Super-resolution reconstruction method and system based on diffusion model

Abstract

The invention relates to a super-resolution reconstruction system method and system based on a diffusion model. The method comprises the following steps: constructing a data set; training a u-net network through the data set to obtain a diffusion model; inputting a to-be-degraded image into the diffusion model, and performing model reasoning on the diffusion model to obtain a degraded image; training based on the diffusion model and the degraded image to obtain a super-resolution reconstruction model; and reconstructing an image to be reconstructed through the super-resolution reconstruction model to obtain a reconstruction result. According to the method, the precision of super-resolution reconstruction is improved, and the quality of a reconstruction result is improved.

Description

Technical Field

[0001] The invention belongs to the field of image processing, and in particular relates to a super-resolution reconstruction system method and system based on a diffusion model. Background Art

[0002] Super-resolution reconstruction is a computer vision technique that aims to reconstruct a high-resolution image from a low-resolution image. Its main goal is to improve the clarity and visual quality of an image by increasing the number and details of pixels. Super-resolution technology plays an important role in many application scenarios, such as medical image processing, satellite image analysis, video enhancement, and security monitoring.

[0003] At present, super-resolution reconstruction technology based on deep learning mainly uses high-resolution-low-resolution (HR-LR) image data pairs for training. There are two main ways to obtain them: one is to degrade the image by methods such as bicubic downsampling, thereby synthesizing training data; the other is to collect HR-LR image data pairs from actual scenes.

[0004] For the above methods, the typical algorithms in the first method include SRCNN, RealESRGAN, etc. SRCNN performs bicubic downsampling on the image to obtain low-resolution image training data. RealESRGAN uses multiple degradation types of input methods in order to simulate more realistic degradation effects. The degradation methods use Gaussian, generalized Gaussian, sinc filtering, jpeg compression, etc. Different hyperparameters are used for each degradation. The advantage of the first method is that it is relatively simple to obtain data, and HR and LR can be aligned pixel by pixel. However, the disadvantage of the low-resolution training images generated by the first method is that it uses one or several pre-defined downsampling methods, which are quite different from the data degradation methods in real scenes. The data degradation in real scenes is often more complicated, and different data sets often have different degradation methods due to different acquisition devices, which are difficult to summarize using simple degradation modeling.

[0005] For the second method, typical data sets include RealSR and DRealSR, which use cameras to shoot paired HR-LR data pairs and use software for registration. The advantage is that both high-resolution and low-resolution image data are obtained from the real world, which conforms to the actual degradation process. However, it requires a lot of manpower to obtain, and the data cost is high. In addition, the obtained HR-LR image data pairs may have misalignment problems due to reliance on post-registration. At the same time, in order to obtain images of different resolution multiples of the same scene, different focal lengths need to be used for shooting, which introduces problems such as lens distortion. Summary of the invention

[0006] In view of the above deficiencies in the prior art, an object of the present invention is to provide a method and system for super-resolution reconstruction based on a diffusion model.

[0007] The present invention provides a super-resolution reconstruction method based on a diffusion model, comprising:

[0008] S1: Construct dataset;

[0009] S2: train the u-net network through the data set to obtain a diffusion model;

[0010] S3: inputting the image to be degraded into the diffusion model, and obtaining a degraded image by performing model reasoning on the diffusion model;

[0011] S4: obtaining a super-resolution reconstruction model based on the diffusion model and the degraded image training;

[0012] S5: Reconstruct the image to be reconstructed by using the super-resolution reconstruction model to obtain a reconstruction result.

[0013] According to a super-resolution reconstruction method based on a diffusion model provided by the present invention, the data set in step S1 includes:

[0014] High-resolution images;

[0015] A low-resolution image of a field of view corresponding to each downsampling multiple of the high-resolution image. According to a super-resolution reconstruction method based on a diffusion model provided by the present invention, step S2 further comprises:

[0016] S21: selecting a set of data pairs from the data set, wherein the data pairs include a high-resolution image and a single low-resolution image;

[0017] S22: downsampling the high-resolution image in the data pair by a downsampling method according to the downsampling multiple of the low-resolution image in the data pair to obtain a first downsampled image;

[0018] S23: inputting the first down-sampled image and the low-resolution image in the data pair into a potential encoder to obtain a first down-sampled code and a first low-resolution code respectively;

[0019] S24: Based on the positive time steps, starting from the first low-resolution encoding, the u-net network is gradient updated until the u-net network converges to obtain a diffusion model.

