Terahertz image restoration method, system and equipment based on diffusion probability model

Abstract

The invention relates to a terahertz image restoration method, device and equipment based on a diffusion probability model. Firstly, an input terahertz image is subjected to preprocessing and noise adding to obtain a noise graph, then the optimal mean value and the optimal variance of the noise graph in the back diffusion process are estimated in real time through a pre-trained diffusion probability model, and finally rapid sampling is performed to generate a restored image. Compared with a traditional method (such as filtering processing and super-resolution reconstruction technologies), the restoration method based on the diffusion probability model is better in performance when processing noise and artifacts, interference can be removed more effectively, and image definition and contrast are improved. Moreover, an optimal variance estimation and rapid sampling strategy is introduced, so that the sampling steps can be reduced, an image with higher quality can be generated, and better detail reservation and texture optimization can be realized, thereby remarkably improving the quality and efficiency of terahertz image restoration.

Description

Technical Field

[0001] The present application relates to the technical field of terahertz image processing, and in particular to a terahertz image restoration method, system and device based on a diffusion probability model. Background Art

[0002] With the rapid development of modern science and technology, terahertz imaging technology has been widely used in security inspection, non-destructive testing, biomedicine and other fields due to its unique non-ionizing and penetrating properties, and its ability to ensure that the detected objects and human body are not harmed.

[0003] Existing terahertz imaging technology still faces problems such as low resolution and image quality. These problems make it difficult for imaging results to meet the needs of practical applications. For example, images are often accompanied by strong noise, streak artifacts, low contrast, and blurred edges. In order to solve these problems, a variety of image denoising and restoration algorithms have been proposed to improve the quality of terahertz images. Common methods include filtering and super-resolution reconstruction techniques, which improve the quality and usability of images by removing noise, enhancing contrast, and reconstructing details.

[0004] However, the above-mentioned traditional terahertz image restoration method has technical problems of low restoration quality and low restoration efficiency. Summary of the invention

[0005] Based on this, it is necessary to provide a terahertz image restoration method based on a diffusion probability model, a terahertz image restoration system based on a diffusion probability model and a computer device to address the above technical problems.

[0006] In order to achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0007] On the one hand, a terahertz image restoration method based on a diffusion probability model is provided, comprising:

[0008] A terahertz image is input, and preprocessed and noise-added are performed on the terahertz image to obtain a noise map;

[0009] Based on the pre-trained diffusion probability model, the optimal mean and optimal variance of the noise map in the reverse diffusion process are estimated in real time through the score function; the score function is one of the outputs of the pre-trained diffusion probability model;

[0010] According to the optimal mean and optimal variance, the noise map is quickly sampled through the diffusion probability model solver to obtain the restored image.

[0011] In one of the embodiments, the method further includes: extracting prior information of the restored image, and performing texture refinement on the restored image according to the prior information using an unconditional diffusion model and a truncated sampling strategy to obtain a detail enhanced image.

[0012] In one of the embodiments, the method further includes: performing adaptive fusion processing on multimodal feature information of the detail enhanced image to obtain an optimized restored image.

[0013] On the other hand, a terahertz image restoration system based on a diffusion probability model is also provided, comprising:

[0014] An input module, used for inputting a terahertz image;

[0015] A processing module, used for preprocessing and adding noise to the terahertz image to obtain a noise map;

[0016] An estimation module, which is used to estimate the optimal mean and optimal variance of the noise map in the reverse diffusion process in real time based on the pre-trained diffusion probability model through a scoring function; the scoring function is one of the outputs of the pre-trained diffusion probability model;

[0017] The sampling module quickly samples the noise image through the diffusion probability model solver according to the optimal mean and optimal variance to obtain the restored image;

[0018] The output module is used to output the restored image.

[0019] On the other hand, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned terahertz image restoration method based on the diffusion probability model when executing the computer program.

