Super-resolution method and system of video, electronic equipment and storage medium

Abstract

The invention provides a video super-resolution method and system, electronic equipment and a storage medium, and relates to the technical field of video processing, and the method comprises the steps: carrying out the character and background separation of a target video, and obtaining a target character region and a target background region; performing geometric flow and texture flow feature enhancement according to the target character region to obtain character region super-division features; performing feature enhancement on the target background region by using an optical flow supervised convolutional neural network to obtain background region super-division features; and performing feature fusion according to the character region super-division features and the background region super-division features to obtain a super-resolution video. According to the method, the characters and the background area of the target video are separated, the structural features and semantic information of the characters are fully considered for super-resolution processing, the background area is also subjected to super-resolution processing, the super-resolution video which is globally clear and rich in character details is finally generated, and the video quality is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of video processing technology, and more specifically to a video super-resolution method, system, electronic device, and storage medium. Background Technology

[0002] In the context of digital transformation, video has become a core medium for information dissemination, widely used in education, scientific research, film and television production, and digital entertainment. As users' demands for video quality continue to rise, subtitle clarity, as a key quality indicator, is receiving increasing attention. An ideal video requires consistent lighting timing and clear text and background details. Video super-resolution technology is an important technique that uses algorithms to reconstruct low-resolution video sequences into high-resolution sequences with rich detail and coherent content.

[0003] However, current video super-resolution technologies still face significant technical bottlenecks when processing videos containing text. As a key carrier of information, text is characterized by its complex structure and delicate strokes, and often forms multi-level spatial interactions and occlusion relationships with the background. Traditional super-resolution methods based on global processing struggle to achieve precise optimization of text regions, easily leading to visual defects such as blurring, jagged edges, and broken strokes, severely impacting information transmission efficiency and video viewing quality. With the development of deep learning technology, data-driven super-resolution methods have made significant progress. Through training on large-scale datasets, architectures such as convolutional neural networks, generative adversarial networks, and Transformers can effectively learn the mapping relationship from low resolution to high resolution. However, existing methods generally suffer from insufficient modeling of text semantic information and contextual relevance, resulting in limited performance in terms of text detail restoration and semantic consistency maintenance. Therefore, in text-dense video scenarios, how to solve the problems of insufficient clarity and readability caused by existing super-resolution methods has become a key technical challenge that urgently needs to be overcome in the field of video processing. Summary of the Invention

[0004] In view of the above-mentioned shortcomings of the prior art, the present invention provides a video super-resolution method, system, electronic device and storage medium, which effectively solves the problems of insufficient clarity and readability caused by the existing super-resolution methods.

[0005] In a first aspect, the present invention provides a video super-resolution method, the method comprising: Separate the target video from the text and background to obtain the target text region and the target background region; Geometric flow and texture flow feature enhancements are performed on the target text region to obtain super-resolution features of the text region. An optical flow-supervised convolutional neural network is used to enhance the features of the target background region, thereby obtaining super-resolution features of the background region; Feature fusion is performed based on the super-resolution features of the text region and the super-resolution features of the background region to obtain a super-resolution video.

[0006] In an optional implementation, separating the target video into text and background regions to obtain target text regions and target background regions includes: Extract text structures of different sizes from each video frame of the target video to obtain initial text features; The initial text features are subjected to progressive upsampling and channel weighting operations to obtain enhanced text features; Adjust the channel weights of the enhanced text features to obtain the target text region and the target background region.

[0007] In an optional implementation, the step of performing geometric flow and texture flow feature enhancement on the target text region to obtain super-resolution features of the text region includes: Adjust the shape of the local receptive field of the text features in the target text region to obtain the text geometric features; Adversarial training is performed on the text features in the target text region to constrain the realism of high-frequency detail generation and obtain text texture features; Adjust the fusion weights of the text geometric features and the text texture features to obtain text fusion features; The semantic rationality of the text fusion features is verified to obtain the super-resolution features of the text region.

[0008] In an optional implementation, the optically flow-supervised convolutional neural network enhances the features of the target background region to obtain super-resolution features of the background region, including: Calculate the initial optical flow field of adjacent video frames in the target background region, and optimize the initial optical flow field to obtain the target optical flow field; Feature extraction is performed on each video frame in the target background region to obtain the video frame super-resolution result; Based on the target optical flow field, the super-resolution results of the video frame are optimized for spatiotemporal consistency to obtain the super-resolution features of the background region.

[0009] In an optional implementation, the step of performing feature fusion based on the super-resolution features of the text region and the super-resolution features of the background region to obtain a super-resolution video includes: Perform geometric alignment and feature alignment on the super-resolution features of the text region and the super-resolution features of the background region; The fusion weights of the super-resolution features of the text region and the super-resolution features of the background region are dynamically assigned based on the text detection confidence and optical flow field. Based on the fusion weights, the super-resolution features of the text region and the super-resolution features of the background region are fused to obtain fused super-resolution features; The super-resolution video is obtained by masking the overlapping areas of text and background in the fused super-resolution features.

[0010] In an optional implementation, the step of geometrically aligning and feature aligning the super-resolution features of the text region and the super-resolution features of the background region includes: Predict the affine transformation matrix of the target text region and the target background region; The coordinates of the super-resolution features of the text region are mapped to the global coordinate system of the super-resolution features of the background region according to the affine transformation matrix; Align the edge contours, character strokes, and semantic information of the super-resolution features of the text region and the super-resolution features of the background region.

[0011] In an optional implementation, the step of performing mask compensation on the overlapping areas of text and background in the fused super-resolution features to obtain the super-resolution video includes: Semantic segmentation is performed on the fused super-resolution features to obtain low-level features and high-level features; The low-level features are inserted into dilated convolutions to expand the receptive field, and the high-level features are upsampled and aligned with the low-level features step by step to obtain the master mask. Calculate the optical flow consistency difference between consecutive frames in the fused super-resolution features, and obtain the error mask based on the optical flow consistency difference; The fused super-resolution features are occlusion compensated based on the master mask and the error mask to obtain the super-resolution video.