[0020] According to a super-resolution reconstruction method based on a diffusion model provided by the present invention, the downsampling method in step S22 includes bicubic downsampling and nearest neighbor downsampling.

[0021] According to a super-resolution reconstruction method based on a diffusion model provided by the present invention, the expression for the gradient update of the u-net network in step S24 is:

[0022]

[0023] Among them, θ is the model parameter, is the gradient operator, ε is the sampling noise, ε θ is the diffusion model, Z′ HR is the first down-sampling code, γ is the noise scheduling parameter, is the first low-resolution encoding at time step t=0, and ||·||| is the vector norm.

[0024] According to a super-resolution reconstruction method based on a diffusion model provided by the present invention, step S3 further comprises:

[0025] S31: inputting the image to be degraded, and downsampling the image to be degraded by a downsampling method to obtain a second downsampled image;

[0026] S32: inputting the second down-sampled image into a potential encoder to obtain second down-sampled codes respectively;

[0027] S33: sampling from a preset Gaussian distribution to obtain a second low-resolution code;

[0028] S34: Based on the reverse time step, starting from the second low-resolution code, the diffusion model is cyclically updated to obtain a third low-resolution code corresponding to the initial time step;

[0029] S35: Input the third low-resolution code into a potential decoder to obtain a low-resolution result image.

[0030] According to a super-resolution reconstruction method based on a diffusion model provided by the present invention, when the diffusion model is cyclically updated in step S34, when the current time step t>1, random noise is sampled from a preset Gaussian distribution, and when the current time step t≤1, the random noise is equal to 0.

[0031] According to a super-resolution reconstruction method based on a diffusion model provided by the present invention, the expression for cyclically updating the diffusion model in step S34 is:

[0032]

[0033] Where t is the sampling time step, is the second lowest resolution code at time step t, is the updated low-resolution code at time step t-1, ε θ is the diffusion model, αt is the adjustment factor at time step t, γ t is the noise scheduling parameter at time step t, Z′ is the second downsampling code, and z is the random noise obtained by Gaussian distribution sampling.

[0034] According to a super-resolution reconstruction method based on a diffusion model provided by the present invention, step S4 further comprises:

[0035] S41: re-collect high-resolution images for training;

[0036] S42: inputting the high-resolution training image into the diffusion model to obtain a low-resolution training image;

[0037] S43: A super-resolution reconstruction model is obtained by training a training data set consisting of multiple groups of training high-resolution images and training low-resolution images.

[0038] The present invention also provides a super-resolution reconstruction system based on a diffusion model, which is used to execute a super-resolution reconstruction method based on a diffusion model as described in any one of the above items, comprising:

[0039] Building module: used to build data sets;

[0040] The first training module is used to train the u-net network through the data set to obtain a diffusion model;

[0041] Reasoning module: used for inputting the image to be degraded into the diffusion model, and obtaining the degraded image by performing model reasoning on the diffusion model;

[0042] A second training module: used for obtaining a super-resolution reconstruction model based on the diffusion model and the degraded image training;

[0043] Reconstruction module: used to reconstruct the image to be reconstructed through the super-resolution reconstruction model to obtain a reconstruction result.

[0044] The present invention provides a super-resolution reconstruction method and system based on a diffusion model. By constructing a data set including high-resolution images and their corresponding low-resolution images with different downsampling multiples, and using these data to train a u-net network to obtain a diffusion model, the mapping relationship between high-resolution images and low-resolution images can be effectively learned. The super-resolution reconstruction model trained based on the diffusion model can more accurately reconstruct high-quality images, thereby improving the accuracy of super-resolution reconstruction. Secondly, when training the u-net network, the present invention adopts a variety of downsampling methods, which helps the model learn more abundant image degradation features, thereby enhancing the generalization ability of the model. When encountering new degraded images that have not appeared in the training set, the model can still better process and reconstruct high-quality images. In addition, through the diffusion model, the present invention can flexibly simulate the degradation process of the image. The present invention can input The image to be degraded is subjected to downsampling, encoding and other processing, so as to generate a low-resolution image similar to the actual degradation process, which provides input data closer to the actual situation for subsequent super-resolution reconstruction, and helps to improve the quality of the reconstruction result; when training the super-resolution reconstruction model, the present invention uses degraded images based on the diffusion model as training data, which not only simplifies the preparation process of the training data, but also the degraded images generated by the diffusion model are closer to the actual situation, which helps the model to learn a more effective reconstruction strategy. At the same time, through reasonable gradient update and loop update strategies, the training process of the model is further optimized; the super-resolution reconstruction system based on the diffusion model provided by the present invention realizes a complete process from data set construction to model training to image reconstruction through the collaborative work of various modules, and the systematic design not only improves the overall performance of the system, but also makes the entire process more efficient and stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings are only used to illustrate specific embodiments and are not considered to limit the present invention. In the entire drawings, the same reference symbols represent the same components. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0046] Figure 1 A schematic flow chart of a super-resolution reconstruction method based on a diffusion model provided by an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of the structure of a super-resolution reconstruction system based on a diffusion model provided by an embodiment of the present invention.

