Scattering Imaging Image Reconstruction Method Based on Biscore Matching Model in Wavelet Domain

Abstract

The present invention discloses a method for scattering imaging image reconstruction based on a wavelet domain dual-score matching model. This method utilizes trained full-frequency and high-frequency prior information of the target object using a full-frequency and high-frequency prior information. Under the dual constraints of the full-frequency and high-frequency prior information, the speckle image of the target object behind the scattering medium acquired by the image sensor is processed using Wiener deconvolution, support domain constraints, and bilateral filtering as a fidelity term for image reconstruction. The reconstruction process includes inputting noise, prediction, correction, sub-image segmentation, updating sub-images, merging sub-images, inverse wavelet transform, and denoising. The denoised image is used as input, and the reconstruction process is repeated a predicted number of times before outputting a reconstructed image. The present invention reconstructs the target object behind the scattering medium through the efficient synergy of high-dimensional, low-rank prior information constraints and data fidelity. This method effectively addresses the difficult problem of reconstructing complex targets in scattering medium scenarios and significantly improves imaging quality.

Description

Technical Field

[0001] The present invention relates to the field of electronic information technology, and in particular to a scattering imaging image reconstruction method based on a wavelet domain bi-score matching model. Background Art

[0002] Imaging through scattering media plays a crucial role in numerous key areas, including medical imaging, deep-sea exploration, and national defense security. In these complex applications, light propagation is often disrupted by scattering media (such as fog, smoke, and biological tissue), causing random refraction. This phenomenon not only makes direct imaging extremely difficult but also severely impacts image quality, significantly limiting the performance of imaging systems.

[0003] Existing scattering imaging solutions mainly include pure hardware solutions, algorithm-based solutions, and diffusion model-based solutions.

[0004] Pure hardware solution: Digital optical phase conjugation technology uses multiple digital components such as electro-optic modulators and spatial light modulators to generate phase conjugate light, thereby achieving focusing and imaging through scattering media.

[0005] Algorithmic solutions: Traditional image reconstruction methods, such as traditional deconvolution, mainly use the point spread function of the measurement system to deconvolve the observed image.

[0006] Diffusion model-based solution: The observation image is processed using the diffusion model.

[0007] In the process of implementing the technical solution of the present invention, the inventor of this patent discovered at least the following technical problems in the prior art:

[0008] (1) Pure hardware solution: The addition of multiple digital components in digital optical phase conjugation technology often makes the imaging system extremely complex, limiting the scope of practical application.

[0009] (2) Algorithmic solutions: Traditional image reconstruction methods are mainly used to reconstruct simple objects. The lack of prior information constraints often leads to the introduction of artifacts and noise in the reconstructed images.

[0010] (3) Diffusion model-based solutions: Although diffusion models have greatly improved image processing, they cannot effectively recover complex targets behind scattering media. This type of solution still does not solve the problem of how to combine physical models and mine prior information more efficiently.

[0011] In summary, most existing scattering imaging solutions focus on optimizing the imaging effects of simple targets, but there is still a large gap between them and the imaging requirements of complex targets in actual scenes, and they cannot meet actual imaging needs. Summary of the Invention

[0012] The present invention provides a scattering imaging image reconstruction method based on a wavelet domain dual-fractional matching model. The trained full-frequency fractional matching sub-model and high-frequency fractional matching sub-model can extract high-dimensional, low-rank prior information of the target object, including full-frequency and high-frequency prior information, and reconstruct the target object behind the scattering medium through the efficient collaboration of high-dimensional, low-rank prior information constraints and data fidelity. It can effectively penetrate the scattering medium and obtain a clear image, solving the problem that most existing scattering imaging solutions focus on optimizing the imaging effects of simple targets, resulting in the inability to meet the imaging requirements of complex targets in actual scenes, and significantly improving the imaging quality.

[0013] The present invention provides a scattering imaging image reconstruction method based on a wavelet domain bi-score matching model, the method comprising:

[0014] In the prior information extraction stage, the trained full-frequency score matching sub-model and high-frequency score matching sub-model are used to extract the full-frequency and high-frequency prior information of the target object;

[0015] In the image reconstruction stage, under the dual constraints of full-frequency and high-frequency prior information, the speckle image of the target object behind the scattering medium acquired by the image sensor is processed as the fidelity term after Wiener deconvolution, support region constraint and bilateral filtering. The reconstruction process includes:

[0016] In the first step, four two-dimensional Gaussian noise images are randomly generated;

[0017] In the second step, the Gaussian noise image is input into the full-frequency fractional matching sub-model for image reconstruction;