[0012] In a second aspect, the present invention provides a video super-resolution system, the system comprising: The text-background separation module is used to separate the text and background of the target video to obtain the target text area and the target background area; The text region super-resolution module is used to perform geometric flow and texture flow feature enhancement on the target text region to obtain text region super-resolution features; The background region super-resolution module is used to perform feature enhancement on the target background region using an optical flow-supervised convolutional neural network to obtain background region super-resolution features; The super-resolution feature fusion module is used to perform feature fusion based on the super-resolution features of the text region and the super-resolution features of the background region to obtain a super-resolution video.

[0013] Thirdly, the present invention provides an electronic device including 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 video super-resolution method as described in the first aspect of the present invention.

[0014] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the video super-resolution method as described in the first aspect of the present invention.

[0015] This invention provides a video super-resolution method, system, electronic device, and storage medium. By designing a neural network specifically for recognizing and separating text regions, it can accurately locate the text position and boundaries in video frames. For text regions, super-resolution processing fully considers the structural features and semantic information of the text, effectively improving text clarity and readability, allowing viewers to see the text content in the video more clearly and avoiding information retrieval issues caused by blurry text. For background regions, super-resolution processing is also performed, focusing on improving text quality while not neglecting background image processing, ensuring that background image quality is not affected and maintaining the overall harmony and aesthetics of the video. The processed text and background regions are effectively merged to generate the final high-definition video frame, achieving differentiated processing of text and background in the video frame. The final result is a globally clear super-resolution video with rich text details, improving overall video quality, providing users with a better visual experience, meeting users' demand for high-quality video, and playing a positive role in various application scenarios such as video playback and information display. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a first schematic diagram of the video super-resolution method provided in this embodiment of the invention; Figure 2 This is a second schematic diagram of the video super-resolution method provided in this embodiment of the invention; Figure 3 This is a third schematic diagram of the video super-resolution method flow provided in this embodiment of the invention; Figure 4 This is the fourth schematic diagram of the video super-resolution method provided in this embodiment of the invention; Figure 5This is the fifth schematic diagram of the video super-resolution method flow provided in the embodiments of the present invention; Figure 6 This is the sixth schematic diagram of the video super-resolution method provided in this embodiment of the invention; Figure 7 This is the seventh schematic diagram of the video super-resolution method provided in the embodiments of the present invention; Figure 8 This is a schematic diagram of the super-resolution video system structure provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0018] Description of main component symbols: 200. Video super-resolution system; 210. Text and background separation module; 220. Text region super-resolution module; 230. Background region super-resolution module; 240. Super-resolution feature fusion module; 300. Electronic device; 310. Processor; 320. Communication interface; 330. Memory; 340. Communication bus. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be further described clearly and completely below with reference to the accompanying drawings of the embodiments of this invention. It should be noted that the described embodiments are merely some embodiments of this invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Existing video super-resolution methods face numerous challenges when processing images containing text. Text, as a crucial element conveying key information in video, is paramount for its clarity and readability. However, text typically possesses complex structures and intricate strokes, and may be intertwined with and occluded by the background in video. Traditional video super-resolution methods mostly employ global processing strategies, which struggle to accurately handle text regions, leading to problems such as blurring, jagged edges, and stroke loss, severely impacting the transmission of textual information and the overall video quality. With the rise of deep learning, models are generally trained using large amounts of data to learn the mapping relationship between low-resolution and high-resolution images. For example, convolutional neural networks automatically extract image features, and the Transformer architecture, with its powerful parallel computing capabilities and long sequence modeling abilities, has also been applied in video super-resolution. However, most existing methods do not fully consider the semantic information and contextual relationships of text, failing to effectively restore text details and semantic consistency. Therefore, in text-dense video scenarios, addressing the insufficient clarity and readability issues caused by existing super-resolution methods has become a critical technical challenge that urgently needs to be overcome in the current video processing field.

[0023] Example 1 This invention provides a video super-resolution method that effectively solves the problems of insufficient clarity and readability caused by existing super-resolution methods. Figure 1 This is a first schematic diagram of the video super-resolution method provided in this embodiment of the invention, as shown below. Figure 1 As shown, the method includes the following steps: S100. Separate the text and background of the target video to obtain the target text area and the target background area.

[0024] In this embodiment of the invention, for the target video, a convolutional neural network integrating dilated convolutional layers, pyramid feature fusion strategy and channel attention mechanism can be used to separate text and background regions, thereby improving the detection accuracy of text regions in the video. Figure 2 This is a second schematic diagram of the video super-resolution method provided in this embodiment of the invention, as shown below. Figure 2 As shown, separating the text and background areas specifically includes the following steps: S110. Extract text structures of different sizes from each video frame of the target video to obtain initial text features.

[0025] In this embodiment of the invention, dilated convolution is used to extract low-resolution text structures of different sizes from each video frame of the target video through multi-scale receptive fields and sparse sampling, thereby obtaining initial text features. The multi-scale receptive fields employ convolutional kernels with different dilation rates to extract text structural features in parallel. When using convolutional kernels with small dilation rates, it can focus on the fine strokes of the text, accurately capturing delicate structures such as left-falling and right-falling strokes, which is crucial for preserving text details. Using convolutional kernels with large dilation rates can cover the contextual information of large fonts. In text recognition, understanding the overall environment of the text helps to more accurately determine the text content. The design of multi-scale receptive fields allows dilated convolution to take into account both local details and overall structure of the text, comprehensively extracting text features and providing rich and accurate information for subsequent processing.

