An image super-resolution reconstruction method, system, computer device and storage medium

Abstract

This application discloses an image super-resolution reconstruction method, system, computer device, and storage medium. First, the image to be super-resolution is acquired and its features are extracted to obtain a first semantic feature vector. Then, a set of high-resolution images is acquired, and downsampled according to the target magnification to obtain a sampled image set. Features are extracted from these sampled images to construct a candidate feature library containing a second semantic feature vector. The similarity between each feature vector in the candidate feature library and the first semantic feature vector is calculated. Based on the similarity score, matching high-resolution images are selected to form a dynamic training set. The dynamic training set is input into a diffusion model for training, resulting in a trained diffusion model. The image to be super-resolution is then input into the trained diffusion model to generate the super-resolution image. This scheme filters homogeneous high-resolution samples through semantic feature matching and constructs a dynamic training set, reducing interference from irrelevant data, effectively improving the semantic consistency and detail texture restoration of the reconstructed image, and efficiently optimizing the quality and processing accuracy of image super-resolution reconstruction.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image super-resolution reconstruction method, system, computer device, and storage medium. Background Technology

[0002] Image super-resolution technology aims to recover high-resolution images from low-resolution image observations. In recent years, with the rise of generative artificial intelligence, diffusion-based super-resolution methods have gradually become mainstream. As an advanced image restoration framework, diffusion models utilize pre-trained generative priors to generate realistic texture details through progressive denoising. However, this powerful generative ability is prone to producing false textures that do not conform to the semantics of the original image when there is a lack of sufficient prior information constraints, i.e., "model illusion." This problem is particularly severe when facing high-magnification tasks such as 8x super-resolution. In such tasks, the input image retains only a very small amount of pixel information from the original image, and high-frequency details are almost completely lost. Current training strategies usually rely on fixed general datasets (such as DIV2K). Due to the difference in distribution between training data and test scenarios, when diffusion models face complex semantic scenes they have never seen before, they often fail to find suitable reference features, leading to distorted generation results.

[0003] To address this issue, it is crucial to select high-resolution samples from massive amounts of external data that are semantically consistent with the test images as "mentor data." However, traditional data selection methods primarily rely on pixel-level metrics. But under extreme conditions of 8x downsampling, the images become extremely blurry, and pixel-to-pixel comparisons can no longer reflect the true structure of the image, leading to significant deviations in the selection results. Summary of the Invention

[0004] To overcome the problem of insufficient training accuracy of diffusion models for image super-resolution in the prior art, this invention provides an image super-resolution reconstruction method, system, computer equipment, and storage medium.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: An image super-resolution reconstruction method includes the following steps: S1: Obtain the image to be super-resolution, and perform feature extraction processing on the image to be super-resolution to obtain the first semantic feature vector; S2: Obtain a high-definition image set, downsample the high-definition image set with a preset super-resolution magnification ratio as the sampling ratio to obtain a sampled image set, perform feature extraction processing on the sampled image set to obtain a second semantic feature vector set, and construct a candidate feature library based on the second semantic feature vector set; S3: Calculate the similarity between each second semantic feature vector and the first semantic feature vector in the candidate feature library to obtain the similarity score between the image to be super-resolution and each of the high-definition images; S4: Select several high-definition images as a dynamic training set based on the similarity scores, and input the dynamic training set into a preset diffusion model for training to obtain a trained diffusion model; S5: Input the image to be super-resolution into the trained diffusion model and output the super-resolution image.

[0006] Preferably, the feature extraction process for the image to be super-resolution includes: The image to be super-resolution is divided into multiple images Patch according to a preset size. The multiple images Patch are input into a preset feature extraction model for feature extraction to obtain the semantic feature vector of each image Patch. The semantic feature vectors of each image Patch are stacked in order to obtain the first semantic feature vector corresponding to the image to be super-resolution. Feature extraction processing of the sampled image set includes: For each sampled image in the sampled image set, the sampled image is divided into multiple sampled image patches according to a preset size. The multiple sampled image patches are input into a preset feature extraction model for feature extraction to obtain the semantic feature vector of each sampled image patch. The semantic feature vectors of each sampled image patch are stacked in order to obtain the second semantic feature vector corresponding to the sampled image. The second semantic feature vector corresponding to each sampled image in the sampled image set is integrated and processed to obtain the second semantic feature vector set.