[0018] The third step is to introduce the constraint of full-frequency prior information and use the fidelity term to enhance data consistency to obtain the predicted image;

[0019] The fourth step is to use the annealed Langevin equation as a corrector and the fidelity term as a data consistency maintainer to input the predicted image into the inner loop for correction to obtain the image to be reconstructed;

[0020] The fifth step is to divide the image to be reconstructed into the original image of the high-frequency component and the original image of the low-frequency component;

[0021] In the sixth step, the high-frequency component original image is input into the high-frequency score matching sub-model, and the high-frequency prior information is used for constraint, while the fidelity term is used as a data consistency maintainer to obtain the high-frequency component updated image;

[0022] In the seventh step, the high-frequency component updated image and the high-frequency component original image are weighted and summed, and then merged and optimized with the low-frequency component original image. The optimized wavelet domain image is subjected to inverse wavelet transform and total variation regularization term denoising to obtain the denoised spatial domain image.

[0023] In the eighth step, the denoised spatial domain image is processed by wavelet and used as input again. Steps 2 to 7 are repeated for a preset number of times and then the loop ends. The spatial domain image output after the last loop is used as a reconstructed image close to the target object.

[0024] One or more technical solutions provided by the present invention have at least the following technical effects or advantages:

[0025] This method precisely captures high-dimensional, low-rank prior information (including full-frequency and high-frequency prior information) of the target as a constraint. The speckle image of the target behind the scattering medium, acquired by the image sensor, is processed using Wiener deconvolution, support region constraints, and bilateral filtering as the fidelity term. This method effectively synergizes the high-dimensional, low-rank prior information constraint with data fidelity to reconstruct the target behind the scattering medium. During the image reconstruction process, a series of collaborative optimization steps ensure that the reconstructed image achieves an optimal balance between the high-dimensional, low-rank prior information constraint and data fidelity. This method effectively solves the challenge of reconstructing complex targets in scattering medium scenarios and significantly improves imaging quality. Compared with existing hardware-only solutions, this method eliminates the need for multiple digital components and only uses an image sensor as a linear scattering imaging system, resulting in a simple imaging system with a wide range of applications. Compared with existing algorithmic solutions and diffusion model-based solutions, this method significantly reduces artifacts, produces clearer and richer images, and significantly improves overall quality when reconstructing complex targets behind scattering media. The present invention not only meets the high-precision requirements in complex imaging environments, but also effectively bridges the gap between existing technologies and practical applications, providing reliable technical support for the efficient reconstruction of complex targets.

[0026] The present invention has achieved a significant breakthrough in the reconstruction of complex targets in scattering media. Experimental results show that compared with the images obtained by the existing diffusion model scheme, the peak signal-to-noise ratio of the reconstructed image finally output by the present invention is improved by nearly 2.67dB, and the structural similarity value reaches 0.7136, significantly improving the overall image quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 2. Schematic diagram of the optical path of a linear scattering imaging system according to an embodiment of the present invention;

[0028] Figure 2 Flowchart of a scattering imaging image reconstruction method based on a wavelet domain bi-score matching model in one embodiment of the present invention;

[0029] Figure 3 Result diagrams of simulation experiments using three different methods in one embodiment of the present invention;

[0030] In the figure, sub-model 1 represents the full-frequency score matching sub-model, sub-model 2 represents the high-frequency score matching sub-model, P represents the predictor, C represents the corrector, DC represents the data consistency maintainer, and TV represents the total variation regularization term. DETAILED DESCRIPTION

[0031] The present invention provides a scattering imaging image reconstruction method based on a wavelet domain dual-fractional matching model. The trained full-frequency fractional matching sub-model and high-frequency fractional matching sub-model can extract high-dimensional, low-rank prior information of the target object, including full-frequency and high-frequency prior information, and reconstruct the target object behind the scattering medium through the efficient collaboration of high-dimensional, low-rank prior information constraints and data fidelity. It can effectively penetrate the scattering medium and obtain a clear image, solving the problem that most existing scattering imaging solutions focus on optimizing the imaging effects of simple targets, resulting in the inability to meet the imaging requirements of complex targets in actual scenes, and significantly improving the imaging quality.

[0032] First, technical terms used in the present invention are explained.

[0033] The wavelet domain dual-score matching model is designed to address the problem of insufficient prior information extraction from single-channel spatial domain data. This invention proposes a prior learning model for the field of scatter imaging. Through wavelet transform, the original single-channel training data is decomposed into four-channel full-frequency transform domain data and three-channel high-frequency transform domain data. On this basis, the score matching model is trained on the four-channel full-frequency transform domain data and the three-channel high-frequency transform domain data, respectively, ultimately obtaining a four-channel wavelet domain full-frequency score matching sub-model (referred to as the full-frequency score matching sub-model) and a three-channel wavelet domain high-frequency score matching sub-model (referred to as the high-frequency score matching sub-model). The wavelet domain dual-score matching model can efficiently learn the high-dimensional, low-rank prior information of the target object, thereby extracting high-dimensional, low-rank prior information from complex data and accurately capturing its global and local features.