[0026] Sparse sampling effectively expands the field of view by inserting holes (empty spaces) into the convolution kernel, without increasing the number of parameters. This allows dilated convolution to detect distant text edges. For low-resolution text, edge information is often not clear enough, but sparse sampling allows dilated convolution to capture edge information across a certain distance, enhancing the ability to detect text edges and thus better extracting text structure, improving the effect of low-resolution text structure extraction.

[0027] S120. Perform step-by-step upsampling and channel weighting operations on the initial text features to obtain enhanced text features.

[0028] In this embodiment of the invention, a pyramid feature fusion network is employed to enhance the text details of the initial text features through multi-level feature alignment and channel weighting, thereby obtaining enhanced text features. First, deep and shallow features of the initial text features are extracted separately. Deep features contain global semantic information, including the position of text blocks, while shallow features retain high-frequency details such as stroke edges. Then, deconvolution is used to upsample the deep features to a shallow resolution, aligning them channel-by-channel with the shallow features. Finally, 1x1 convolution is used to generate weight coefficients for each channel, reducing the weights of smooth regions to suppress background noise channels and increasing the weights of regions with large gradient changes to enhance text edges.

[0029] S130. Adjust the channel weights to enhance text features and obtain the target text region and target background region.

[0030] In this embodiment of the invention, a channel attention mechanism is employed to dynamically adjust the weights of each channel that enhances text features, strengthening text-related feature channels and suppressing irrelevant feature channels such as background features. This allows for more precise focusing on the target text region while also reasonably distinguishing the target background region. First, a global average pooling operation is performed on the feature map of each channel, compressing the feature map in spatial dimensions (height and width) and converting each channel's feature map into a single numerical value, i.e., a channel descriptor. For example, a feature map of size...H × W × C The feature map, where H Indicates altitude, W Indicates width, C The number of channels is represented by a 1×1×C vector after global average pooling, where each element represents the feature description of the corresponding channel.

[0031] The generated channel descriptors are then input into a fully connected layer, which performs a linear transformation on the channel descriptors to capture the complex relationships between channels. The output of the fully connected layer is then normalized using the Softmax function to obtain the weight coefficients for each channel. The weight coefficients range from 0 to 1, and the sum of the weight coefficients for all channels is 1. These weight coefficients reflect the importance of each channel to the text detection task. For example, the weight coefficient for channels with obvious text edge features tends to be close to 1, while the weight coefficient for channels dominated by background features tends to be close to 0.

[0032] Finally, before feature map reuse, the obtained weight coefficients are multiplied with the original channel features. For feature channels related to text, such as those containing high-frequency components of horizontal and vertical strokes, the multiplication enhances their features due to their larger weight coefficients, making it easier to capture key information about the text. For example, in detecting Chinese characters, horizontal and vertical strokes are the basic elements constituting the characters; by enhancing the channel features corresponding to these strokes, the shape and structure of the characters can be identified more accurately. For channels related to background or irrelevant noise, the multiplication suppresses their features due to their smaller weight coefficients. For example, for background channels such as uniform texture regions, their weight coefficients are close to 0; after multiplication, the features of these channels are almost ignored, thus reducing noise interference.

[0033] In this embodiment of the invention, the dilated convolutional layer employs multi-scale convolutional kernels with various dilation rates to enhance the receptive field of the convolutional neural network (CNN) to capture text structures of different sizes. The pyramid feature fusion strategy integrates multi-level feature maps from the CNN, enhancing the response capability to small text through progressive upsampling and channel weighting operations. The CNN introduces a channel attention mechanism, which adaptively adjusts channel weights to strengthen text edge contour features, thereby enhancing text-related features. Simultaneously, because irrelevant noise channels are suppressed, the difference between the background and text regions becomes more pronounced, thus obtaining the target text region and the target background region.

[0034] S200. Perform geometric flow and texture flow feature enhancement on the target text region to obtain super-resolution features of the text region.

[0035] In this embodiment of the invention, super-resolution processing is performed on the target text region. A dual-stream feature enhancement architecture is adopted to process the structural and detailed information of the text in parallel using the geometric and texture streams. Then, based on an adaptive feature fusion strategy, the fusion weights of geometric and texture features are dynamically adjusted through a gating mechanism. Considering that the semantics may change after super-resolution of blurred text, a pre-trained language model is finally introduced to constrain the semantic rationality of the generated text, thereby obtaining the super-resolution features of the text region. Figure 3 This is a third schematic diagram of the video super-resolution method flow provided in this embodiment of the invention, as shown below. Figure 3 As shown, the super-resolution processing of text specifically includes the following steps: S210. Adjust the shape of the local receptive field of the text features in the target text region to obtain the geometric features of the text.

[0036] In this embodiment of the invention, a deformable convolutional network is used to adjust the shape of the local receptive field of text features in the target text region, accurately capturing the geometric features of strokes such as horizontal, vertical, left-falling, and right-falling strokes to obtain the text's geometric features. The deformable convolutional network introduces learnable offsets, which can be adjusted for the sampling points of each convolutional kernel. Specifically, before the convolution operation, an additional network branch can learn the offset of each sampling point. Then, the sampling position of the convolutional kernel is adjusted according to these offsets, thereby changing the shape of the local receptive field.

[0037] For example, for horizontal strokes, a deformable convolutional network can adjust the offset to expand the receptive field horizontally and shrink it vertically, thus more closely matching the shape of the horizontal stroke and more accurately extracting features such as its edges and length. Similarly, for vertical strokes, a deformable convolutional network can expand the receptive field vertically and shrink it horizontally to accurately capture the features of vertical strokes. For left-falling strokes, a deformable convolutional network can tilt the receptive field to the lower left to better capture features such as the starting point and curvature of left-falling strokes. For right-falling strokes, a deformable convolutional network can tilt the receptive field to the lower right to accurately extract features such as the starting point and curvature of right-falling strokes.