[0020] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0021] The above-mentioned terahertz image restoration method, system and device based on the diffusion probability model first pre-process and add noise to the input terahertz image to obtain a noise map, and then, through the pre-trained diffusion probability model, the optimal mean and optimal variance of the noise map in the reverse diffusion process are estimated in real time, and finally the restored image is generated by fast sampling. Traditional methods (such as filtering and super-resolution reconstruction technology) often perform poorly in dealing with noise and artifacts, resulting in impaired image quality, while the restoration method based on the diffusion probability model can better capture and remove these interferences by accurately estimating the noise characteristics, thereby improving the clarity and contrast of the image; and by introducing the optimal variance estimation and fast sampling strategy, it can generate higher quality images while reducing the sampling steps, achieve better detail preservation and texture optimization, and greatly improve the quality and efficiency of terahertz image restoration. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the conventional technology, the drawings required for use in the embodiments or the conventional technology descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 is a schematic flow chart of a terahertz image restoration method based on a diffusion probability model in one embodiment;

[0024] Figure 2 Schematic diagram of the effect after using IRM (identity restoration module) and TEM (texture enhancement module) in one embodiment, wherein (a) represents an input image, (b) represents an image processed by IRM (identity restoration module), and (c) represents an image processed by TEM (texture enhancement module);

[0025] Figure 3 This is a diagram showing the effect of introducing multimodal information fusion in an embodiment, wherein (a) represents the input multimodal feature information, and (b) represents the output optimized restored image. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0028] It should be noted that the reference to "embodiment" herein means that the specific features, structures or characteristics described in conjunction with the embodiment may be included in at least one embodiment of the present invention. The presentation of the phrase at various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments.

[0029] The following will describe the implementation of the present invention in detail with reference to the accompanying drawings in the embodiment diagram of the present invention.

[0030] In one embodiment, Figure 1 As shown, the embodiment of the present application provides a terahertz image restoration method based on a diffusion probability model, including the following processing steps S12-16:

[0031] S12, inputting a terahertz image, and preprocessing and adding noise to the terahertz image to obtain a noise map.

[0032] It can be understood that the acquired terahertz image data set as input may have characteristics such as low resolution, blurred edges, and obvious noise, and can be of different sizes, such as 200×200 pixels, 500×500 pixels, etc., to adapt to various practical application scenarios. When the image data is input, it will also be pre-processed, such as normalization and removal of some stray signals to improve the effect of subsequent processing. After pre-processing, noise is added. Specifically, Gaussian noise, Poisson noise or white noise can be added. In this embodiment and the following text, Gaussian noise is taken as an example: The pixel channel value of the mixed image: Where ∈ is Gaussian noise, β is a random value, and x is the image information. Finally, the noise map is obtained.

[0033] S14, based on the pre-trained diffusion probability model, estimates the optimal mean and optimal variance of the noise map in the reverse diffusion process in real time through the scoring function; the scoring function is one of the outputs of the pre-trained diffusion probability model.

[0034] It can be understood that based on the pre-trained diffusion probability model, the optimal mean and optimal variance of the noise image in the reverse diffusion process are estimated in real time through the scoring function. As one of the outputs of the pre-trained model, the scoring function is used to quantify the feature distribution of the current noise image. First, the relationship between the noise image and the original image obtained by the system during the noise addition process is used to derive the mean and variance. This process can be achieved through Analytic-DPM (Analytical Diffusion Probability Model). The model uses the known distribution in the noise addition process to ensure that during reverse denoising, the generated distribution is consistent with the conditional distribution in the forward process, so that the denoising process is reasonable and feasible.

[0035] Specifically, the proportion of noise and the original image in the mixture increases and decreases, and βt gradually increases from a number closest to 0 to 1 in the end. Because the diffusion speed is getting faster and faster, the image at time t is expressed as:

[0036]

[0037] Can be rewritten as:

[0038]

[0039] α t =1-β t

[0040] Among them, α t Represents β t A variation of ∈ tIt obeys the normal distribution N(0,1), that is, the noise has a mean of 0 and a variance of 1. The forward process is from the original image x0 after adding noise to any x t The relationship between the moment images, the reverse process goal is to get from the last x t The original image at time x0 is restored from the noisy image at time x0. According to the noise adding process, p(x t |x0) and q(x t-1 |x0,x t ), reverse denoising uses p(x t-1 |x t ), we need to estimate the mean and variance of the distribution, and through derivation, we find that the constraint mean needs to be and q(x t-1 |x0,x t ) are consistent with the mean.