[0048] Reference numerals:

[0049] 100, construction module; 200, first training module; 300, reasoning module; 400, second training module; 500, reconstruction module. DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making creative work should fall within the scope of protection of the present invention.

[0051] Furthermore, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily obscuring the concepts disclosed in the present invention.

[0052] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the orientation or position relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inside", "outside", etc. is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. The terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0053] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of methods and systems consistent with some aspects of the present invention as detailed in the appended claims.

[0054] Combine the following Figure 1 and Figure 2 Embodiments of the present invention are described.

[0055] like Figure 1As shown, the present invention provides a super-resolution reconstruction method based on a diffusion model, comprising:

[0056] S1: Build the dataset.

[0057] The data set in step S1 includes:

[0058] A high-resolution image; and a low-resolution image of a field of view corresponding to each downsampling multiple of the high-resolution image.

[0059] In step S1, a real-world super-resolution reconstruction dataset is first constructed, such as public datasets such as RealSR and DRealSR, or it can be collected by yourself. The dataset has the following structure: high-resolution (HR) images; for each downsampling factor s, there is a low-resolution (LR) image of the corresponding field of view. After the dataset is constructed, it is also necessary to divide the dataset into training set, test set, and validation set according to 7:2:1, and determine that the downsampling factor for the model training is s.

[0060] S2: Train the u-net network using the data set to obtain a diffusion model.

[0061] The overall goal of step S2 is to train a u-net network with a given data set so that it can learn the mapping relationship from low-resolution images to high-resolution images, and then generate a diffusion model. The diffusion model is a generative model that gradually generates data from noise. In the present invention, it generates high-resolution images from low-resolution images.

[0062] Wherein, step S2 further comprises:

[0063] S21: Selecting a set of data pairs from the data set, wherein the data pairs include a high-resolution image and a single low-resolution image.

[0064] S22: Downsampling the high-resolution image in the data pair by a downsampling method according to the downsampling multiple of the low-resolution image in the data pair to obtain a first down-sampled image.

[0065] The downsampling method in step S22 includes bicubic downsampling and nearest neighbor downsampling.

[0066] In order to simulate the degradation process from high resolution to low resolution, in step S22, the high resolution image is downsampled using the same downsampling multiple as the low resolution image. The downsampling method can be bicubic downsampling or nearest neighbor downsampling, etc. The method determines the way the image is reduced. Bicubic downsampling usually produces a smoother image, while nearest neighbor downsampling retains more edges and details.

[0067] S23: Input the first down-sampled image and the low-resolution image in the data pair into a potential encoder to obtain a first down-sampled code and a first low-resolution code, respectively.

[0068] The above-mentioned latent encoder is a pre-trained neural network module, which is used to convert an image into a code. In step S23, the first down-sampled image and the low-resolution image are respectively input into the latent encoder to obtain their code representations. The obtained code is the basis for subsequent u-net network processing.

[0069] S24: Based on the positive time steps, starting from the first low-resolution encoding, the u-net network is gradient updated until the u-net network converges to obtain a diffusion model.

[0070] In step S24, the first low-resolution encoding is used as a starting point, and higher-resolution image representations are gradually generated through the u-net network until the target resolution (i.e., the resolution of the original high-resolution image) is reached to obtain a diffusion model. In the specific process, the u-net network updates its parameters by gradient descent to reduce the difference between the generated image and the true high-resolution image. When the network performance is no longer significantly improved (i.e., converged), the training process ends, and the obtained u-net network is the desired diffusion model.