[0038] By adjusting the shape of the local receptive field, deformable convolutional networks can more accurately capture the geometric features of character strokes, which helps improve the accuracy of subsequent character recognition and detection tasks. The learnable offsets of deformable convolutional networks enable them to automatically adapt to different fonts and styles of characters, enhancing the versatility and adaptability of character recognition.

[0039] S220. Perform adversarial training on the text features in the target text region to constrain the realism of high-frequency detail generation and obtain text texture features.

[0040] In this embodiment of the invention, an adversarial generative network (GAN) is used to perform adversarial training on the text features in the target text region to constrain the realism of high-frequency detail generation, thereby obtaining text texture features. The generator of the GAN is responsible for generating text texture features in the target text region, while the discriminator judges the text texture features generated by the generator, distinguishing whether they are real data from a real text image or generated data. Through this adversarial training, the generator continuously adjusts its parameters to generate more realistic text texture features, making it difficult for the discriminator to distinguish them. The presence of the discriminator prompts the generator to generate text texture features with higher realism. It can capture subtle differences and patterns in real text textures, such as the connection methods and curvature of strokes in different fonts.

[0041] For example, when generating handwritten text textures, the generator may initially produce some blurry and unnatural stroke details, which the discriminator can easily identify. However, during adversarial training, the generator gradually learns high-frequency detail features such as the thickness variations of real handwritten strokes and ink distribution, generating textures that are closer to real handwritten text. When generating Song typeface text textures, the generator learns the characteristics of Song typeface strokes—thin horizontal strokes, thick vertical strokes, and sharp angles—making the generated textures more realistic.

[0042] In this embodiment of the invention, the loss function of the adversarial generative network includes a weighted sum of adversarial loss, content loss, and perceptual loss. Weighting these three losses allows for a comprehensive consideration of different requirements. The adversarial loss is the core loss function of the adversarial generative network, used to measure the difference between the text texture features generated by the generator and the real data. It encourages the generator to generate data that can deceive the discriminator, while also prompting the discriminator to continuously improve its judgment ability. For example, as the text texture features generated by the generator become increasingly closer to the real data, the adversarial loss gradually decreases, indicating that the generator's performance is improving. The content loss typically measures the similarity between the text texture features generated by the generator and the target content based on pixel-level differences. It ensures that the texture generated by the generator is consistent with the target in content, without missing or incorrect content. For example, when generating a texture for a specific character, the content loss ensures that the shape and structure of the characters in the generated texture are consistent with the target character. The perceptual loss extracts high-level features from the text texture features generated by the generator and the real data, and then calculates the differences between these high-level features. It focuses on the semantic information and overall perceptual quality of the texture, capturing features that the human visual system is more interested in. For example, perceptual loss can ensure that the generated text texture is similar to the real data in terms of color, texture quality, etc., making the generated texture look more natural and beautiful.

[0043] By adjusting the weights of each loss, the generator's performance in terms of realism, content accuracy, and perceptual quality can be balanced. For example, if a greater emphasis is placed on the realism of the generated text textures, the weight of the adversarial loss can be appropriately increased; if greater emphasis is placed on the consistency between the generated content and the target, the weight of the content loss can be increased. This weighted summation method can be flexibly adjusted according to specific task requirements to generate text texture features that better meet the requirements.

[0044] S230. Adjust the fusion weights of text geometric features and text texture features to obtain text fusion features.

[0045] In this embodiment of the invention, an adaptive feature fusion strategy is adopted. Based on the characteristics of the character stroke structure, the fusion weights of the character geometric features and the character texture features are dynamically adjusted to achieve character feature fusion. The adaptive feature fusion strategy is implemented through a gating mechanism, which uses learnable dynamic weight parameters to adaptively control the feature fusion process, thereby generating the optimal feature representation under different conditions.

[0046] Specifically, a 1x1 convolutional layer is introduced to process the input text geometric and text texture features, generating a gating signal. The value range of the gating signal is restricted to [0,1], representing the relative importance of different feature flows in the fusion process. For example, when the gating signal is close to 1, it means that the text geometric features dominate the fusion, while when the gating signal is close to 0, the text texture features are more important. Based on the generated gating signal, the text geometric and text texture features are linearly weighted and fused. The linear weighting method is simple and effective, and can flexibly adjust the contribution ratio of the two features according to the value of the gating signal.

[0047] Optionally, the weight parameters can be dynamically adjusted based on the stroke results of the characters. When the character stroke structure is simple, the gate signal approaches 1, emphasizing the structural fidelity of the character's geometric features. When the character stroke is complex, the gate signal approaches 0.5, balancing geometric and texture features.

[0048] For example, the stroke structure of printed text is usually relatively regular and simple, with a clear geometric shape and structure. In this case, the structural information provided by the geometric features of the text is crucial for accurately reproducing the shape and structure of the text. When the gating signal approaches 1, the geometric features of the text have a larger weight in the fusion process, while the text texture features have a smaller weight. The fused features retain more of the structural information of the geometric features, better maintaining the shape and structure of the text and avoiding text deformation or distortion caused by interference from texture features. The stroke structure of handwritten text has greater variability and diversity. Different people have different writing styles, and strokes may be curved or connected. In this case, relying solely on the structural information of the geometric features of the text may not be able to fully and accurately express the characteristics of handwritten text. The stroke details and style information contained in the text texture features become particularly important. When the gating signal approaches 0.5, the weights of the geometric features and the text texture features are roughly equal, and they play a balancing role in the fusion process. Thus, the fused features retain both the structural information of the geometric features of the text and incorporate the details and style information of the text texture features, better reproducing the true form and style of handwritten text.

[0049] In this embodiment of the invention, the adaptive feature fusion strategy dynamically adjusts the fusion weights of text geometric features and text texture features through a gating mechanism, and flexibly fuses them according to the characteristics of the text stroke structure, effectively improving the quality of text super-resolution reconstruction and enhancing adaptability and generalization ability.