[0041] Among them, p(x t |x0) represents the image x at time point t under the condition of the pre-processed terahertz image x0. t The probability distribution of the noise image x0 to the noise image x t The relationship between q(x t-1 |x0,x t ) means that between x0 and x t Under the condition of t-1 The probability distribution of p(x t-1 |x t ) represents the image x at time point t t Under the condition, at time t-1, the image x t-1 The conditional probability distribution of describes how to transform the image x with more noise into t First, we use a step of denoising to get x closer to the original image. t-1 . Since q(x t-1 |x0,x t ) This distribution is known during the noise addition process, and we expect the p(x t-1 |x t ) and the conditional distribution q(x) during the noise addition process t-1 |x0,x t ) to ensure that the denoising process is reasonable. Analytic-DPM found that the optimal mean is consistent with the optimal mean of the diffusion probability model, and the optimal variance It can also be calculated analytically.

[0042] The core of this step is to introduce the Analytic-DPM method, which estimates the feature distribution of the input image in real time through the score function, and uses the relationship between image pixels to calculate the optimal mean and variance upper and lower bounds in the diffusion process. After the optimal variance estimation, an optimal variance matrix Σ is generated for each pixel. t , providing guidance for the subsequent sampling process, which significantly improves the image restoration effect.

[0043] S16, according to the optimal mean and the optimal variance, the noise image is quickly sampled through the diffusion probability model solver to obtain a restored image.

[0044] It can be understood that fast sampling is a key step in gradually recovering the original image from the noisy image using the diffusion probability model solver. In this process, the optimal mean and the optimal variance are the core elements (already obtained in the previous step). Through their estimation, the model can effectively guide the sampling of each step. Specifically, DPM-Solver is used, which is an efficient solver that reformulates the diffusion process as a problem of ordinary differential equations (ODE). Through the cleverly designed high-order ODE solver, DPM-Solver can generate high-quality samples within fewer sampling steps (for example, within 10-20 steps), thereby significantly reducing the sampling steps and ensuring high-quality restored images with minimal computation.

[0045] The uniqueness of DPM-Solver lies in its high-order ODE solver. Through a cleverly designed sampling path, it can complete the conversion from a completely noisy image to a clear restored image within 10-20 steps. Compared with the traditional diffusion model, this greatly reduces the sampling steps and significantly improves the computational efficiency. The noise prediction function is calculated at each moment based on the state variables of the input image, and combined with the guidance of the optimal variance, the noise amplitude is gradually reduced to ensure that each step is close to the true distribution of the target image. Finally, through the combination of iterative calculation and optimal variance, the model quickly samples a clear restored image.

[0046] The above-mentioned terahertz image restoration method based on the diffusion probability model first preprocesses and adds noise to the input terahertz image to obtain a noise map, and then estimates the optimal mean and optimal variance of the noise map in the reverse diffusion process in real time through the pre-trained diffusion probability model, and finally generates a restored image by fast sampling. Traditional methods (such as filtering and super-resolution reconstruction technology) often perform poorly in dealing with noise and artifacts, resulting in impaired image quality, while the restoration method based on the diffusion probability model can better capture and remove these interferences by accurately estimating the noise characteristics, thereby improving the clarity and contrast of the image; and by introducing the optimal variance estimation and fast sampling strategy, it can generate higher quality images while reducing the sampling steps, achieving better detail preservation and texture optimization, which greatly improves the quality and efficiency of terahertz image restoration.

[0047] In one embodiment, the above-mentioned terahertz image restoration method based on the diffusion probability model further includes the steps of: extracting prior information of the restored image, and performing texture refinement on the restored image according to the prior information using an unconditional diffusion model and a truncated sampling strategy to obtain a detail enhanced image.