[0034] Prior information extraction is to learn the physical properties of images such as edges, textures, and shapes from a large number of image data sets. In the present invention, the proposed wavelet domain dual-score matching model is used to extract high-dimensional, low-rank prior information of the target object, provide accurate and effective regularization constraints for the image reconstruction process, effectively suppress the overfitting phenomenon of the model, and significantly improve the accuracy and stability of image reconstruction. Among them, the full-frequency prior information is a kind of high-dimensional, low-rank prior information extracted from the target object by the full-frequency score matching sub-model. The present invention uses the full-frequency prior information to constrain the generation of the image to be reconstructed. High-frequency prior information is a kind of high-dimensional, low-rank prior information extracted from the target object by the high-frequency score matching sub-model. The present invention uses high-frequency prior information to enhance the details and edge information of the image. Through the dual constraints of full-frequency prior information and high-frequency prior information, high-quality reconstruction of complex targets is achieved.

[0035] Next, the linear scattering imaging system involved in the present invention is introduced.

[0036] Scatter imaging is an imaging technology that uses speckle information to restore the image of objects behind the scattering medium. Compared with the traditional pure hardware solution that requires the use of multiple digital components such as electro-optical modulators and spatial light modulators, this invention builds a linear scatter imaging system that only requires an image sensor. Its optical path principle is as follows: Figure 1 As shown in the figure, the light emitted by the LED passes through the collimating lens and illuminates the target object. When the target light field passes through the scattering medium (such as fog, smoke, biological tissue, etc.), the light will be scattered multiple times, and the image sensor will capture the speckle image. It can be regarded as the point spread function of the linear scattering imaging system With the target Convolution, that is ,in Denotes the convolution operator. By measuring the system's point spread function and speckle pattern using an image sensor, deconvolution can be used to reconstruct the scattered image. However, unconstrained reconstructed images are subject to numerous artifacts and low clarity. Therefore, the present invention requires efficient collaboration between high-dimensional, low-rank prior information constraints and data fidelity to reconstruct the target behind the scattering medium and obtain a clear image.

[0037] To better understand the present invention's scatter imaging image reconstruction method based on a wavelet-domain bifractional matching model, the following detailed description is provided in conjunction with the accompanying drawings and specific implementation methods. Obviously, the embodiments described herein are only a subset of the embodiments, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments herein without inventive effort are also within the scope of protection of the present invention.

[0038] The scattering imaging image reconstruction method based on the wavelet domain bi-score matching model of the present invention mainly includes three stages: a model training stage, a priori information extraction stage and an image reconstruction stage.

[0039] (1) Model training stage;

[0040] In the model training stage, the single-channel spatial domain image dataset of the target object is first converted into a full-frequency image dataset in the four-channel wavelet domain and a high-frequency image dataset in the three-channel wavelet domain using wavelet transform. Then, the two datasets obtained after wavelet transform are used to train the first original score matching model and the second original score matching model respectively to obtain the trained full-frequency score matching sub-model and high-frequency score matching sub-model.

[0041] It should be noted that the wavelet basis function used in the wavelet transform process in this embodiment is the Haar function. In practical applications, there is no restriction on the selected wavelet basis function.

[0042] Among them, the construction process of the four-channel wavelet domain full-frequency image dataset and the three-channel wavelet domain high-frequency image dataset, in the specific construction process, for example: select the LSUN-Church dataset as the target object original dataset, crop and adjust each selected image in the original dataset to 56×56 pixels, then use the adjusted image as the center and evenly fill the black background area around it, and finally generate a single-channel spatial domain image with a size of 256×256 pixels to complete the preprocessing of the original dataset. Next, select 10,000 images from the preprocessed dataset and use wavelet transform to convert these 10,000 single-channel spatial domain images into a four-channel wavelet domain full-frequency image dataset. and high-frequency image datasets in the three-channel wavelet domain ,in, 、 、 、 Represents single-channel spatial domain image datasets The low-frequency-low-frequency subband image data subset, the low-frequency-high-frequency subband image data subset, the high-frequency-low-frequency subband image data subset, and the high-frequency-high-frequency subband image data subset are obtained after wavelet transform.