[0050] S240. The semantic rationality of the text fusion features is verified and processed to obtain the super-resolution features of the text region.

[0051] In this embodiment of the invention, a pre-trained language model is used to verify the semantic rationality of the text fusion features. Through a dual mechanism of semantic penalty and post-processing verification, it is ensured that the generated text conforms to linguistic norms while achieving visual enhancement, thus obtaining super-resolution features of the text region. The pre-trained language model is integrated into the loss function to calculate the semantic similarity between the text sequence of the text fusion features and the original low-resolution text. This pre-trained language model has powerful language understanding and semantic analysis capabilities, enabling accurate evaluation of the semantics of the text. For example, if the original low-resolution text is "HD," and the generated text is "Shangqing," the pre-trained language model can identify the semantic difference between the two. If the generated text contains typos or semantic contradictions, the loss value will increase. During backpropagation, the pre-trained language model will adjust the generator's parameters based on the increased loss value, making the generator more focused on semantic accuracy in subsequent generation processes. For example, when generating "Shangqing" leads to an increased loss value, the generator will adjust its parameters to avoid generating similar erroneous text again.

[0052] Post-processing verification detects grammatical errors in the generated text by comparing it with the trained language model. This analyzes the grammatical structure of the text and determines whether grammatical errors exist, such as subject-verb disagreement or inappropriate collocation. If verification fails, i.e., a grammatical error is detected, a rollback mechanism is triggered. The rollback mechanism re-processes the target text region using super-resolution to regenerate the text, aiming to obtain a result that is both semantically and grammatically correct, thus acquiring the super-resolution features of the text region.

[0053] In this embodiment of the invention, the pre-trained language model effectively solves the problem of semantic distortion of text through a dual mechanism of semantic penalty term and post-processing verification, thereby improving the quality and reliability of text generation.

[0054] S300: An optically flow-supervised convolutional neural network is used to enhance the features of the target background region, thereby obtaining super-resolution features of the background region.

[0055] In this embodiment of the invention, efficient super-resolution enhancement of the target background region is achieved through optical flow iterative optimization and perception quality enhancement techniques. Figure 4 This is the fourth schematic diagram of the video super-resolution method flow provided in the embodiments of the present invention, as shown below. Figure 4 As shown, the super-resolution enhancement of the target background region specifically includes the following steps: S310. Calculate the initial optical flow field of adjacent video frames in the target background region, and optimize the initial optical flow field to obtain the target optical flow field.

[0056] In this embodiment of the invention, for a series of consecutive low-resolution video frames in a target background region, a pre-trained optical flow network model such as RAFT (Rapid Arithmetic Flow Optimization) can be used to calculate the initial optical flow field between adjacent video frames. Simultaneously, local feature matching constraints are introduced to extract local features between adjacent video frames, optimizing the optical flow field by minimizing the differences in local features between adjacent frames. Specifically, for each pixel, its corresponding position in adjacent frames is found through feature matching, and the optical flow vector is adjusted according to the feature differences. Local feature matching constraints can correct local errors in the initial optical flow field, improving the accuracy of the optical flow field, especially in regions with complex textures or rapid motion.

[0057] Optionally, spatiotemporal consistency regularization and Laplacian matrix regularization terms can be introduced to ensure the long-term consistency and spatial smoothness of the optical flow field in the video sequence, avoiding discontinuous or abrupt optical flow vectors. The spatiotemporal consistency regularization term limits the variation amplitude of the optical flow field between adjacent frames, while the Laplacian matrix regularization term penalizes abrupt changes and discontinuities in the optical flow field by calculating its second derivative. In addition to relying on adjacent frames, the optical flow fields from multiple consecutive frames can be jointly optimized to enhance the global consistency of motion estimation.

[0058] In this embodiment of the invention, by introducing various constraints and regularization terms to optimize the initial optical flow field, higher quality reconstruction results can be achieved in the super-resolution task of the target background region.

[0059] S320. Extract features from each video frame in the target background region to obtain the video frame super-resolution result.

[0060] In this embodiment of the invention, a convolutional neural network can be used to extract deep and shallow features from low-resolution background regions. Deep features focus on high semantic information, such as identifying whether an object in the image is a person, landscape, or other object. Shallow features focus on high-resolution detail features, such as the edge contours and texture details of objects. The deep and shallow features are then concatenated along the channel dimension to generate fused features, thereby integrating information from different levels. This results in fused features that contain rich semantic information while retaining sufficient detail information, leading to video frame super-resolution results.

[0061] In the training process of a convolutional neural network, structural similarity loss is used as one of the optimization objectives. During backpropagation, the generator parameters are adjusted to make the structure of the generated frame more similar to that of the original frame. Structural similarity loss is a metric that measures the structural similarity between two images, taking into account the brightness, contrast, and structural information of the images. By calculating the structural similarity loss between the generated frame and the original frame, the generated frame can be constrained to maintain structural similarity with the original frame.

[0062] Adding a bandpass filter to the generator of a convolutional neural network allows signals within a specific frequency range to pass through while suppressing signals in other frequency ranges, thus preserving only specific frequency details in video frames and removing high-frequency noise. Introducing a high-frequency consistency constraint into the discriminator of the convolutional neural network forces the generator to avoid generating unrealistic high-frequency noise.

[0063] S330. Based on the target optical flow field, perform spatiotemporal consistency optimization on the video frame super-resolution results to obtain background region super-resolution features.

[0064] In this embodiment of the invention, the target optical flow field obtained by calculating adjacent frames is inversely deformed into the coordinate system of the current frame to obtain the deformation results of adjacent frames. The feature difference between the super-resolution result of the current frame and the deformation results of adjacent frames is calculated, and this feature difference is added to the loss function as a regularization term. By minimizing this feature difference, the spatiotemporal consistency between frames is maintained during the super-resolution reconstruction process, thereby effectively avoiding discontinuities in the time dimension of the super-resolution results, such as inter-frame flickering or jumps, thus generating coherent and stable background region super-resolution features.