[0048] It is understandable that the IRM (Identity Restoration Module) and TEM (Texture Enhancement Module) may be combined to add a detail enhancement step to further improve the image quality. Figure 2 A schematic diagram showing the effect of using IRM (identity restoration module) and TEM (texture enhancement module) in one embodiment, wherein (a) represents an input image, (b) represents an image processed by IRM (identity restoration module), and (c) represents an image processed by TEM (texture enhancement module).

[0049] First, by extracting the prior information of the restored image, such as edge features and high-frequency details, the foundation is laid for subsequent detail enhancement. In specific implementation, a convolutional neural network (CNN) is used to extract multi-scale features from the input image to ensure that key features of different scales can be captured. In this process, the kernel size and number of the convolution layer can be adjusted to achieve comprehensive extraction of details of different scales. These extracted edge and detail features will be embedded as guiding information in the inverse process of the diffusion model, thereby providing more accurate guidance in the image restoration process.

[0050] The texture of the restored image is further refined through the unconditional diffusion model and truncated sampling strategy. Specifically, the sampling strategy is adjusted using different time steps (such as 5, 10, and 20 steps) to enhance the image's performance in high-frequency details and overall texture. In each step, the model adjusts the generation process based on the previously extracted prior information so that the refined image is closer to the real data distribution. This meticulous adjustment not only improves the visual effect of the image, but also effectively retains the important features in the image, making the final detail-enhanced image significantly better in quality than traditional restoration methods.

[0051] In one embodiment, the above-mentioned terahertz image restoration method based on the diffusion probability model further includes the step of: performing adaptive fusion processing on the multimodal feature information of the detail enhanced image to obtain an optimized restored image.

[0052] It can be understood that in the above-mentioned terahertz image restoration method based on the diffusion probability model, after optimal variance estimation, fast sampling, and detail enhancement processing, multimodal information fusion is further introduced to improve the accuracy and realism of the image. Figure 3 The figure shows the effect after introducing multimodal information fusion in one embodiment, wherein (a) represents the input multimodal feature information, and (b) represents the output optimized restored image.

[0053] First, adaptive fusion processing is performed on the multimodal feature information (such as image content, detail texture, etc.) extracted from the detail enhanced image. In the specific implementation, a dynamic diffuser is used to optimize the image restoration effect by adaptively adjusting the contribution of each modality in combination with the importance of each modality feature.

[0054] During this process, the dynamic prediction mechanism generates an influence function to evaluate the influence of each mode on the restoration result. According to the characteristic characteristics of different modes in the input image, the influence function will guide the dynamic diffuser to adjust the weights of each mode during the restoration process, thereby achieving the organic fusion of different information sources. In this way, the restoration process can make full use of the complementarity of multimodal information, so that the final optimized restored image is more accurate in details and overall texture, improving the visual effect and realism. This multimodal information fusion mechanism significantly enhances the effect of image restoration and ensures that the generated image is more in line with the actual situation.

[0055] In one embodiment, the optimal mean value of the terahertz image restoration method based on the diffusion probability model is:

[0056]

[0057] in, represents the score function, x n Representing the noise map at step n:

[0058]

[0059] α n =1-β n

[0060] Among them, ∈ n Indicates that Gaussian noise conforms to the normal distribution N(0,1), β is a random value, and α is a variation of β;

[0061] The optimal variance is:

[0062]

[0063] in, represents the variance parameter, which can be directly taken for the diffusion probability model represents the expectation of the noise distribution, and d represents the dimension of the noise map.

[0064] In the above-mentioned terahertz image restoration method based on the diffusion probability model, Gaussian noise is used for noise addition, which can effectively simulate the noise interference in the real environment, thereby improving the generalization ability and restoration effect of the model in the denoising process. At the same time, the reversibility of Gaussian noise enhances the accuracy of restoring the original image during the reverse diffusion process.