[0071] Among them, the expression of the gradient update of the u-net network in step S24 is:

[0072]

[0073] Among them, θ is the model parameter, is the gradient operator, ε is the sampling noise, ε θ is the diffusion model, Z′ HR is the first down-sampling code, γ is the noise scheduling parameter, is the first low-resolution encoding at time step t=0, and ||·||| is the vector norm.

[0074] In step S2 and corresponding steps S21 to S24, firstly, a pair of HR-LR data is taken from the training set, denoted as I HR and I LR , and to I HR Perform downsampling to obtain I′ HR , the downsampling operation randomly selects one of the bicubic downsampling and the nearest neighbor downsampling, and the downsampling multiple is s. At this time, I′ HR and I LR The same size, I′ HR and I LR Input into the pre-trained latent encoder and output the latent code Z′ HR and ZLR .

[0075] From Z LR At the beginning, the time step is t = 0, recorded as The time step t is sampled from a uniform distribution (1, T), the noise ε is sampled from a Gaussian distribution N(0, I), and a noise scheduling parameter γ is sampled from a predefined distribution. This parameter is related to t and is used to control the intensity of the noise.

[0076] Then, the gradient descent update is performed according to the above formula, where ε θ For the conditional diffusion model, repeat the above steps until the model converges and obtain the diffusion model.

[0077] S3: inputting the image to be degraded into the diffusion model, and obtaining a degraded image by performing model inference on the diffusion model.

[0078] Wherein, step S3 further comprises:

[0079] S31: inputting the image to be degraded, and downsampling the image to be degraded by a downsampling method to obtain a second downsampled image.

[0080] In step S31, a downsampling method is first applied to the image to be degraded, and the downsampling method is consistent with the downsampling method in step S2 to obtain an image version with lower resolution, namely the second downsampled image. This step is to simulate a preliminary stage of image degradation.

[0081] S32: Input the second down-sampled image into a potential encoder to obtain second down-sampled codes.

[0082] S33: Sampling from a preset Gaussian distribution to obtain a second low-resolution code.

[0083] In step S33, a preset Gaussian distribution is randomly sampled to obtain an initial low-resolution code, ie, a second low-resolution code, which will serve as the starting point of the inverse process of the diffusion model to gradually generate a coarser image representation.

[0084] S34: Based on the reverse time step, starting from the second low-resolution code, the diffusion model is cyclically updated to obtain a third low-resolution code corresponding to the initial time step.

[0085] Wherein, when the diffusion model is cyclically updated in step S34, when the current time step t>1, random noise is sampled from a preset Gaussian distribution, and when the current time step t≤1, the random noise is equal to 0.

[0086] Step S34, starting from the second low-resolution encoding, the diffusion model is cyclically updated in reverse time steps, that is, from coarser to finer resolutions. In each step, the low-resolution encoding of the current time step is updated according to the adjustment factor of the current time step, the noise scheduling parameter and the random noise sampled from the Gaussian distribution (at the last time step, the random noise is 0). The cyclic update process continues until the initial time step is reached. The low-resolution encoding obtained at this time (that is, the third low-resolution encoding mentioned above) represents the representation of the final degraded image in the encoding space.

[0087] The expression for cyclically updating the diffusion model in step S34 is:

[0088]

[0089] Where t is the sampling time step, is the second lowest resolution code at time step t, is the updated low-resolution code at time step t-1, ε θ is the diffusion model, α t is the adjustment factor at time step t, γ t is the noise scheduling parameter at time step t, Z′ is the second downsampling code, and z is the random noise obtained by Gaussian distribution sampling.

[0090] S35: Input the third low-resolution code into a potential decoder to obtain a low-resolution result image.

[0091] In step S35, the third low-resolution code is input into the inverse process of the potential encoder, ie, the decoder, to obtain a degraded low-resolution image. The obtained image is the final result obtained after the image to be degraded is inversely processed by the diffusion model.

[0092] Step S3 and the corresponding steps S31 to S35 input the target image to be degraded, perform downsampling operation on the image to be degraded, input the downsampled image into the pre-trained latent encoder, and output the latent code Z ′ ; Subsequent sampling from Gaussian distribution N(0,I) Subsequently, from time step t = T to 1, the following loop is executed: if the current time step t>1, sample z from the Gaussian distribution N(0,I), otherwise set z = 0 and update the above expression; finally, the final result obtained by loop update Input into the pre-trained potential decoder, output return This is the generated low-resolution training image I D .

[0093] S4: Obtain a super-resolution reconstruction model based on the diffusion model and the degraded image training.