[0007] Preferably, calculating the similarity between each second semantic feature vector and the first semantic feature vector in the candidate feature library to obtain the similarity score between the image to be super-resolution and each of the high-resolution images includes: For each second semantic feature vector in the candidate feature library, the semantic feature vector of each of its patches is compared with the semantic feature vector of each patch in the first semantic feature vector to calculate the pairwise similarity, so as to obtain the original similarity between each pair of patches. The original similarity scores are accumulated and then divided by the number of patches in the corresponding high-definition image to obtain the similarity score between the image to be super-resolution and the corresponding high-definition image.

[0008] Preferably, the calculation of the original similarity is expressed as follows:

[0009] in, Let be the semantic feature vector of the m-th patch of the image to be super-resolution. This is the semantic feature vector of the nth patch corresponding to the high-resolution image. For vector dot product operation, This is a vector transpose operation. and They are respectively and The modulus length; The similarity score between the i-th high-resolution image and the image to be super-resolution is calculated as follows:

[0010] in, Let be the number of patches for the i-th high-resolution image. Let be the semantic feature vector of the nth patch of the i-th high-resolution image.

[0011] Preferably, the preset feature extraction model is a shared-weight DINOv2 model.

[0012] Preferably, selecting a number of the high-resolution images as a dynamic training set based on the similarity score includes: Sort all similarity scores in descending order, and select the high-resolution images corresponding to the top few similarity scores to construct a dynamic training set.

[0013] Preferably, selecting several first feature vectors with the highest similarity scores and obtaining their corresponding sampled images, and obtaining high-resolution images of the corresponding sampled images as a dynamic training set further includes: If the acquired high-resolution image has already been selected as a dynamic training set by other images to be super-resolution, then skip the high-resolution image and continue filtering according to the similarity score.

[0014] An image super-resolution reconstruction system, applying any one of the above, includes: Feature extraction module: used to acquire the image to be super-resolution, perform feature extraction processing on the image to be super-resolution, and obtain the first semantic feature vector; Candidate feature library construction module: used to acquire a set of high-definition images, downsample the set of high-definition images at a preset super-resolution magnification ratio to obtain a set of sampled images, perform feature extraction processing on the set of sampled images to obtain a second set of semantic feature vectors, and construct a candidate feature library based on the second set of semantic feature vectors; Similarity calculation module: used to calculate the similarity between each second semantic feature vector and the first semantic feature vector in the candidate feature library, and to obtain the similarity score between the image to be super-resolution and each of the high-definition images; Model training module: used to select several high-definition images as a dynamic training set based on the similarity score, input the dynamic training set into a preset diffusion model for training, and obtain a trained diffusion model; Super-resolution module: Used to input the image to be super-resolution into the trained diffusion model and output a super-resolution image.

[0015] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the methods described above.

[0017] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This method extracts semantic features from the image to be super-resolution, then constructs a candidate feature library composed of semantic features from a set of high-resolution images. Based on the semantic features of the super-resolution image and the high-resolution image set, it performs similarity matching, quickly selecting high-resolution images similar to the image to be super-resolution to form a dynamic training set. This makes the training of the diffusion model more targeted and improves the model's adaptability to the current image to be processed. Using the dynamic training set to specifically train the preset diffusion model before performing super-resolution processing ensures that the output super-resolution image more closely matches the real content of the image to be super-resolution in terms of semantic structure and detail texture. This effectively improves the generation quality and semantic consistency of the super-resolution image, while reducing the interference of irrelevant data on model training, thus improving the accuracy and reliability of super-resolution processing. Attached Figure Description

[0018] Figure 1 This is a flowchart of an image super-resolution reconstruction method in Example 1; Figure 2 This is a flowchart of an image super-resolution reconstruction method in Example 1; Figure 3 This is a flowchart illustrating the process of constructing a dynamic training set for an image super-resolution reconstruction method in Example 1. Figure 4 This is one of the 8x super-resolution effect images of an image super-resolution reconstruction method in Example 1; Figure 5This is the second image showing the 8x super-resolution effect of an image super-resolution reconstruction method in Example 1. Figure 6 This is a system framework diagram of an image super-resolution reconstruction system in Example 2. Detailed Implementation

[0019] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this application. To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions; It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Example 1 Flowchart as follows Figure 1 As shown, an image super-resolution reconstruction method includes the following steps: S1: Obtain the image to be super-resolution, and perform feature extraction processing on the image to be super-resolution to obtain the first semantic feature vector; S2: Obtain a high-definition image set, downsample the high-definition image set with a preset super-resolution magnification ratio as the sampling ratio to obtain a sampled image set, perform feature extraction processing on the sampled image set to obtain a second semantic feature vector set, and construct a candidate feature library based on the second semantic feature vector set; S3: Calculate the similarity between each second semantic feature vector and the first semantic feature vector in the candidate feature library to obtain the similarity score between the image to be super-resolution and each of the high-definition images; S4: Select several high-definition images as a dynamic training set based on the similarity scores, and input the dynamic training set into a preset diffusion model for training to obtain a trained diffusion model; S5: Input the image to be super-resolution into the trained diffusion model and output the super-resolution image.

