A hyperspectral image denoising diffusion model construction method and system

Abstract

The application relates to a hyperspectral image denoising diffusion model construction method and system, and belongs to the technical field of computer vision. The method generates a noisy low-resolution hyperspectral image for testing by adding noise to a clean image; the number of compressed channels is dynamically determined by using an elbow point method, spectrum compression is performed, and a coefficient matrix is obtained; each band complex noise image after spectrum compression is converted into an approximate Gaussian noise image by using a noise conversion technology; the number of inverse diffusion embedding steps of each band after noise conversion is dynamically determined by using a pre-trained diffusion model, and a more accurate denoising process is performed through the determined number of inverse diffusion steps. The application can dynamically determine the number of spectrum compression channels of the noisy hyperspectral image, thereby effectively improving the recovery accuracy of the hyperspectral image, can adapt to various noisy hyperspectral image scenes, can dynamically determine the number of inverse diffusion embedding steps according to the noise intensity, can sufficiently denoise high-noise channels and moderately constrain low-noise channels, and can dynamically balance the denoising intensity and spectrum fidelity.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision, and particularly relates to a hyperspectral image denoising diffusion model construction method and system. BACKGROUND

[0002] Hyperspectral imaging is an imaging technology that finely samples in both spatial and spectral dimensions. Hyperspectral images (HSI) have shown significant application value and broad development potential in various fields such as agricultural practice, mineral exploration, environmental monitoring, and military applications due to their ability to capture high-dimensional spectral information. However, the data quality of HSI directly affects its analysis efficiency and application effect and significantly impacts the performance of subsequent tasks. Therefore, hyperspectral image denoising, as a core technology to ensure data integrity and improve overall performance, is very important.

[0003] Traditional model-driven denoising methods often construct denoising as an inverse problem solving framework from noisy images to pure images. It ensures consistency with the original data through a data fidelity term, and uses regularization techniques (such as sparsity, low-rank characteristics, tensor structure, etc. prior) to constrain the solution space. However, this method has limitations: it relies on fixed mathematical assumptions and is difficult to adapt to the dynamic complexity of hyperspectral images; it uses global static modeling and cannot effectively handle local heterogeneous structures in images.

[0004] To address the shortcomings of traditional model methods, deep learning technology has made breakthroughs in hyperspectral image denoising. According to the training method, it can be divided into supervised and unsupervised methods. Among them, the supervised method achieves denoising by constructing a noisy-pure image mapping model, which requires paired data for training. It learns feature relationships with the help of deep neural networks, denoises and preserves spectral information by minimizing loss, and although it has good performance, it has obvious limitations in practical applications due to its reliance on a large amount of paired labeled data, which is costly, scarce, and difficult to label.

[0005] In this context, unsupervised denoising methods provide a new direction for hyperspectral image denoising. They do not rely on paired ground truth and often use zero-shot deep priors or blind spot / mask reconstruction self-supervised strategies to learn and recover from noisy data. However, handcrafted priors have limited ability to characterize complex mixed noise, and self-supervised methods rely on noise independence assumptions, which can easily degrade when encountering strong structured noise. Therefore, diffusion models are introduced as learnable generative priors. In recent years, diffusion models have been a typical representative of unsupervised methods and have shown excellent performance in image denoising, super-resolution reconstruction, image inpainting, and other restoration tasks. Unconditional diffusion models relying on large-scale natural image pre-training effectively alleviate the problem of insufficient data samples.

[0006] However, the existing diffusion model is difficult to effectively handle the following three major problems in hyperspectral denoising: 1) The number of bands of the hyperspectral image is greatly different (from dozens to hundreds), and it is not reasonable to compress to a fixed number of channels (such as 3), and when facing a hyperspectral image containing hundreds of channels, a large amount of ground object and spectral effective information will be discarded, resulting in information loss and effect decline during recovery; 2) The complexity of noise sources, the noise of the hyperspectral image is often mixed, for example, Gaussian noise mixed with stripe noise. However, there is an assumption of Gaussian distribution in the diffusion model, which makes it not achieve the best effect when dealing with complex noise; 3) The heterogeneity of band noise, the type and intensity of noise in different bands are heterogeneous, and the same diffusion step will lead to incomplete denoising of high-noise channels and over-denoising of low-noise channels. In the modeling and training process, a large number of traditional methods and deep learning methods generally assume that the noise of the full spectrum of the hyperspectral image conforms to the same distribution, which is inconsistent with the actual situation, and is difficult to cope with the dual challenges of mixed noise and band noise heterogeneity. SUMMARY