[0094] Wherein, step S4 further comprises:

[0095] S41: Re-collect high-resolution images for training.

[0096] S42: Inputting the high-resolution training image into the diffusion model to obtain a low-resolution training image.

[0097] S43: A super-resolution reconstruction model is obtained by training a training data set consisting of multiple groups of training high-resolution images and training low-resolution images.

[0098] In step S4 and the corresponding steps S41 to S43, image data is first collected for super-resolution model training. The data collected here does not need to be HR-LR image data pairs under the same field of view. Only one high-resolution image is required for the same field of view. The diffusion model trained above is used to infer each of the above data to generate a low-resolution image, and then HR-LR data pairs are constructed, and all processed data pairs are regarded as super-resolution reconstruction data sets. Finally, the super-resolution reconstruction model is trained using the super-resolution reconstruction data set.

[0099] S5: Reconstruct the image to be reconstructed by using the super-resolution reconstruction model to obtain a reconstruction result.

[0100] like Figure 2 As shown, the present invention also provides a super-resolution reconstruction system based on a diffusion model, comprising:

[0101] Construction module 100: used to construct a data set;

[0102] The first training module 200 is used to train the u-net network through the data set to obtain a diffusion model;

[0103] Reasoning module 300: used for inputting the image to be degraded into the diffusion model, and obtaining the degraded image by performing model reasoning on the diffusion model;

[0104] The second training module 400 is used to obtain a super-resolution reconstruction model based on the diffusion model and the degraded image training;

[0105] Reconstruction module 500: used to reconstruct the image to be reconstructed by using the super-resolution reconstruction model to obtain a reconstruction result.

[0106] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0107] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0108] The present invention proposes a super-resolution reconstruction method and system based on a diffusion model, which can obtain a large amount of high-quality training data at a low cost and improve the diversity and realism of the data. Through this algorithm, the following problems can be solved at the same time:

[0109] In previous solutions, if you want to ensure the authenticity of low-resolution degraded image data, you usually need to collect high-resolution (HR) and low-resolution (LR) image pairs at the same time. This process not only consumes a lot of manpower and material resources, but also involves complex image registration problems (ensuring that LR and HR images correspond pixel by pixel). The synthesis method based on the diffusion model proposed in the present invention does not need to collect HR-LR image pairs, but can automatically generate corresponding LR images through HR images. This greatly simplifies the data acquisition process and significantly reduces the cost of super-resolution model training data.

[0110] In previous solutions, if you want to use data synthesis methods to generate training data in batches at low cost, you usually use predefined degradation operations, such as bicubic downsampling, fuzzy filtering, noise addition, etc., but these degradation methods are often different from the image degradation in actual scenes, resulting in poor performance of the trained model in real applications. In contrast, the present invention is based on a diffusion model, which directly generates LR images from HR images by learning the degradation patterns in real data. This generation process is not a simple mathematical downsampling operation, but a diffusion model that simulates the degradation process of the image in a real environment. Therefore, the generated LR image is closer to the image quality distribution in actual applications.

[0111] The present invention proposes a super-resolution reconstruction method and system based on a diffusion model. First, a diffusion model is used to generate a low-resolution training image: a process of generating a low-resolution (LR) image from a high-resolution (HR) image using a diffusion model. Different from the traditional predefined degradation methods (such as bicubic interpolation, adding noise, etc.), the diffusion model simulates the degradation process in a real environment by gradually adding and removing noise. The model can flexibly adjust the generated image details and noise levels, and the generated low-resolution image is more consistent with the actual image degradation distribution, thereby improving the reconstruction of the super-resolution model in real scenes; secondly, low-cost data acquisition of HR-LR image pairs is not required: in traditional super-resolution training, in order to obtain real degraded pictures, considering that the previous synthesis method cannot well summarize all degradation methods, it is usually necessary to collect HR and LR image pairs in reality. The present invention can automatically generate LR images from HR images through a synthesis technology based on a diffusion model, and obtain images that are more consistent with the actual degradation distribution without the need to collect HR and LR images at the same time. This technology reduces the complexity of registering high-resolution and low-resolution images, and also reduces the cost of collecting images. It also improves data diversity and coverage: the diffusion model can generate multiple different LR images for the same HR image through random sampling of noise. By changing the intensity and characteristics of the noise, a variety of different degradation conditions can be synthesized. These different low-resolution images differ in details, which enhances the diversity of training data, improves the generalization ability of the model, and enables the model to perform better in complex scenes.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the above embodiments, or replace some of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or replacements that can be easily thought of by those skilled in the art within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. A super-resolution reconstruction method based on a diffusion model, characterized in that: include: S1: Construct dataset; S2: train the u-net network through the data set to obtain a diffusion model; S3: inputting the image to be degraded into the diffusion model, and obtaining a degraded image by performing model reasoning on the diffusion model; S4: obtaining a super-resolution reconstruction model based on the diffusion model and the degraded image training; S5: Reconstruct the image to be reconstructed by using the super-resolution reconstruction model to obtain a reconstruction result.