[0065] In one embodiment, in the above-mentioned terahertz image restoration method based on the diffusion probability model, the sampling path of fast sampling is:

[0066]

[0067] It can be understood that the ODE (ordinary differential equation) of the noise prediction model can be defined as:

[0068]

[0069] Among them, h θ (x t ,t) means at the current time t x t The state change of f(t) represents the linear term control of state transfer, and g(t) represents the adjustment term of noise characteristics; ∈ θ (x t ,t) represents the noise prediction function, x t It follows a Gaussian distribution with a mean of 0 in time T and a covariance of

[0070] By solving the ODE from T to 0 to draw samples, the solution at time t can be exactly expressed by the constant change as:

[0071]

[0072] The high-order ordinary differential equation solver in DPM-Solver is used to calculate the sampling path according to the optimal variance matrix in each step:

[0073]

[0074] t represents the time variable, sampling from T to 0, ∈ θ (x t ,t) represents the noise prediction function, x t Represents the noise map at the current time t; α t represents the random noise β at the current time t t Variant of ; σ t Represents the optimal standard deviation at the current time t, according to the optimal variance Get;h i Indicates time interval Represents the time difference between two adjacent steps in the sampling process; represents a higher-order error term.

[0075] Given an initial value and a time step, reduce from t0 = T to t M =0, from At the beginning, the sequence is iteratively calculated, gradually approaching the true distribution of the target image.

[0076] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0077] In one embodiment, a terahertz image restoration system based on a diffusion probability model is provided, comprising: an input module for inputting a terahertz image;

[0078] A processing module, used for preprocessing and adding noise to the terahertz image to obtain a noise map;

[0079] An estimation module, which is used to estimate the optimal mean and optimal variance of the noise map in the reverse diffusion process in real time based on the pre-trained diffusion probability model through a scoring function; the scoring function is one of the outputs of the pre-trained diffusion probability model;

[0080] The sampling module quickly samples the noise image through the diffusion probability model solver according to the optimal mean and optimal variance to obtain the restored image;

[0081] The output module is used to output the restored image.

[0082] The above-mentioned terahertz image restoration system based on the diffusion probability model first pre-processes and adds noise to the input terahertz image to obtain a noise map, and then estimates the optimal mean and optimal variance of the noise map in the reverse diffusion process in real time through the pre-trained diffusion probability model, and finally generates a restored image by fast sampling. Traditional methods (such as filtering and super-resolution reconstruction technology) often perform poorly in dealing with noise and artifacts, resulting in impaired image quality, while the restoration method based on the diffusion probability model can better capture and remove these interferences by accurately estimating the noise characteristics, thereby improving the clarity and contrast of the image; and by introducing the optimal variance estimation and fast sampling strategy, it can generate higher quality images while reducing the sampling steps, achieving better detail preservation and texture optimization, which greatly improves the quality and efficiency of terahertz image restoration.

[0083] In one embodiment, a terahertz image restoration system based on a diffusion probability model is provided, further comprising: a detail preservation module for extracting prior information of the restored image;

[0084] The texture optimization module is used to refine the texture of the restored image according to the prior information using the unconditional diffusion model and the truncated sampling strategy to obtain the detail enhanced image; the output module outputs the detail enhanced image.

[0085] In one embodiment, a terahertz image restoration system based on a diffusion probability model is provided, further comprising: a multimodal information fusion module for adaptively fusing multimodal feature information of a detail enhanced image to obtain an optimized restored image; and an output module for outputting the optimized restored image.

[0086] For the specific definition of the terahertz image restoration system based on the diffusion probability model, please refer to the definition of the terahertz image restoration method based on the diffusion probability model in the above text, which will not be repeated here. Each module in the above-mentioned terahertz image restoration system based on the diffusion probability model can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0087] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0088] A terahertz image is input, and preprocessed and noise-added are performed on the terahertz image to obtain a noise map;

[0089] Based on the pre-trained diffusion probability model, the optimal mean and optimal variance of the noise map in the reverse diffusion process are estimated in real time through the score function; the score function is one of the outputs of the pre-trained diffusion probability model;

[0090] According to the optimal mean and optimal variance, the noise map is quickly sampled through the diffusion probability model solver to obtain the restored image.