[0022] The feature extraction process for the image to be super-resolution includes: The image to be super-resolution is divided into multiple images Patch according to a preset size. The multiple images Patch are input into a preset feature extraction model for feature extraction to obtain the semantic feature vector of each image Patch. The semantic feature vectors of each image Patch are stacked in order to obtain the first semantic feature vector corresponding to the image to be super-resolution. Feature extraction processing of the sampled image set includes: For each sampled image in the sampled image set, the sampled image is divided into multiple sampled image patches according to a preset size. The multiple sampled image patches are input into a preset feature extraction model for feature extraction to obtain the semantic feature vector of each sampled image patch. The semantic feature vectors of each sampled image patch are stacked in order to obtain the second semantic feature vector corresponding to the sampled image. The second semantic feature vector corresponding to each sampled image in the sampled image set is integrated and processed to obtain the second semantic feature vector set.

[0023] Calculating the similarity between each second semantic feature vector and the first semantic feature vector in the candidate feature library to obtain the similarity score between the image to be super-resolution and each of the high-resolution images includes: For each second semantic feature vector in the candidate feature library, the semantic feature vector of each of its patches is compared with the semantic feature vector of each patch in the first semantic feature vector to calculate the pairwise similarity, so as to obtain the original similarity between each pair of patches. The original similarity scores are accumulated and then divided by the number of patches in the corresponding high-definition image to obtain the similarity score between the image to be super-resolution and the corresponding high-definition image.

[0024] The calculation of the original similarity is expressed as follows:

[0025] in, Let be the semantic feature vector of the m-th patch of the image to be super-resolution. This is the semantic feature vector of the nth patch corresponding to the high-resolution image. For vector dot product operation, This is a vector transpose operation. and They are respectively and The modulus length; The similarity score between the i-th high-resolution image and the image to be super-resolution is calculated as follows:

[0026] in, Let be the number of patches for the i-th high-resolution image. Let be the semantic feature vector of the nth patch of the i-th high-resolution image.

[0027] The preset feature extraction model is the DINOv2 model with shared weights.

[0028] Selecting a number of the high-resolution images as a dynamic training set based on the similarity scores includes: Sort all similarity scores in descending order, and select the high-resolution images corresponding to the top few similarity scores to construct a dynamic training set.

[0029] Selecting several first feature vectors with the highest similarity scores and obtaining their corresponding sampled images, and obtaining high-resolution images of the corresponding sampled images as a dynamic training set, further includes: If the acquired high-resolution image has already been selected as a dynamic training set by other images to be super-resolution, then skip the high-resolution image and continue filtering according to the similarity score.

[0030] In the specific implementation process, the process framework diagram is as follows: Figure 2 As shown, firstly, a low-resolution image to be super-resolution is acquired and divided into multiple patches according to a preset size. In this embodiment, the preset size is set to 140×140 with a step size of 70 pixels to ensure that there is no overlap between the patches before the image to be super-resolution is divided. Then, each patch of the image to be super-resolution is input into the encoder of the DINOv2 model with shared weights for feature extraction to obtain the semantic feature vector of each patch of the image to be super-resolution. The semantic feature vectors of each patch are stacked in order to obtain the first semantic feature vector.