[0043] After constructing the full-frequency image dataset in the four-channel wavelet domain and high-frequency image datasets in the three-channel wavelet domain After that, the training process begins, and the training process includes steps S11 to S13.

[0044] Step S11: The full-frequency image dataset in the four-channel wavelet domain The first original score matching model is input for training to learn the full-frequency prior information of the overall structure of the image and obtain the full-frequency score matching sub-model.

[0045] The process of obtaining the full-frequency score matching sub-model includes steps S111 to S112.

[0046] Step S111: Configure the stochastic differential equation for variance explosion used in the model training phase. The expression is as follows:

[0047] ,

[0048] Where, represents a full-frequency image dataset in the four-channel wavelet domain, represents a full-frequency monotonically increasing function, represents the full-frequency diffusion coefficient, is the uniform sampling time, Indicates Brownian motion in dimensional real space, represents a real number, Represents the vector dimension.

[0049] Stochastic differential equations with variance explosion can effectively disperse the data distribution into a wide range of noise levels, thereby more effectively learning the underlying data structure and improving the quality of generated samples.

[0050] Step S112: Through training, use the full-frequency continuous time correlation score function To estimate the gradient of the full-frequency logarithmic data distribution Taking into account the true value of Still unknown, choose to use replace .

[0051] The full-frequency continuous time correlation fraction function will be used To estimate the gradient of the replaced full-frequency logarithmic data distribution This estimation process is modeled as solving the following full-frequency kernel objective function in a fractional-based stochastic differential equation environment:

[0052] ,

[0053] Where, represents the optimal parameters for full-frequency neural network training, represents the full-frequency neural network training parameters, Express expectations, represents a positive weight function, Represented as full-frequency image training samples in four-channel wavelet domain, Therefore The full-frequency Gaussian perturbation kernel centered at represents the full-frequency continuous time correlation fractional function, for The full-frequency logarithmic data distribution gradient.

[0054] Step S12: The high-frequency image dataset in the three-channel wavelet domain The second original score matching model is input for training to learn the high-frequency prior information of local details and texture features of the image and obtain a high-frequency score matching sub-model.

[0055] The process of obtaining the high-frequency score matching sub-model includes steps S121 and S122.

[0056] Step S121: Configure the stochastic differential equation for variance explosion used in the model training phase. The expression is as follows:

[0057] ,

[0058] Where, represents a high-frequency image dataset in the three-channel wavelet domain, represents a high-frequency monotonically increasing function, represents the high-frequency diffusion coefficient, is the uniform sampling time, Indicates Brownian motion in dimensional real space, represents a real number, Represents the vector dimension.

[0059] Stochastic differential equations with variance explosion can effectively disperse the data distribution into a wide range of noise levels, thereby more effectively learning the underlying data structure and improving the quality of generated samples.

[0060] Step S122: Through training, use the high-frequency continuous time correlation score function To estimate the gradient of the high-frequency logarithmic data distribution Taking into account the true value of Still unknown, choose to use replace .

[0061] A high frequency continuous time correlation score function will be used To estimate the gradient of the replaced high-frequency logarithmic data distribution This estimation process is modeled as solving the following high-frequency core objective function in a fractional-based stochastic differential equation environment:

[0062] ,

[0063] Where, represents the optimal parameters for high-frequency neural network training, represents the high-frequency neural network training parameters, Express expectations, represents a positive weight function, Represented as high-frequency image training samples in the three-channel wavelet domain, Therefore The high-frequency Gaussian perturbation kernel centered at represents the high-frequency continuous-time correlation fractional function, yes The high-frequency logarithmic data distribution gradient.

[0064] Step S13: Configure training parameters. First, add Gaussian noise to perturb the data distribution. For example, add 2000 Gaussian noises with noise values ​​between 0.01 and 384. Next, optimize using the Adam optimizer. For example, set the learning rate to 0.0002. Finally, set the number of iterations. For example, set it to 500,000.

[0065] (2) Prior information extraction stage;

[0066] like Figure 2 As shown, in the prior information extraction stage, the full-frequency score matching sub-model trained in the model training stage (i.e. Figure 2 sub-model 1) and the high-frequency score matching sub-model (i.e. Figure 2 Sub-model 2) in

[15] is used to extract the full-frequency and high-frequency prior information of the target object.

[0067] (3) Image reconstruction stage;

[0068] like Figure 2 As shown in the figure, in the image reconstruction stage, under the dual constraints of full-frequency and high-frequency prior information, the speckle image of the target behind the scattering medium acquired by the image sensor is converted into After Wiener deconvolution, support region constraint and bilateral filtering, it is used as the fidelity item , and perform image reconstruction.