[0007] The purpose of the present application is to overcome the problem that the existing hyperspectral image denoising method ignores the inconsistency of the noise intensity and type of each band in the real scene, and to provide a hyperspectral image denoising diffusion model construction method and system, which improves the reconstruction quality by introducing adaptive and dynamic determination of the number of compressed channels in spectral compression and introducing adaptive and dynamic determination of the number of inverse diffusion embedding steps in the inverse diffusion process.

[0008] To achieve the above purpose, the technical scheme of the present application is: a hyperspectral image denoising diffusion model construction method, comprising:

[0009] Step 1, generate a test noisy low-resolution hyperspectral image by a real clean hyperspectral image;

[0010] Step 2, dynamically determine the number of compressed channels for the noisy low-resolution hyperspectral image by the elbow point method, and perform spectral compression to obtain a coefficient matrix;

[0011] Step 3, respectively perform noise conversion on each band after compression, and convert the complex noise image into an approximate Gaussian noise image;

[0012] Step 4, use a pre-trained diffusion model to dynamically determine the number of inverse diffusion embedding steps for each band after noise conversion, and perform denoising processing on each band;

[0013] Step 5, sequentially splice each band after denoising, and perform high-dimensional reconstruction through the pseudo-inverse matrix of the coefficient matrix.

[0014] Further, in step 1, the noisy low-resolution hyperspectral image is obtained by adding Gaussian noise and stripe noise to the hyperspectral image obtained by cropping and normalizing the real clean hyperspectral image in the spatial dimension, so that the hyperspectral image becomes a noisy low-resolution hyperspectral image with complex noise.

[0015] Further, in step 2, the number of compression channels is dynamically determined by the elbow point method, that is, the number of compression channels K is dynamically determined by the elbow point of the singular value curve. By dynamically determining the number of compression channels by the elbow point method, the core features of the hyperspectral image with multiple bands and overlapping and redundant information can be accurately dealt with, the data recovery effect defects caused by fixed channel number compression can be effectively made up, and finally the efficient balance between compression efficiency and data integrity is realized, so that the compression result is more in line with the demand for data quality in actual application scenarios.

[0016] Further, step 2 specifically includes:

[0017] The noisy low-resolution hyperspectral image is singular value decomposed into the product of three matrices:

[0018]

[0019] wherein the left singular matrix , the singular value matrix , and the right singular matrix , B is the number of bands of the hyperspectral image, H is the height of the image, and W is the width of the image;

[0020] The diagonal elements of the singular value matrix S, that is, the singular values, are logarithmically converted:

[0021]

[0022] wherein represents the i-th singular value, is the value obtained by logarithmic transformation of the i-th singular value;

[0023] After the singular values arranged in descending order are logarithmically converted, a straight line is constructed with the first point and the last point of the singular value logarithmic curve as the endpoints, and the straight line is used as a reference line to measure the deviation of other points on the curve. The distance between all points on the singular value logarithmic curve and the reference line is calculated:

[0024]

[0025] wherein is the first point of the singular value logarithmic curve, is the last point of the singular value logarithmic curve, is the i-th point on the singular value logarithmic curve;

[0026] The vertical distance of all points is traversed, the point with the largest distance is found, which is the elbow point, and the index corresponding to the elbow point (i.e. the serial number of the singular value arranged in descending order from zero) is added by 1 to obtain the optimal compression channel number K;

[0027] After dynamically determining the optimal compression channel number K, spectral compression is performed to obtain the coefficient matrix E and the low-dimensional feature matrix A (i.e. the low-dimensional feature matrix obtained after spectral compression of the hyperspectral image):

[0028]

[0029]

[0030] The pseudo-inverse matrix of the matrix E is:

[0031] .

[0032] Further, in step 3, the noise conversion method is: training the network architecture composed of the decoder and the encoder through the inner loop, and optimizing the auxiliary variable through the Gaussian denoiser trained through the outer loop, and iteratively optimizing the inner and outer to make the decoder output approximate Gaussian noise images to restore the complex noise images.