[0031] In this embodiment, the Liu4k_v2, UHDIQA, and Aes datasets are used as high-resolution image sets. Images in these datasets possess realistic high-resolution textures consistent with the scene to be super-resolutiond. Even in high-magnification super-resolution tasks where high-frequency information is extremely scarce, this method can provide accurate texture feature guidance, effectively constraining the generation space of the diffusion model and significantly improving reconstruction fidelity in high-magnification scenes. Next, the high-resolution image is downsampled at a preset super-resolution magnification of 8x to obtain a degraded pseudo-low-resolution sampled image. An edge reflection filling strategy is used to fill in edge pixels, thereby unifying the dimensions of the high-resolution image and the image to be super-resolutiond. The image to be super-resolution is then segmented and feature extracted in the same manner to obtain a second semantic feature vector set. A candidate feature library is constructed based on this second semantic feature vector set. At this point, the first and second semantic feature vectors have the same dimension and the same scale between patches, meeting the conditions for similarity calculation. Traditional screening strategies typically use pixel-level metrics such as MSE or PSNR to measure image differences. However, under extreme blurring conditions caused by high-magnification downsampling, pixel information is severely lost, often rendering pixel-to-pixel comparisons ineffective. This can easily lead to significant semantic discrepancies between the selected data and the test image. The present invention uses a first semantic feature vector and a second semantic feature vector with the same dimension and the same scale across patches. The similarity calculation overcomes the limitation of blurred pixels in low-resolution images, accurately identifying deep texture semantics and structural features. This ensures that the selected high-resolution images and the images to be super-resolutiond maintain a high degree of consistency in distribution, thereby significantly improving the targeting and effectiveness of model training.

[0032] For each second semantic feature vector in the candidate feature library, the semantic feature vector of each of its patches is compared with the semantic feature vector of each patch in the first semantic feature vector to calculate the pairwise similarity, so as to obtain the original similarity between each pair of patches. The original similarity scores are accumulated and then divided by the number of patches in the corresponding high-definition image to obtain the similarity score between the image to be super-resolution and the corresponding high-definition image.

[0033] The calculation of the original similarity is expressed as follows:

[0034] in, Let be the semantic feature vector of the m-th patch of the image to be super-resolution. This is the semantic feature vector of the nth patch corresponding to the high-resolution image. For vector dot product operation, This is a vector transpose operation. and They are respectively and The modulus length; The similarity score between the i-th high-resolution image and the image to be super-resolution is calculated as follows:

[0035] in, Let be the number of patches for the i-th high-resolution image. Let be the semantic feature vector of the nth patch of the i-th high-resolution image.

[0036] Based on the similarity scores, all similarity scores are sorted in descending order using a greedy deduplication selection strategy. That is, if the high-resolution image corresponding to a certain similarity score has been selected as a dynamic training set by other images to be super-resolution, then the high-resolution image is skipped and the selection continues to be carried out according to the similarity scores.

[0037] The flowchart for constructing a dynamic training set is as follows: Figure 3 As shown, a dynamic training set is constructed by selecting high-resolution images corresponding to the top few similarity scores. In this embodiment, the top 3 high-resolution images corresponding to the top similarity scores are selected to construct the dynamic training set. Then, the dynamic training set is input into a preset diffusion model for training to obtain a trained diffusion model. In this embodiment, the preset diffusion model is the HYPIR diffusion model. For each image to be super-resolution processed, its dynamic dataset is constructed, and its corresponding dynamic dataset is input into the preset diffusion model for training. The trained diffusion model is then used to perform super-resolution processing on the image to be super-resolution, resulting in a super-resolution image. In this embodiment, two images to be super-resolution processed are used for testing, and the effect is shown in the figure. Figure 4 , Figure 5 As shown, the method of this invention significantly improves the visual performance of blind super-resolution tasks. Compared with other traditional generative models that lack real references, the super-resolution images generated by this invention have higher fidelity and more realistic details and textures, effectively suppressing artifacts and "hallucinations" common in generative models. It exhibits strong adaptability and robustness, especially in super-resolution reconstruction tasks of complex real-world images. Through this image super-resolution method based on semantic data source tracing and diffusion models, combined with the construction of a dynamic training set, the texture accuracy and visual fidelity of super-resolution images are significantly improved, providing an efficient and reliable solution for blind super-resolution reconstruction in complex degraded scenarios.

[0038] Example 2 System framework diagram as follows Figure 6 As shown, an image super-resolution reconstruction system includes: Feature extraction module 201: used to acquire the image to be super-resolution, perform feature extraction processing on the image to be super-resolution, and obtain a first semantic feature vector; Candidate feature library construction module 202: used to acquire a set of high-definition images, downsample the set of high-definition images at a preset super-resolution magnification ratio to obtain a set of sampled images, perform feature extraction processing on the set of sampled images to obtain a second set of semantic feature vectors, and construct a candidate feature library based on the second set of semantic feature vectors; Similarity calculation module 203: used to calculate the similarity between each second semantic feature vector and the first semantic feature vector in the candidate feature library, and to obtain the similarity score between the image to be super-resolution and each of the high-definition images; Model training module 204: used to select several high-definition images as a dynamic training set based on the similarity score, input the dynamic training set into a preset diffusion model for training, and obtain a trained diffusion model; Super-resolution module 205: used to input the image to be super-resolution into the trained diffusion model and output the super-resolution image.