[0065] S400: Feature fusion is performed based on the super-resolution features of the text region and the super-resolution features of the background region to obtain super-resolution video.

[0066] In this embodiment of the invention, through strategies such as multimodal feature alignment, dynamic weight allocation, and overlapping region processing, seamless fusion of super-resolution features of text region and super-resolution features of background region is achieved. Figure 5 This is the fifth schematic diagram of the video super-resolution method flow provided in the embodiments of the present invention, as shown below. Figure 5 As shown, feature fusion specifically includes the following steps: S410. Perform geometric alignment and feature alignment on the super-resolution features of the text region and the super-resolution features of the background region.

[0067] Figure 6 This is the sixth schematic diagram of the video super-resolution method flow provided in the embodiment of the present invention, as shown below. Figure 6 As shown, geometric alignment and feature alignment specifically include the following steps: S411, Predict the affine transformation matrix of the target text region and the target background region.

[0068] In this embodiment of the invention, a spatial transformation network is used to predict the affine transformation matrix of the target text region and the target background region. This affine transformation matrix describes the linear transformation from the target text region to the target background region, including operations such as translation, rotation, and scaling. By minimizing semantic feature differences, the spatial transformation network can learn the optimal transformation parameters.

[0069] S412. Map the coordinates of the super-resolution features of the text region to the global coordinate system of the super-resolution features of the background region according to the affine transformation matrix.

[0070] A sampling grid covering the entire video frame is generated based on the affine transformation matrix. This sampling grid defines the new position of each pixel in the video frame after the transformation. Optionally, bilinear interpolation or nearest neighbor interpolation methods are used to perform pixel-level transformation on the text region super-resolution features based on the generated sampling grid. These interpolation methods can smoothly align the text region super-resolution features to the global coordinate system of the background region super-resolution features. The transformed text region super-resolution features and background region super-resolution features are spatially aligned, achieving pixel-level geometric alignment.

[0071] S413. Align the edge contours, character strokes, and semantic information of the super-resolution features of the text region and the super-resolution features of the background region.

[0072] In this embodiment of the invention, a feature pyramid fusion architecture is employed to align the super-resolution features of the text region and the super-resolution features of the background region. The bottom layer of the feature pyramid fusion architecture focuses on the fusion of edge contours; these bottom-layer features typically contain rich detail information, such as edges and textures, suitable for maintaining the clear outline of the text. The middle layer of the feature pyramid fusion architecture strengthens the stroke features of the text; these middle-layer features can capture the structural information of the text, such as the direction and connection of strokes, helping to maintain the readability of the text. The high layer of the feature pyramid fusion architecture constrains the semantic consistency between the fused features and the original background; these high-level features contain rich semantic information, ensuring that the fused background features are semantically consistent with the original background.

[0073] S420. Dynamically allocate the fusion weights of super-resolution features of text region and super-resolution features of background region based on text detection confidence and optical flow field.

[0074] In this embodiment of the invention, a spatiotemporal joint optimization strategy is employed to address the weight allocation of text and background regions during the fusion of super-resolution features from text and background regions. Horizontally, fusion weights are allocated based on text detection confidence, while vertically, the weight transfer between adjacent frames is constrained to ensure inter-frame coherence of the generated results. Weight allocation is based on text detection confidence; when text detection confidence is sufficient, the weight of text features increases, while the weight of background features decreases, prioritizing the preservation of text details. Simultaneously, the weight transfer constraint between adjacent frames is guided by optical flow field. By calculating the difference in optical flow fields between adjacent frames, the amplitude of optical flow changes is constrained, preventing flickering and jittering of text in the video.

[0075] S430. Based on the fusion weight, the super-resolution features of the text region and the super-resolution features of the background region are fused to obtain the fused super-resolution features.

[0076] In the embodiments of the present invention, the super-resolution features of the text region and the super-resolution features of the background region are weighted and summed according to the calculated fusion weights to obtain the fused super-resolution features.

[0077] S440. Mask compensation is performed on the overlapping areas of text and background in the fused super-resolution features to obtain super-resolution video.

[0078] In this embodiment of the invention, a dual-stream mask generation network is designed for the overlapping area of ​​text and background in the fused super-resolution features. The main mask network marks the invisible area based on semantic segmentation, and the auxiliary mask network filters dynamic occlusion errors through optical flow consistency test. Finally, the fused mask is normalized to achieve pixel-level occlusion compensation. Figure 7 This is the seventh schematic diagram of the video super-resolution method provided in this embodiment of the invention, as shown below. Figure 7 As shown, mask compensation specifically includes the following steps: S441. Perform semantic segmentation on the fused super-resolution features to obtain low-level and high-level features.

[0079] In this embodiment of the invention, a pre-trained semantic segmentation model can be used to extract multi-scale features that fuse super-resolution features, thereby obtaining low-level features and high-level features.

[0080] S442. Insert low-level features into dilated convolutions to expand the receptive field, and align high-level features with low-level features step by step through upsampling to obtain the master mask.

[0081] Dilated convolutions are inserted into low-level features to expand the receptive field and capture global structural information of the text. High-level features are then progressively aligned with low-level features through upsampling operations, enhancing the ability to recognize small characters and obtaining semantic segmentation results. Upsampling can employ methods such as deconvolution or interpolation to match the spatial resolution of high-level features with that of low-level features. Occluded or invisible text regions are marked using the semantic segmentation results, generating a master mask.

[0082] S443. Calculate the optical flow consistency difference between consecutive frames in the fused super-resolution features, and obtain the error mask based on the optical flow consistency difference.