[0069] Among them, in order to further improve the underlying physical consistency of the reconstructed image and take into account the area of ​​interest in the actual imaging field The existence of the speckle image does not directly As a fidelity item, the acquired speckle image needs to be After processing with Wiener deconvolution, support region constraint and bilateral filtering, it is used as the fidelity item . Fidelity Acts as a data consistency maintainer throughout the reconstruction process (i.e. Figure 2 DC in the collected speckle image. The processing specifically includes steps S21 to S22.

[0070] Step S21: Use Wiener deconvolution to deconvolve the speckle image After processing, the deconvolution image is obtained , which is expressed as follows:

[0071] ,

[0072] Where, is the image after deconvolution, represents the Fourier transform, is the speckle image, is the point spread function in the Fourier domain, yes The conjugate of represents an adjustable parameter of noise;

[0073] Step S22: Deconvolute the image Perform support domain constraint and bilateral filtering to obtain the fidelity term , the expression is as follows:

[0074] ,

[0075] Where, Represents the image after deconvolution A pixel position on Represents the image after deconvolution Another pixel position on Represents fidelity item Medium pixel The pixel value at Represents the image after deconvolution Medium pixel The pixel value at Represents the image after deconvolution Medium pixel The pixel value at is a normalization factor used to ensure that the weights sum to 1, It's a pixel Neighborhood, is a Gaussian function in the spatial domain, is the standard deviation in the spatial domain, is a Gaussian function in the intensity domain, is the standard deviation of the intensity domain, is the region of interest in the actual imaging field.

[0076] The reconstruction process includes steps one to eight.

[0077] In the first step, four two-dimensional Gaussian noise images are randomly generated.

[0078] In the second step, the Gaussian noise image is input into the full-frequency fractional matching sub-model for image reconstruction.

[0079] Since the full-frequency score matching sub-model uses a stochastic differential equation with variance explosion in the training phase, the second step is to use the obtained full-frequency continuous time correlation score function Solve the stochastic differential equation for the inverse time variance explosion of the full-frequency fractionally matched submodel:

[0080] ,

[0081] Where, represents the full-frequency image in the wavelet domain at the image reconstruction stage, represents a full-frequency monotonically increasing function, is the uniform sampling time, represents the full-frequency logarithmic data distribution gradient, represents the inverse Brownian motion, represents a real number, Represents the vector dimension.

[0082] The third step is to introduce the constraint of full-frequency prior information and use the fidelity term to enhance data consistency to obtain the first discrete step lengths of the predicted image .

[0083] In order to ensure the reconstruction quality of complex targets behind the scattering medium, the constraint of full-frequency prior information is introduced to avoid large errors in the evolution of the stochastic differential equation with variance explosion in the discretization direction. Since the full-frequency fractional matching sub-model uses the stochastic differential equation with variance explosion in the training phase, the target image estimation process at each time step is constrained by introducing full-frequency prior information when solving the stochastic differential equation with inverse time variance explosion. The constraint of full-frequency prior information is introduced in the third step to constrain the predictor (i.e. Figure 2 The prediction process of P in sub-model 1 is constrained, and the fidelity term is used to enhance data consistency. Its expression is as follows:

[0084] ,

[0085] Where, Indicates the discrete step length prediction image, is a factor that balances the data consistency term and the regularization term, represents the fidelity term, represents the wavelet transform, The total time step is External index of Indicates the The full-frequency image in the wavelet domain with discrete steps, Indicates the The noise intensity at discrete steps, express The square of Indicates the The noise intensity at discrete steps, express The square of To convert the full-frequency continuous time correlation fraction function The function obtained after discretization is is Gaussian noise.

[0086] The fourth step is to use the annealed Langevin equation as a corrector (i.e. Figure 2 In the sub-model 1 of C), the fidelity term is used as the data consistency maintainer. The predicted image with discrete steps is input into the inner loop for correction to obtain the discrete steps of the image to be reconstructed .

[0087] Among them, the annealed Langevin equation is used as the corrector, and the fidelity term is used as the data consistency maintainer. The predicted image is input into the inner loop for correction. Its expression is as follows:

[0088] ,

[0089] Where, Indicates As the inner loop input The full-frequency image in the wavelet domain with discrete steps is represents the factor that balances the data consistency term and the regularization term, represents the wavelet transform, represents the fidelity term, Indicates As the inner loop input The full-frequency image in the wavelet domain with discrete steps is The total step length is The internal index of Indicates the The full frequency step length when the discrete step length is To convert the full-frequency continuous time correlation fraction function The function obtained after discretization is is Gaussian noise.

[0090] When the inner loop is finished, the discrete steps of the image to be reconstructed .