[0033] Further, in step 4, the inverse diffusion embedding step number of each band after noise conversion is dynamically determined by using the pre-trained diffusion model, which is: calculating the noise intensity of the approximate Gaussian noise image, and calculating the noise intensity of the noisy image of the t-th step of the pre-trained diffusion model, and selecting the diffusion time step closest to 1 as the inverse diffusion embedding step number. By adjusting the step number, the inverse process of the diffusion model starts from the starting point consistent with the true noise intensity, maximizing the denoising accuracy.

[0034] Further, the inverse diffusion embedding step number calculation formula is as follows:

[0035]

[0036] Wherein, represents the noise intensity of the approximate Gaussian noise image, represents the noise intensity of the noisy image of the t-th step of the diffusion model.

[0037] The application also provides a hyperspectral image denoising diffusion model construction system for executing the hyperspectral image denoising diffusion model construction method.

[0038] The data acquisition module generates a test noisy low-resolution hyperspectral image by using a real clean hyperspectral image;

[0039] ​The data processing module uses the elbow method to perform dynamic spectral compression on noisy low-resolution hyperspectral images. After noise conversion, the inverse diffusion embedding step is dynamically determined so that the inverse process of the diffusion model starts from a starting point consistent with the real noise intensity, maximizing the denoising accuracy. Finally, high-dimensional reconstruction is performed to obtain the denoised hyperspectral image.

[0040] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform a hyperspectral image denoising diffusion model construction method as described above.

[0041] The present invention also provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements a hyperspectral image denoising diffusion model construction method as described above.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. This invention proposes an adaptive and collaborative robust denoising paradigm that integrates adaptive spectral channel compression and band-level dynamic diffusion step embedding to construct an end-to-end intelligent framework that can automatically adjust the processing strategy according to the hyperspectral data structure and noise characteristics.

[0044] 2. The adaptive spectral channel compression method based on singular value curve elbow detection in this invention can automatically determine the optimal number of compressed channels, capture signal energy attenuation elbows, remove redundancy and noise while retaining key information, and obtain efficient low-dimensional representation.

[0045] 3. This invention proposes a band-adaptive diffusion step embedding, which dynamically allocates the number of steps based on the noise intensity of the approximate Gaussian noise image obtained after noise conversion for each band. The high-noise channel is fully denoised, and the low-noise channel is moderately constrained, so as to achieve a dynamic balance between denoising intensity and spectral fidelity, suppressing noise and preserving subtle spectral features and spatial structure information.

[0046] Therefore, the method of the present invention is more competitive in the field of hyperspectral image restoration, especially in applications involving complex noisy hyperspectral images. Attached Figure Description

[0047] Figure 1 This is a flowchart of a hyperspectral image denoising diffusion model construction method according to the present invention;

[0048] Figure 2 This is a schematic diagram illustrating the principle of the elbow point method for adaptively and dynamically determining the number of compression channels.

[0049] Figure 3 This is a schematic diagram illustrating the principle of adaptively and dynamically determining the number of inverse diffusion embedding steps for each band.

[0050] Figure 4 These are application effect diagrams of the present invention on the CAVE dataset, Houston dataset, Washington DC (WDC) Mall dataset, and PaviaU dataset; Figure 4 In the figure, (a), (c), (e), and (g) are the original noisy pseudo-RGB images of the CAVE dataset, Houston dataset, Washington DC (WDC) Mall dataset, and PaviaU dataset, respectively; (b), (d), (f), and (h) are the corresponding pseudo-RGB images of the CAVE dataset, Houston dataset, Washington DC (WDC) Mall dataset, and PaviaU dataset after denoising processing, respectively. Detailed Implementation

[0051] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] This invention provides a method for constructing a hyperspectral image denoising diffusion model, comprising:

[0053] Step 1: Generate a noisy, low-resolution hyperspectral image for testing using a real, clean hyperspectral image;

[0054] Step 2: Dynamically determine the number of compressed channels for noisy low-resolution hyperspectral images using the elbow method, and perform spectral compression to obtain the coefficient matrix;

[0055] Step 3: Perform noise conversion on each band after compression to convert the complex noise image into an approximate Gaussian noise image;

[0056] Step 4: Using the pre-trained diffusion model, dynamically determine the number of inverse diffusion embedding steps for each band after noise conversion, and perform denoising processing for each band.