[0083] In this embodiment of the invention, an optical flow estimation network is used to calculate the optical flow field between consecutive frames in the fused super-resolution features, and the consistency of optical flow between consecutive frames is analyzed. If the optical flow difference in a certain region is too large, it may indicate that there is occlusion, motion blur, or other anomalies in that region, and therefore it is identified as a potential error region. An appropriate difference threshold is set, and regions with optical flow differences exceeding the difference threshold are marked as error regions, generating an error mask. This error mask is a binary image, where error regions are represented by 1 and non-error regions by 0.

[0084] S444. Perform occlusion compensation on the fused super-resolution features based on the master mask and error mask to obtain the super-resolution video.

[0085] In this embodiment of the invention, the master mask and the error mask are combined, and logical operations are used to suppress missegmentation of text in occluded areas. Specifically, if a region is marked as an error region in the error mask, its corresponding pixel value is set to 0 (i.e., background) in the master mask, thereby avoiding missegmentation. The fused mask is normalized so that its pixel values ​​are within the range of [0,1]. The normalized mask can be used to guide super-resolution reconstruction, achieving pixel-level occlusion compensation and obtaining super-resolution video.

[0086] As a further embodiment of the present invention, a video super-resolution model can be constructed based on the above embodiments. This video super-resolution model requires training with multiple pairs of low-resolution and high-resolution videos. The learning objective is to minimize the difference between the model's super-resolution video and the real high-resolution video, which can be specifically measured from aspects such as high-fidelity reconstruction of text and background regions and global consistency.

[0087] Optionally, low-resolution videos can be collected using both real and degraded data. Low-resolution footage, such as videos recorded with a mobile phone, can be used directly, preserving real noise and compression artifacts. Alternatively, bicubic interpolation can be performed on original high-definition videos, such as 4K videos, to reduce degradation and generate low-resolution videos. Motion-blurred scene data, such as racing and dance videos, can be added to the training data to improve the model's ability to handle complex motions. Text region annotation data can be generated using manual annotation or semantic segmentation models to create text masks. Data augmentation techniques such as random cropping and color dithering can be introduced to improve the generalization of the training data. The training data is then divided into training, validation, and test sets in a 7:2:1 ratio.

[0088] In this embodiment of the invention, the loss function of the video super-resolution model can adopt a multi-task joint optimization loss function, including adversarial loss, content loss, perceptual loss, structural similarity loss, and inter-frame consistency loss, etc., and achieve multi-objective collaborative optimization through dynamic weight allocation. In the initial training, adversarial loss and content loss are emphasized, and in the later stage, perceptual loss, structural similarity loss, and inter-frame consistency loss are added to improve the realism of details.

[0089] Optionally, during the training of the video super-resolution model, the optimizer uses the Adam optimizer, and gradient clipping is employed to limit the gradient norm to no more than 1 to prevent training instability. A cosine annealing strategy is used during training, gradually reducing the learning rate in the later stages, and L2 regularization is added to prevent overfitting during model training.

[0090] The video super-resolution method provided in this invention, through the design of a specialized neural network for recognizing and separating text regions, can accurately locate the text position and boundaries in video frames. For text regions, super-resolution processing fully considers the structural features and semantic information of the text, effectively improving text clarity and readability, allowing viewers to see the text content in the video more clearly and avoiding information retrieval difficulties caused by blurry text. For background regions, super-resolution processing is also performed, focusing on improving text quality while not neglecting background image processing, ensuring that background image quality is not affected and maintaining the overall harmony and aesthetics of the video.

[0091] Example 2 Based on the same technical concept as the above embodiments, this embodiment of the invention provides a video super-resolution system. Figure 8 This is a schematic diagram of the super-resolution video system structure provided in an embodiment of the present invention, as shown below. Figure 8 As shown, the super-resolution system 200 for this video includes: The text-background separation module 210 is used to separate the text and background of the target video to obtain the target text area and the target background area.

[0092] The text region super-resolution module 220 is used to perform geometric flow and texture flow feature enhancement on the target text region to obtain text region super-resolution features; Background region super-resolution module 230 is used to perform feature enhancement on the target background region using an optical flow-supervised convolutional neural network to obtain background region super-resolution features.

[0093] The super-resolution feature fusion module 240 is used to perform feature fusion based on the super-resolution features of the text region and the super-resolution features of the background region to obtain super-resolution video.

[0094] The video super-resolution system provided in this invention can generate super-resolution videos with global clarity and rich text details, thereby improving the overall video quality, providing users with a better visual experience, meeting users' demand for high-quality videos, and playing a positive role in multiple application scenarios such as video playback and information display.

[0095] It is understood that the implementation method of the video super-resolution method described in Embodiment 1 above is also applicable to this embodiment and can achieve the same technical effect, so it will not be described again here.

[0096] Example 3 Based on the same concept, embodiments of the present invention also provide an electronic device. Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 9 As shown, the electronic device 300 may include a processor 310, a communication interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute the steps of the video super-resolution method as described in the above embodiments. For example, this includes: S100: Separate the text and background of the target video to obtain the target text area and the target background area; S200. Perform geometric flow and texture flow feature enhancement on the target text region to obtain super-resolution features of the text region; S300: An optical flow-supervised convolutional neural network is used to enhance the features of the target background region and obtain super-resolution features of the background region. S400: Feature fusion is performed based on the super-resolution features of the text region and the super-resolution features of the background region to obtain super-resolution video.

[0097] The processor 310 can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0098] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] The memory 330 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0100] Example 4 Based on the same concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program containing at least one piece of code executable by a master control device to control the master control device to implement the video super-resolution method as described in the above embodiments. For example, it includes: S100: Separate the text and background of the target video to obtain the target text area and the target background area; S200. Perform geometric flow and texture flow feature enhancement on the target text region to obtain super-resolution features of the text region; S300: An optical flow-supervised convolutional neural network is used to enhance the features of the target background region and obtain super-resolution features of the background region. S400: Feature fusion is performed based on the super-resolution features of the text region and the super-resolution features of the background region to obtain super-resolution video.