[0091] Step 5: discrete steps of the image to be reconstructed Divided into The high-frequency component of the original image when the discrete step length Hedi The original image of the low-frequency component at discrete steps .

[0092] Step 6: The high-frequency components of the original image with discrete steps Input to the high-frequency score matching sub-model (i.e. Figure 2 2 in the sub-model), and use high-frequency prior information for constraints, and use the fidelity term as a data consistency maintainer to obtain the first The high-frequency components of the discrete steps update the image .

[0093] By using high-frequency prior information to constrain the detail optimization process, that is, by using high-frequency prior information to constrain the stochastic differential equation for solving the inverse time variance explosion in the prediction and correction stage, the complex details and edge information of the image can be fully reconstructed while maintaining the basic characteristics of the reconstruction result. In the sixth step, the obtained high-frequency continuous time correlation score function is used When solving the stochastic differential equation of the inverse time variance explosion of the high-frequency score matching model, the high-frequency component original image is input into the high-frequency score matching sub-model (i.e. Figure 2 2 in the sub-model), and use the high-frequency prior information to match the predictors in the high-frequency score sub-model (i.e. Figure 2 P in sub-model 2) and the corrector (i.e. Figure 2 The prediction correction process of C) in sub-model 2 is constrained, and the fidelity term is used as a data consistency maintainer to maintain fidelity. Its expression is as follows:

[0094] ,

[0095] Where, Indicates the The high-frequency components of the wavelet domain with discrete steps are used to update the image. Indicates As the inner loop input The high-frequency component of the wavelet domain updates the image with discrete steps. 、 denote the factors that balance the data consistency term and the regularization term, Represents fidelity item High-frequency image after wavelet transform , 、 、 Fidelity terms After wavelet transform, we get low-frequency-high-frequency sub-band image, high-frequency-low-frequency sub-band image, and high-frequency-high-frequency sub-band image. For the The high-frequency component of the original image at discrete steps is Indicates the A long walk away, Indicates the The noise intensity at discrete steps, express The square of Indicates the The noise intensity at discrete steps, express The square of Indicates As the inner loop input The image is updated by the high-frequency components in the wavelet domain with discrete steps.

[0096] Step 7: The high-frequency components of the discrete steps update the image Hedi The high-frequency components of the original image with discrete steps After weighted summation, The original image has a low-frequency component with discrete steps The merge optimization is performed, and the optimized wavelet domain image is subjected to inverse wavelet transform and total variation regularization term denoising to obtain the denoised spatial domain image.

[0097] In the seventh step, the optimized wavelet domain image is inversely transformed to obtain the The spatial domain image with discrete steps. And using the total variation regularization term ( Figure 2 TV in the denoising) to achieve further denoising, the denoised discrete steps of the spatial domain image The expression is as follows:

[0098] ,

[0099] Where, Represents the first The spatial domain image with discrete steps, represents the inverse wavelet transform, Indicates the The high-frequency components of the wavelet domain with discrete steps are used to update the image. Indicates the The high-frequency component of the original image with discrete steps, Indicates the The original image with low-frequency components of discrete steps, is the weight factor.

[0100] In the eighth step, the denoised spatial domain image is processed by wavelet and used as input again. The second to seventh steps are repeated for a preset number of times and the loop ends. The spatial domain image output after the last loop is used as the reconstructed image close to the target object. In a specific implementation, for example, when the preset number of times is 1000, the eighth step outputs a high-definition reconstructed image with minimal artifacts.

[0101] To evaluate the performance of the scatter imaging image reconstruction method based on the wavelet-domain bifractional matching model, 50 images were randomly selected from the preprocessed dataset as a test set. To simulate actual speckle images, each image in the test set was convolved with the measured value of the point spread function to generate a speckle pattern. A traditional deconvolution algorithm and a diffusion model algorithm were used as comparison algorithms, with peak signal-to-noise ratio and structural similarity selected as evaluation metrics. The comparison results are shown in Table 1.

[0102] Table 1 Comparison of peak signal-to-noise ratio and structural similarity

[0103]

[0104] Image reconstruction effect is as follows Figure 3 shown. Figure 3 (a) in the figure is 6 real images. Figure 3 (b) in the figure is the speckle image collected after each real image passes through the scattering medium. Figure 3 (c), (d) and (e) are the reconstructed images output by the traditional deconvolution algorithm, the diffusion model algorithm and the present invention for each speckle image. In order to better demonstrate the quality of the reconstructed image, only the region of interest of 56×56 pixels is displayed. Figure 3 As can be seen, for complex objects, the reconstructed images output by both the traditional deconvolution algorithm and the diffusion model algorithm exhibit obvious horizontal and vertical streaks. In contrast, the reconstructed images output by the present invention display fewer artifacts and retain clearer details. Test results show that the reconstructed image quality of the present invention is superior to the two existing methods.