2. The super-resolution reconstruction method based on the diffusion model according to claim 1, characterized in that: The data set in step S1 includes: High-resolution images; A low-resolution image of a field of view corresponding to each downsampling multiple of the high-resolution image.

3. The super-resolution reconstruction method based on a diffusion model according to claim 1, characterized in that: Step S2 further comprises: S21: selecting a set of data pairs from the data set, wherein the data pairs include a high-resolution image and a single low-resolution image; S22: downsampling the high-resolution image in the data pair by a downsampling method according to the downsampling multiple of the low-resolution image in the data pair to obtain a first downsampled image; S23: inputting the first down-sampled image and the low-resolution image in the data pair into a potential encoder to obtain a first down-sampled code and a first low-resolution code respectively; S24: Based on the positive time steps, starting from the first low-resolution encoding, the u-net network is gradient updated until the u-net network converges to obtain a diffusion model.

4. The super-resolution reconstruction method based on the diffusion model according to claim 3, characterized in that: The downsampling method in step S22 includes bicubic downsampling and nearest neighbor downsampling.

5. The super-resolution reconstruction method based on diffusion model according to claim 3, characterized in that: The expression for the gradient update of the u-net network in step S24 is: Among them, θ is the model parameter, is the gradient operator, ε is the sampling noise, ε θ is the diffusion model, Z ′ HR is the first down-sampling code, γ is the noise scheduling parameter, is the first low-resolution encoding at time step t=0, and ||·||| is the vector norm.

6. The super-resolution reconstruction method based on a diffusion model according to claim 1, characterized in that: Step S3 further comprises: S31: inputting the image to be degraded, and downsampling the image to be degraded by a downsampling method to obtain a second downsampled image; S32: inputting the second down-sampled image into a potential encoder to obtain second down-sampled codes respectively; S33: sampling from a preset Gaussian distribution to obtain a second low-resolution code; S34: Based on the reverse time step, starting from the second low-resolution code, the diffusion model is cyclically updated to obtain a third low-resolution code corresponding to the initial time step; S35: Input the third low-resolution code into a potential decoder to obtain a low-resolution result image.

7. The super-resolution reconstruction method based on the diffusion model according to claim 6, characterized in that: When the diffusion model is cyclically updated in step S34, when the current time step t>1, random noise is sampled from a preset Gaussian distribution, and when the current time step t≤1, the random noise is equal to 0.

8. The super-resolution reconstruction method based on diffusion model according to claim 6, characterized in that: The expression for cyclically updating the diffusion model in step S34 is: Where t is the sampling time step, is the second lowest resolution code at time step t, is the updated low-resolution code at time step t-1, ε θ is the diffusion model, α t is the adjustment factor at time step t, γ t is the noise scheduling parameter at time step t, Z′ is the second downsampling code, and z is the random noise obtained by Gaussian distribution sampling.

9. The super-resolution reconstruction method based on diffusion model according to claim 1, characterized in that: Step S4 further comprises: S41: re-collect high-resolution images for training; S42: inputting the high-resolution training image into the diffusion model to obtain a low-resolution training image; S43: A super-resolution reconstruction model is obtained by training a training data set consisting of multiple groups of training high-resolution images and training low-resolution images.

10. A super-resolution reconstruction system based on a diffusion model, used to execute a super-resolution reconstruction method based on a diffusion model according to any one of claims 1 to 9, characterized in that: include: Building module: used to build data sets; The first training module is used to train the u-net network through the data set to obtain a diffusion model; Reasoning module: used for inputting the image to be degraded into the diffusion model, and obtaining the degraded image by performing model reasoning on the diffusion model; A second training module: used for obtaining a super-resolution reconstruction model based on the diffusion model and the degraded image training; Reconstruction module: used to reconstruct the image to be reconstructed through the super-resolution reconstruction model to obtain a reconstruction result.

Images