[0039] In the specific implementation process, the feature extraction module 201 first receives the image to be super-resolution, divides it into image patches according to a preset size, completes feature extraction through the DINOv2 model with shared weights, and stacks them in order to obtain the first semantic feature vector and outputs it; the candidate feature library construction module 202 obtains a set of high-definition images, downsamples them according to the target magnification to obtain a set of sampled images, extracts semantic features in the same way as the feature extraction module 201, integrates them to form a set of second semantic feature vectors, and completes the construction of the candidate feature library; the similarity calculation module 203 reads the first semantic feature vector and the candidate feature library, calculates the original similarity for each patch according to the cosine similarity formula, accumulates and normalizes to obtain the similarity score between the image to be super-resolution and each high-definition image; the model training module 204 sorts the similarity scores in descending order, uses a greedy deduplication strategy to select high-resolution high-definition images to form a dynamic training set, inputs the training set into a preset diffusion model to complete the training, and outputs a trained diffusion model adapted to the current image; the super-resolution module 205 inputs the image to be super-resolution into the trained diffusion model, performs super-resolution reconstruction, and finally outputs a high-definition super-resolution image with semantic consistency and high detail restoration.

[0040] Example 3 This embodiment provides a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set. The at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor, causing the processor to perform some or all of the steps of the method provided in Embodiment 1 of this application.

[0041] It is understood that the storage medium can be transient or non-transient. Exemplarily, the storage medium includes, but is not limited to, 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.

[0042] By way of example, the processor may be a central processing unit (CPU), a microprocessor unit (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0043] By way of example, the read-only memory includes, but is not limited to, MASK ROM, PROM, EPROM, EEPROM, Flash, etc.

[0044] By way of example, the random access memory includes, but is not limited to, DRAM, SRAM, SDRAM, DDR SDRAM, etc.

[0045] In some examples, a computer program product is provided, which can be implemented by hardware, software, or a combination thereof. As a non-limiting example, the computer program product can be embodied in the storage medium, or it can be embodied in a software product, such as an SDK (Software Development Kit).

[0046] As a non-limiting example, a computer program product is provided, comprising a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer program or computer-executable instructions from the computer-readable storage medium, and executes the computer-executable instructions, causing the electronic device to perform some or all of the steps of the method described in the embodiments of this application.

[0047] In some examples, a computer program is provided, including computer-readable code, wherein, when the computer-readable code is run in a computer device, a processor in the computer device performs some or all of the steps for implementing the method.

[0048] This embodiment also proposes an electronic device, including a memory and a processor. The memory stores at least one instruction, at least one program, code set, or instruction set. When the processor executes the at least one instruction, at least one program, code set, or instruction set, it implements some or all of the steps of the method described in Embodiment 1.

[0049] In some examples, a hardware entity of the electronic device is provided, including: a processor, a memory, and a communication interface; wherein the processor typically controls the overall operation of the electronic device; the communication interface is used to enable the electronic device to communicate with other terminals or servers via a network; the memory is configured to store instructions and applications executable by the processor, and may also cache data to be processed or already processed (including but not limited to image data, audio data, voice communication data, and video communication data) to be processed by the processor and various modules in the electronic device, and may be implemented using flash memory (FLASH), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or random access memory (RAM).

[0050] A processor may include one or more processing elements. Therefore, a processor may include one or more integrated circuits (ICs) configured to perform the functions of the processor. Furthermore, each integrated circuit may include circuitry (e.g., a first circuit, a second circuit, and other circuitry) configured to perform the functions of the processor.

[0051] Furthermore, data can be transferred between the processor, communication interface, and memory via a bus, which can include any number of interconnected buses and bridges, connecting various circuits of one or more processors and memories together.

[0052] It is understood that the options in Embodiment 1 above also apply to this embodiment, so they will not be described again here.

[0053] The same or similar labels correspond to the same or similar parts; The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this application. Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. An image super-resolution reconstruction method, characterized in that, Includes the following steps: S1: Obtain the image to be super-resolution, and perform feature extraction processing on the image to be super-resolution to obtain the first semantic feature vector; S2: Obtain a high-definition image set, downsample the high-definition image set with a preset super-resolution magnification ratio as the sampling ratio to obtain a sampled image set, perform feature extraction processing on the sampled image set to obtain a second semantic feature vector set, and construct a candidate feature library based on the second semantic feature vector set; S3: Calculate the similarity between each second semantic feature vector and the first semantic feature vector in the candidate feature library to obtain the similarity score between the image to be super-resolution and each of the high-definition images; S4: Select several high-definition images as a dynamic training set based on the similarity scores, and input the dynamic training set into a preset diffusion model for training to obtain a trained diffusion model; S5: Input the image to be super-resolution into the trained diffusion model and output the super-resolution image.