[0105] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A scattering imaging image reconstruction method based on a wavelet domain bi-score matching model, characterized in that: The method comprises: In the prior information extraction stage, the trained full-frequency score matching sub-model and high-frequency score matching sub-model are used to extract the full-frequency and high-frequency prior information of the target object; In the image reconstruction stage, under the dual constraints of full-frequency and high-frequency prior information, the speckle image of the target object behind the scattering medium acquired by the image sensor is processed as the fidelity term after Wiener deconvolution, support region constraint and bilateral filtering. The reconstruction process includes: In the first step, four two-dimensional Gaussian noise images are randomly generated; In the second step, the Gaussian noise image is input into the full-frequency fractional matching sub-model for image reconstruction; The third step is to introduce the constraint of full-frequency prior information and use the fidelity term to enhance data consistency to obtain the predicted image; The fourth step is to use the annealed Langevin equation as a corrector and the fidelity term as a data consistency maintainer to input the predicted image into the inner loop for correction to obtain the image to be reconstructed; The fifth step is to divide the image to be reconstructed into the original image of the high-frequency component and the original image of the low-frequency component; In the sixth step, the high-frequency component original image is input into the high-frequency score matching sub-model, and the high-frequency prior information is used for constraint, while the fidelity term is used as a data consistency maintainer to obtain the high-frequency component updated image; In the seventh step, the high-frequency component updated image and the high-frequency component original image are weighted and summed, and then merged and optimized with the low-frequency component original image. The optimized wavelet domain image is subjected to inverse wavelet transform and total variation regularization term denoising to obtain the denoised spatial domain image. In the eighth step, the denoised spatial domain image is processed by wavelet and used as input again. Steps 2 to 7 are repeated for a preset number of times and then the loop ends. The spatial domain image output after the last loop is used as the reconstructed image close to the target object. Before the prior information extraction stage, the model training stage is also included, including: The single-channel spatial domain image dataset of the target object is converted into a full-frequency image dataset in a four-channel wavelet domain and a high-frequency image dataset in a three-channel wavelet domain using wavelet transform; The two data sets obtained after wavelet transformation are used to train the first raw score matching model and the second raw score matching model respectively, so as to obtain the trained full-frequency score matching sub-model and high-frequency score matching sub-model.

2. The method according to claim 1, wherein The full-frequency score matching sub-model is obtained by: Configure the stochastic differential equation for variance explosion used in the model training phase. The expression is as follows: ， Where, represents a full-frequency image dataset in the four-channel wavelet domain, represents a full-frequency monotonically increasing function, represents the full-frequency diffusion coefficient, is the uniform sampling time, Indicates Brownian motion in dimensional real space, represents a real number, Represents vector dimension; Through training, the gradient of the full-frequency logarithmic data distribution is estimated using the full-frequency continuous time correlation score function. The estimation process is modeled as solving the following full-frequency core objective function in a fractional-based stochastic differential equation environment: ， Where, represents the optimal parameters for full-frequency neural network training, represents the full-frequency neural network training parameters, Express expectations, represents a positive weight function, Represented as full-frequency image training samples in four-channel wavelet domain, Therefore The full-frequency Gaussian perturbation kernel centered at represents the full-frequency continuous time correlation fractional function, for The full-frequency logarithmic data distribution gradient.

3. The method according to claim 1, wherein The high-frequency score matching sub-model is obtained by: Configure the stochastic differential equation for variance explosion used in the model training phase. The expression is as follows: ， Where, represents a high-frequency image dataset in the three-channel wavelet domain, represents a high-frequency monotonically increasing function, represents the high-frequency diffusion coefficient, is the uniform sampling time, Indicates Brownian motion in dimensional real space, represents a real number, Represents vector dimension; Through training, a high-frequency continuous-time correlation score function is used to estimate the gradient of the high-frequency logarithmic data distribution. The estimation process is modeled as solving the following high-frequency core objective function in a fractional-based stochastic differential equation environment: ， Where, represents the optimal parameters for high-frequency neural network training, represents the high-frequency neural network training parameters, Express expectations, represents a positive weight function, Represented as high-frequency image training samples in the three-channel wavelet domain, Therefore The high-frequency Gaussian perturbation kernel centered at represents the high-frequency continuous-time correlation fractional function, yes The high-frequency logarithmic data distribution gradient.

