includes the following steps: The low-resolution image input at the current time is divided into edge regions and non-edge regions to obtain the corresponding edge region pixel set and non-edge region pixel set; A depth pattern classification is performed on the target pixel block selected from the edge region pixel set and the neighboring pixel blocks of the target pixel block to derive the edge direction vector of the target pixel block and generate the corresponding direction code based on the edge direction vector; The lookup table pre-built using the directional encoding index is used to output the first index result, and the coverage sensing filter corresponding to the edge is constructed using the first index result. For the brightness difference of each pixel block in the non-edge region pixel set, a corresponding gradient vector is constructed and a corresponding feature code is generated based on the gradient vector; The lookup table is indexed using the feature encoding to output a second index result, and a reconstruction filter for the corresponding feature is constructed using the second index result; The coverage-aware filter and the reconstruction filter are used to reconstruct high-resolution pixels in the corresponding region of the low-resolution image to generate an anti-aliased high-resolution image. The initial high-resolution image at the current moment is fused with the high-resolution image at the previous moment over time to update each pixel with weights, thereby outputting a high-resolution reconstructed image.

2. The image super-resolution edge reconstruction method based on morphology and coverage awareness according to claim 1, characterized in that, The process of segmenting the low-resolution image is as follows: The depth values ​​of four pixels within a target 2×2 pixel block are compared. When the depth difference between pixels exceeds a preset threshold, the target pixel block is determined to be an edge region. When the depth difference does not exceed the threshold, the target pixel block is determined to be a non-edge region.

3. The image super-resolution edge reconstruction method based on morphology and coverage awareness according to claim 1, characterized in that, The coverage-aware filter is constructed using a coverage-aware weighted fusion strategy.

4. The image super-resolution edge reconstruction method based on morphology and coverage awareness according to claim 1 or 3, characterized in that, The expression for the coverage-sensing filter is as follows: in, Represents the reconstructed high-resolution pixels The color, Represents the normalization factor. Represents low-resolution neighborhood samples The set, Sampling points representing reconstruction The color, For space core, The amplitude kernel is modulated by the coverage rate.

5. The image super-resolution edge reconstruction method based on morphology and coverage awareness according to claim 1, characterized in that, When deriving the edge direction vector of the target pixel block, the receptive field is set to Low-resolution regions are extracted, and nine overlapping regions are generated within the receptive field. Depth block orientation encoding; Based on the orthogonality and saliency criteria, two neighboring edge direction vectors orthogonal to the edge direction vector of the target pixel block are selected, and the three edge direction vectors are used as a triple to represent the local edge.

6. The image super-resolution edge reconstruction method based on morphology and coverage awareness according to claim 1, characterized in that, An upsampling strategy is employed to obtain the feature encoding of each pixel block within a set of pixels in non-edge regions. This upsampling strategy includes: Calculate the brightness map of a low-resolution image; The brightness difference between the horizontal and vertical directions of each pixel block is calculated based on the brightness map to output the corresponding gradient vector; Without explicitly calculating the angle, the gradient direction and intensity level are obtained from the gradient vector using the ratio relationship, and the non-edge labels, gradient directions and intensity levels are combined into triples to represent the corresponding feature codes.

7. The image super-resolution edge reconstruction method based on morphology and coverage awareness according to claim 1, characterized in that, The time accumulation strategy includes: The super-resolution image from the previous moment Based on rendering motion vectors Reverse deformation to the current viewpoint yields the deformed image from the previous moment. And on Perform bounding box trimming; The pixel-wise fusion weights are calculated based on three types of confidence cues, including motion amplitude, occlusion mask, and distance confidence. Based on the cropped The high-resolution reconstructed image at that moment is calculated by combining pixel-by-pixel fusion weights.

8. The image super-resolution edge reconstruction method based on morphology and coverage awareness according to claim 7, characterized in that, The expression for the high-resolution reconstructed image is as follows: ； ； in, Reconstruct a high-resolution image for the current frame. Indicates mixed weights, Indicates the amplitude of motion. Indicates the masking code. Indicates the distance confidence level. This represents a constant term.

9. An image super-resolution edge reconstruction system, characterized in that, The step of performing the image super-resolution edge reconstruction method based on morphology and coverage awareness as described in any one of claims 1 to 8 includes: an edge region reconstruction unit for anti-aliasing reconstruction of edge regions; Non-edge region upsampling unit, used for high-resolution reconstruction of non-edge regions; The temporal accumulation unit outputs the reconstruction result of the current frame based on the edge region reconstruction unit and the non-edge region upsampling unit. It performs motion vector alignment and pixel-by-pixel weighted fusion with historical frames to output the final high-resolution reconstructed image.

Images