[0057] Step 5: Sequentially stitch together the denoised bands and perform high-dimensional reconstruction using the pseudo-inverse matrix of the coefficient matrix.

[0058] This invention adaptively and dynamically determines the optimal number of compressed channels using the elbow method, preserving key information while removing redundancy and noise. Furthermore, it dynamically allocates the number of steps based on the noise intensity of the approximate Gaussian noise image obtained after noise conversion for each band, ensuring sufficient denoising of high-noise channels and moderate constraint of low-noise channels, achieving a dynamic balance between denoising intensity and spectral fidelity. This invention can recover higher-quality hyperspectral images and perform hyperspectral image denoising tasks more dynamically and flexibly. This invention can be applied to multiple fields such as agricultural practice, mineral exploration, and environmental monitoring.

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0060] like Figure 1 As shown in the figure, this embodiment discloses a method for constructing a hyperspectral image denoising diffusion model, including the following steps:

[0061] Step S1: Generate a noisy, low-resolution hyperspectral image for testing using a real, clean hyperspectral image;

[0062] Step S2: Dynamically determine the number of compressed channels for the noisy hyperspectral image using the elbow method, perform spectral compression, and obtain the coefficient matrix;

[0063] Step S3: Perform noise conversion on each band after compression to convert the complex noise image into an approximate Gaussian noise image;

[0064] Step S4: Using the pre-trained diffusion model, dynamically determine the number of inverse diffusion embedding steps for each band after noise conversion and perform subsequent denoising steps.

[0065] Step S5: Sequentially stitch together the denoised bands and perform high-dimensional reconstruction using the pseudo-inverse matrix of the coefficient matrix.

[0066] Specifically, step S1 is as follows:

[0067] This embodiment uses hyperspectral images from the publicly available CAVE, Houston, Washington DC (WDC) Mall, and PaviaU datasets. Each hyperspectral image retains its channel count, is cropped in the spatial dimension, and normalized to obtain a 256×256 spatial resolution hyperspectral image. The noisy low-resolution hyperspectral image is created by adding Gaussian noise and stripe noise to the cropped image, making it a noisy image with complex noise. For adding Gaussian noise, considering that the noise values ​​of different bands may vary in the real world, different Gaussian noise values ​​are added to each band of the hyperspectral image within the range [10, 16]. For stripe noise, 33% of the bands are randomly selected to add stripe noise. Within each of these 33% selected bands, 10%-15% of the columns are randomly selected as noise interference columns in the spatial dimension. For these selected columns, a random value is added to each pixel. This random value is uniformly sampled from the interval [-0.25, 0.25] to simulate the "band-like interference" characteristics of stripe noise in real-world scenes.

[0068] Step S2 describes the elbow method for dynamically determining the number of compression channels. This involves dynamically determining the number of compression bands K using the "elbow points" of the singular value curve, allowing for the selection of the optimal compression level and thus avoiding the loss of crucial information. Figure 2 As shown. First, the noisy hyperspectral image... Singular value decomposition (SVD) can be performed to decompose the matrix into the product of three matrices:

[0069]

[0070] in, , , ;

[0071] Then, a logarithmic transformation is performed on the diagonal elements, i.e., the singular values, of matrix S:

[0072]

[0073] in, Represents the i-th singular value. It is the value obtained by performing a logarithmic transformation on the i-th singular value. The logarithmic transformation makes it easier to determine the inflection point of the singular value more accurately, makes the difference between the steep and flat parts of the curve more obvious, and accurately captures the first K singular values ​​containing more information.

[0074] After performing a logarithmic transformation on the singular values ​​arranged in descending order, a straight line is constructed using the first and last points of the singular value logarithm curve as endpoints. This line serves as a "baseline" to measure the deviation of other points on the curve from the baseline. The distances between all points on the curve and the baseline are calculated, and efficient batch computation is achieved using vector cross products.

[0075]

[0076] in, It is the first point on the singular value logarithm curve. It is the last point on the singular value logarithm curve. It is the i-th point on the singular value logarithm curve.

[0077] Iterate through the vertical distances of all points and find the point with the largest distance; this point is the elbow. The elbow is the turning point where the curve transitions from a rapid descent to a slow, flattened curve, and it deviates furthest from the baseline. Increment the index (starting from 0) of the elbow point by 1 to obtain the optimal number of compressed channels K.