4. The method according to claim 1, wherein The speckle image is processed by Wiener deconvolution, support region constraint and bilateral filtering as a fidelity item, specifically including: After processing the speckle image using Wiener deconvolution, the deconvolved image is obtained, which is expressed as follows: ， Where, is the image after deconvolution, represents the Fourier transform, is the speckle image, is the point spread function in the Fourier domain, yes The conjugate of represents an adjustable parameter of noise; The deconvolved image is subjected to support region constraint and bilateral filtering to obtain the fidelity term, which is expressed as follows: ， Where, Represents the image after deconvolution A pixel position on Represents the image after deconvolution Another pixel position on Represents fidelity item Medium pixel The pixel value at Represents the image after deconvolution Medium pixel The pixel value at Represents the image after deconvolution Medium pixel The pixel value at is a normalization factor used to ensure that the weights sum to 1, It's a pixel Neighborhood, is a Gaussian function in the spatial domain, is the standard deviation in the spatial domain, is a Gaussian function in the intensity domain, is the standard deviation of the intensity domain, is the region of interest in the actual imaging field.

5. The method according to claim 2, wherein The second step is specifically: using the obtained full-frequency continuous time correlation score function Solve the stochastic differential equation for the inverse time variance explosion of the full-frequency fractionally matched submodel: ， Where, represents the full-frequency image in the wavelet domain at the image reconstruction stage, represents a full-frequency monotonically increasing function, is the uniform sampling time, represents the full-frequency logarithmic data distribution gradient, represents the inverse Brownian motion, represents a real number, Represents the vector dimension.

6. The method according to claim 5, wherein When solving the stochastic differential equation with inverse time variance explosion, the third step introduces the constraint of full-frequency prior information and uses the fidelity term to enhance data consistency. The expression is as follows: ， Where, Indicates the discrete step length prediction image, is a factor that balances the data consistency term and the regularization term, represents the fidelity term, represents the wavelet transform, The total time step is External index of Indicates the The full-frequency image in the wavelet domain with discrete steps, Indicates the The noise intensity at discrete steps, express The square of Indicates the The noise intensity at discrete steps, express The square of To convert the full-frequency continuous time correlation fraction function The function obtained after discretization is is Gaussian noise.

7. The method according to claim 6, wherein In the fourth step, the annealed Langevin equation is used as a corrector, and the fidelity term is used as a data consistency maintainer. The predicted image is input into the inner loop for correction. The expression is as follows: ， Where, Indicates As the inner loop input The full-frequency image in the wavelet domain with discrete steps is represents the factor that balances the data consistency term and the regularization term, represents the wavelet transform, represents the fidelity term, Indicates As the inner loop input The full-frequency image in the wavelet domain with discrete steps is The total step length is The internal index of Indicates the The full frequency step length when the discrete step length is To convert the full-frequency continuous time correlation fraction function The function obtained after discretization is is Gaussian noise.

8. The method according to claim 7, wherein In the sixth step, when using the obtained high-frequency continuous time correlation score function When solving the stochastic differential equation of the inverse time variance explosion of the high-frequency score matching model, the high-frequency component original image is input into the high-frequency score matching sub-model, and the high-frequency prior information is used for constraint. At the same time, the fidelity term is used as a data consistency maintainer for fidelity. Its expression is as follows: ， Where, Indicates the The high-frequency components of the wavelet domain with discrete steps are used to update the image. Indicates As the inner loop input The high-frequency component of the wavelet domain updates the image with discrete steps. 、 denote the factors that balance the data consistency term and the regularization term, Represents fidelity item High-frequency image after wavelet transform , 、 、 Fidelity terms After wavelet transform, we get low-frequency-high-frequency sub-band image, high-frequency-low-frequency sub-band image, and high-frequency-high-frequency sub-band image. For the The high-frequency component of the original image at discrete steps is Indicates the A long walk away, Indicates the The noise intensity at discrete steps, express The square of Indicates the The noise intensity at discrete steps, express The square of Indicates As the inner loop input The image is updated by the high-frequency components in the wavelet domain with discrete steps.

9. The method according to claim 1, wherein In the seventh step, the high-frequency component updated image and the high-frequency component original image are weighted and summed, and then merged and optimized with the low-frequency component original image. The optimized wavelet domain image is subjected to inverse wavelet transform and total variation regularization term denoising, and its expression is as follows: ， Where, Represents the first The spatial domain image with discrete steps, represents the inverse wavelet transform, Indicates the The high-frequency components of the wavelet domain with discrete steps are used to update the image. Indicates the The high-frequency component of the original image with discrete steps, Indicates the The original image with low-frequency components of discrete steps, is the weight factor.

Images