[0101] Based on the same technical concept, this embodiment of the invention also provides a computer program, which, when executed by a master control device, is used to implement the above-described method embodiments.

[0102] The computer program may be stored, in whole or in part, on a computer-readable storage medium packaged with the processor, or in part or in whole on a memory not packaged with the processor.

[0103] Based on the same technical concept, embodiments of the present invention also provide a processor for implementing the above-described method embodiments. The processor may be a chip.

[0104] In summary, the video super-resolution method, system, electronic device, and storage medium provided by this invention, through the design of a neural network specifically for recognizing and separating text regions, can accurately locate the text position and boundaries in video frames. For text regions, super-resolution processing fully considers the structural features and semantic information of the text, effectively improving text clarity and readability, allowing viewers to see the text content in the video more clearly and avoiding information retrieval issues caused by blurry text. For background regions, super-resolution processing is also performed, focusing on improving text quality while not neglecting background image processing, ensuring that background image quality is not affected and maintaining the overall harmony and aesthetics of the video. The processed text and background regions are effectively merged to generate the final high-definition video frame, achieving differentiated processing of text and background in the video frame. The final result is a globally clear super-resolution video with rich text details, improving overall video quality, providing users with a better visual experience, meeting users' demand for high-quality video, and playing a positive role in various application scenarios such as video playback and information display.

[0105] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0106] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications 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.

Claims

1. A video super-resolution method, characterized in that, The method includes: Separate the target video from the text and background to obtain the target text region and the target background region; Geometric flow and texture flow feature enhancements are performed on the target text region to obtain super-resolution features of the text region. An optical flow-supervised convolutional neural network is used to enhance the features of the target background region, thereby obtaining super-resolution features of the background region; Feature fusion is performed based on the super-resolution features of the text region and the super-resolution features of the background region to obtain a super-resolution video.

2. The video super-resolution method according to claim 1, characterized in that, The step of separating the target video into text and background to obtain the target text region and the target background region includes: Extract text structures of different sizes from each video frame of the target video to obtain initial text features; The initial text features are subjected to progressive upsampling and channel weighting operations to obtain enhanced text features; Adjust the channel weights of the enhanced text features to obtain the target text region and the target background region.

3. The video super-resolution method according to claim 1, characterized in that, The step of performing geometric flow and texture flow feature enhancement on the target text region to obtain super-resolution features of the text region includes: Adjust the shape of the local receptive field of the text features in the target text region to obtain the text geometric features; Adversarial training is performed on the text features in the target text region to constrain the realism of high-frequency detail generation and obtain text texture features; Adjust the fusion weights of the text geometric features and the text texture features to obtain text fusion features; The semantic rationality of the text fusion features is verified to obtain the super-resolution features of the text region.

4. The video super-resolution method according to claim 1, characterized in that, The optically supervised convolutional neural network enhances the features of the target background region to obtain super-resolution features of the background region, including: Calculate the initial optical flow field of adjacent video frames in the target background region, and optimize the initial optical flow field to obtain the target optical flow field; Feature extraction is performed on each video frame in the target background region to obtain the video frame super-resolution result; Based on the target optical flow field, the super-resolution results of the video frame are optimized for spatiotemporal consistency to obtain the super-resolution features of the background region.

5. The video super-resolution method according to claim 1, characterized in that, The step of fusing features based on the super-resolution features of the text region and the super-resolution features of the background region to obtain a super-resolution video includes: Perform geometric alignment and feature alignment on the super-resolution features of the text region and the super-resolution features of the background region; The fusion weights of the super-resolution features of the text region and the super-resolution features of the background region are dynamically assigned based on the text detection confidence and optical flow field. Based on the fusion weights, the super-resolution features of the text region and the super-resolution features of the background region are fused to obtain fused super-resolution features; The super-resolution video is obtained by masking the overlapping areas of text and background in the fused super-resolution features.

6. The video super-resolution method according to claim 5, characterized in that, The step of performing geometric alignment and feature alignment on the super-resolution features of the text region and the super-resolution features of the background region includes: Predict the affine transformation matrix of the target text region and the target background region; The coordinates of the super-resolution features of the text region are mapped to the global coordinate system of the super-resolution features of the background region according to the affine transformation matrix; Align the edge contours, character strokes, and semantic information of the super-resolution features of the text region and the super-resolution features of the background region.

7. The video super-resolution method according to claim 5, characterized in that, The step of performing mask compensation on the overlapping areas of text and background in the fused super-resolution features to obtain the super-resolution video includes: Semantic segmentation is performed on the fused super-resolution features to obtain low-level features and high-level features; The low-level features are inserted into dilated convolutions to expand the receptive field, and the high-level features are upsampled and aligned with the low-level features step by step to obtain the master mask. Calculate the optical flow consistency difference between consecutive frames in the fused super-resolution features, and obtain the error mask based on the optical flow consistency difference; The fused super-resolution features are occlusion compensated based on the master mask and the error mask to obtain the super-resolution video.

8. A video super-resolution system, characterized in that, The system includes: The text-background separation module is used to separate the text and background of the target video to obtain the target text area and the target background area; The text region super-resolution module is used to perform geometric flow and texture flow feature enhancement on the target text region to obtain text region super-resolution features; The background region super-resolution module is used to perform feature enhancement on the target background region using an optical flow-supervised convolutional neural network to obtain background region super-resolution features; The super-resolution feature fusion module is used to perform feature fusion based on the super-resolution features of the text region and the super-resolution features of the background region to obtain a super-resolution video.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the video super-resolution method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the video super-resolution method as described in any one of claims 1-7.

Images