[0078] After dynamically determining the optimal K, spectral compression is performed to obtain the coefficient matrix E and the low-dimensional feature matrix A (i.e., the low-dimensional feature matrix obtained after spectral compression of the hyperspectral image):

[0079]

[0080]

[0081] The pseudo-inverse of matrix E is:

[0082]

[0083] This pseudo-inverse matrix is ​​used for subsequent high-dimensional reconstruction:

[0084]

[0085] in, This is A after noise reduction. This is the result of high-dimensional reconstruction.

[0086] In step 3, the noise transformation method is as follows: The network architecture consisting of a decoder and encoder is trained through an inner loop, and an auxiliary variable is optimized through a Gaussian denoiser trained through an outer loop. This iterative optimization process, both internal and external, allows the decoder to output an approximate Gaussian noise image to reconstruct the complex noisy image. The core of the noise transformation is to discover a latent image z that is associated with the input noisy image, and the noise in z conforms to the assumption of additive white Gaussian noise (AWGN). The latent image z is generated by the encoder and must conform to the AWGN assumption. The network architecture is a U-Net network structure consisting of an encoder (en_net) and a decoder (de_net). The encoder is responsible for extracting latent features from the complex noisy image, while the decoder uses these latent features to reconstruct the noisy image itself.

[0087] Regarding iterative optimization, the optimal denoising result is gradually approximated by repeatedly adjusting variables and neural network parameters. The entire iterative optimization process is divided into an inner loop and an outer loop. The inner loop trains the encoder and decoder. The encoder observes an image with complex noise and outputs a matrix mean, which attempts to capture the latent representation of the clean image within the noisy image. A random Gaussian perturbation is superimposed, and then a Gaussian approximation of the noise is generated using z. The decoder uses z to reconstruct the complex noisy image and then judges the reconstruction effect using a loss function. The outer loop optimizes the auxiliary variables using a Gaussian denoiser to further constrain the clean image estimation; this method uses the BM3D Gaussian denoiser.

[0088] In step S3, the noise conversion method converts various types of noise into Gaussian noise. The core of the noise conversion is to minimize the following loss function:

[0089]

[0090] in, For parameters The encoder network of the variational automatic encoder (VAE). For parameters The decoder network, where m represents a clean image, and R(m) is the regularization function and λ is the weight coefficient of the regularization function, and n represents the noisy image. Regarding noise The expectation.

[0091] After noise transformation, the latent image z obtained by the encoder is:

[0092]

[0093] in, Let i be the i-th band in A; This represents the noise conversion function.

[0094] The process of dynamically determining the number of inverse diffusion embedding steps described in step S4 involves adjusting the diffusion step size based on the Gaussian noise distribution obtained from the noise transformation to determine the optimal diffusion time step, such as... Figure 3 As shown, the diffusion model essentially simulates a forward process of progressively adding Gaussian noise to a clean image and a reverse denoising process of recovering a clean image from a noisy input. Each diffusion step corresponds to a specific Gaussian noise intensity; that is, the larger the step size, the stronger the noise. During the adaptive diffusion step size adjustment process, the optimal step size N is determined through a quantitative matching strategy to ensure that the noise intensity generated by the diffusion model is consistent with the approximate Gaussian noise intensity obtained by the noise transformation. This guarantees that the reverse denoising process of the diffusion model accurately corresponds to the actual noise level.

[0095] The forward noise addition process of the diffusion model follows a pre-defined Gaussian noise increment rule. Starting from a clean image, Gaussian noise is added to the image at each step according to fixed parameters, ultimately resulting in pure Gaussian noise after T steps. The noise addition formula for step t of the diffusion model is:

[0096]

[0097] in, For a clean image, For the noisy image at step t, The retention factor for a clean image. The weighting factor for the noise. Standard Gaussian noise;

[0098] From the noise addition formula, we can see that The noise part is Its standard deviation is:

[0099]

[0100] The intensity of the Gaussian approximation noise z obtained during the noise conversion stage is quantified using the standard deviation:

[0101]

[0102] Where std(z) is the standard deviation of z in the spatial dimension, which reflects the overall fluctuation amplitude of the noise.

[0103] The optimal number of steps N needs to satisfy the condition that the ratio of the diffusion model noise intensity to the Gaussian approximation noise intensity is closest to 1:

[0104]

[0105] Where t is the diffusion step number.