2. The image super-resolution reconstruction method according to claim 1, characterized in that, The feature extraction process for the image to be super-resolution includes: The image to be super-resolution is divided into multiple images Patch according to a preset size. The multiple images Patch are input into a preset feature extraction model for feature extraction to obtain the semantic feature vector of each image Patch. The semantic feature vectors of each image Patch are stacked in order to obtain the first semantic feature vector corresponding to the image to be super-resolution. Feature extraction processing of the sampled image set includes: For each sampled image in the sampled image set, the sampled image is divided into multiple sampled image patches according to a preset size. The multiple sampled image patches are input into a preset feature extraction model for feature extraction to obtain the semantic feature vector of each sampled image patch. The semantic feature vectors of each sampled image patch are stacked in order to obtain the second semantic feature vector corresponding to the sampled image. The second semantic feature vector corresponding to each sampled image in the sampled image set is integrated and processed to obtain the second semantic feature vector set.

3. The image super-resolution reconstruction method according to claim 2, characterized in that, Calculating the similarity between each second semantic feature vector and the first semantic feature vector in the candidate feature library to obtain the similarity score between the image to be super-resolution and each of the high-resolution images includes: For each second semantic feature vector in the candidate feature library, the semantic feature vector of each of its patches is compared with the semantic feature vector of each patch in the first semantic feature vector to calculate the pairwise similarity, so as to obtain the original similarity between each pair of patches. The original similarity scores are accumulated and then divided by the number of patches in the corresponding high-definition image to obtain the similarity score between the image to be super-resolution and the corresponding high-definition image.

4. The image super-resolution reconstruction method according to claim 3, characterized in that, The calculation of the original similarity is expressed as follows: in, Let be the semantic feature vector of the m-th patch of the image to be super-resolution. This is the semantic feature vector of the nth patch corresponding to the high-resolution image. For vector dot product operation, This is a vector transpose operation. and They are respectively and The modulus length; The similarity score between the i-th high-resolution image and the image to be super-resolution is calculated as follows: in, Let be the number of patches for the i-th high-resolution image. Let be the semantic feature vector of the nth patch of the i-th high-resolution image.

5. The image super-resolution reconstruction method according to claim 2, characterized in that, The preset feature extraction model is the DINOv2 model with shared weights.

6. The image super-resolution reconstruction method according to claim 1, characterized in that, Selecting a number of the high-resolution images as a dynamic training set based on the similarity scores includes: Sort all similarity scores in descending order, and select the high-resolution images corresponding to the top few similarity scores to construct a dynamic training set.

7. The image super-resolution reconstruction method according to claim 6, characterized in that, Selecting several first feature vectors with the highest similarity scores and obtaining their corresponding sampled images, and obtaining high-resolution images of the corresponding sampled images as a dynamic training set, further includes: If the acquired high-resolution image has already been selected as a dynamic training set by other images to be super-resolution, then skip the high-resolution image and continue filtering according to the similarity score.

8. An image super-resolution reconstruction system, employing any one of claims 1-7, characterized in that, include: Feature extraction module: used to acquire the image to be super-resolution, perform feature extraction processing on the image to be super-resolution, and obtain the first semantic feature vector; Candidate feature library construction module: used to acquire a set of high-definition images, downsample the set of high-definition images at a preset super-resolution magnification ratio to obtain a set of sampled images, perform feature extraction processing on the set of sampled images to obtain a second set of semantic feature vectors, and construct a candidate feature library based on the second set of semantic feature vectors; Similarity calculation module: used to calculate the similarity between each second semantic feature vector and the first semantic feature vector in the candidate feature library, and to obtain the similarity score between the image to be super-resolution and each of the high-definition images; Model training module: used to select several high-definition images as a dynamic training set based on the similarity score, input the dynamic training set into a preset diffusion model for training, and obtain a trained diffusion model; Super-resolution module: Used to input the image to be super-resolution into the trained diffusion model and output a super-resolution image.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 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 steps of the method according to any one of claims 1 to 7.

Images