[0091] It can be understood that in addition to the memory and processor involved above, the above-mentioned computer device also includes other software and hardware components not listed in this specification. The specific components can be determined according to the model of the specific computer device in different application scenarios, and this specification will not list them one by one in detail.

[0092] In one embodiment, when the processor executes the computer program, the processor may also implement the steps or sub-steps added in each embodiment of the above-mentioned terahertz image restoration method based on the diffusion probability model.

[0093] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A terahertz image restoration method based on a diffusion probability model, characterized in that: Includes steps: Inputting a terahertz image, and preprocessing and adding noise to the terahertz image to obtain a noise map; Based on the pre-trained diffusion probability model, an optimal mean and an optimal variance of the noise map in the reverse diffusion process are estimated in real time through a scoring function; the scoring function is one of the outputs of the pre-trained diffusion probability model; According to the optimal mean and the optimal variance, the noise map is quickly sampled through a diffusion probability model solver to obtain a restored image.

2. The terahertz image restoration method based on the diffusion probability model according to claim 1, characterized in that: Also includes the steps: Prior information of the restored image is extracted, and texture refinement of the restored image is performed according to the prior information using an unconditional diffusion model and a truncated sampling strategy to obtain a detail-enhanced image.

3. The terahertz image restoration method based on the diffusion probability model according to claim 2, characterized in that: Also includes the steps: Adaptively fusion processing is performed on the multimodal feature information of the detail enhanced image to obtain an optimized restored image.

4. The terahertz image restoration method based on the diffusion probability model according to any one of claims 1 to 3, characterized in that: The optimal mean is: in, represents the score function, x n Representing the noise graph at step n: α n ＝1-β n Among them, ∈ n Indicates that Gaussian noise conforms to the normal distribution N(0,1), β is a random value, and α is a variation of β; The optimal variance is: in, represents the variance parameter, represents the expectation of the noise distribution, and d represents the dimension of the noise map.

5. The terahertz image restoration method based on the diffusion probability model according to claim 4, characterized in that: The sampling path of the fast sampling is: t represents the time variable, sampling from T to 0; ∈ θ (x t ,t) represents the noise prediction function, x t represents the noise graph at the current time t; α t represents the random noise β at the current time t t Variant of ; σ t Represents the optimal standard deviation at the current time t, according to the optimal variance get;h i Indicates time interval Represents the time difference between two adjacent steps in the sampling process; represents a higher-order error term.

6. A terahertz image restoration system based on a diffusion probability model, characterized in that: include: An input module, used for inputting a terahertz image; A processing module, used for preprocessing and adding noise to the terahertz image to obtain a noise map; An estimation module, for estimating in real time an optimal mean and an optimal variance of the noise map in a reverse diffusion process through a score function based on a pre-trained diffusion probability model; the score function is one of the outputs of the pre-trained diffusion probability model; A sampling module, which quickly samples the noise image through a diffusion probability model solver according to the optimal mean and the optimal variance to obtain a restored image; An output module is used to output the restored image.

7. The terahertz image restoration method based on the diffusion probability model according to claim 6, characterized in that: Also includes: A detail preservation module, used for extracting prior information of the restored image; The texture optimization module is used to refine the texture of the restored image according to the prior information using an unconditional diffusion model and a truncated sampling strategy to obtain a detail-enhanced image; the output module outputs the detail-enhanced image.

8. The terahertz image restoration method based on the diffusion probability model according to claim 7, characterized in that: Also includes: A multimodal information fusion module, used to perform adaptive fusion processing on the multimodal feature information of the detail enhanced image to obtain an optimized restored image; The output module outputs the optimized restored image.

9. A computer device comprising a memory and a processor, characterized in that: The memory stores a computer program, and when the processor executes the computer program, the steps of the terahertz image restoration method based on the diffusion probability model described in any one of claims 1 to 5 are implemented.

Images