[0106] The closest result was obtained through calculation. The diffusion time step is adjusted so that the inverse denoising process of the diffusion model starts from a point consistent with the actual noise intensity, thereby maximizing the denoising accuracy.

[0107] Step S5 describes the high-dimensional reconstruction part, in which the denoised compressed channels are stitched together in sequence, and finally high-dimensional reconstruction is performed through the pseudo-inverse matrix of E to obtain a denoised high-resolution hyperspectral image. Figure 4 The diagram shows the application effect of the method of this invention on the CAVE dataset, Houston dataset, Washington DC (WDC) Mall dataset, and PaviaU dataset.

[0108] This embodiment also provides a hyperspectral image denoising diffusion model construction system based on any one of the methods described in the above embodiments, comprising:

[0109] The data acquisition module generates noisy, low-resolution hyperspectral images for testing from real, clean hyperspectral images.

[0110] The data processing module is responsible for dynamically compressing noisy hyperspectral images using the elbow method. After noise conversion, it dynamically determines the number of inverse diffusion embedding steps so that the inverse process of the diffusion model starts from a starting point consistent with the real noise intensity, maximizing the denoising accuracy. Finally, it performs high-dimensional reconstruction to obtain the denoised hyperspectral image.

[0111] This invention provides a computer-readable storage medium storing a computer program that, when executed in a computer, causes the computer to perform a hyperspectral image denoising diffusion model construction method as described above.

[0112] This invention provides an electronic device, specifically, the electronic device includes a memory and a processor, the memory stores executable code, and when the processor executes the executable code, it implements the method described in any of the embodiments.

[0113] The memory may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk drive. Communication between this system network element and at least one other network element is achieved through at least one communication interface (which can be wired or wireless), such as the Internet, wide area network, local area network, or metropolitan area network.

[0114] The bus can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc.

[0115] The memory is used to store programs. After receiving an execution instruction, the processor executes the program. The method executed by the device for defining the flow process disclosed in any of the foregoing embodiments of the present invention can be applied to the processor or implemented by the processor.

[0116] The processor may be an integrated circuit chip with signal processing capabilities. In implementation, the steps of the above methods can be completed by integrated logic circuits in the processor's hardware or by software instructions. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in the memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above method.

[0117] The computer program product of the readable storage medium provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the foregoing method embodiments. For specific implementation, please refer to the foregoing method embodiments, which will not be repeated here.

[0118] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0119] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0120] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a hyperspectral image denoising diffusion model, characterized in that, include: Step 1: Generate a noisy, low-resolution hyperspectral image for testing using a real, clean hyperspectral image; Step 2: Dynamically determine the number of compressed channels for the noisy low-resolution hyperspectral image using the elbow method, and perform spectral compression to obtain the coefficient matrix; wherein, the number of compressed channels is dynamically determined by the elbow method, that is, the number of compressed channels K is dynamically determined by the elbow of the singular value curve. Step 2 specifically includes: Noisy low-resolution hyperspectral images Perform singular value decomposition, which decomposes the matrix into the product of three matrices: Among them, the left singular matrix Singular value matrix Right singular matrix B is the number of bands in the hyperspectral image, H is the height of the image, and W is the width of the image; Perform a logarithmic transformation on the diagonal elements, i.e., the singular values, of the singular value matrix S: in, Represents the i-th singular value. It is the value obtained by performing a logarithmic transformation on the i-th singular value; After performing a logarithmic transformation on the singular values ​​arranged in descending order, a straight line is constructed using the first and last points of the singular value logarithm curve as endpoints. This line serves as a baseline to measure the deviation of other points on the curve. The distance between all points on the singular value logarithm curve and the baseline is then calculated. in, It is the first point on the singular value logarithm curve. It is the last point on the singular value logarithm curve. It is the i-th point on the singular value logarithm curve; Iterate through the vertical distances of all points, find the point with the largest distance, and this point is the elbow point; increment the index corresponding to the elbow point by 1 to obtain the optimal number of compression channels K; After dynamically determining the optimal number of compressed channels K, spectral compression is performed to obtain the coefficient matrix E and the low-dimensional feature matrix A: The pseudo-inverse of matrix E for: Step 3: Perform noise conversion on each band after compression to convert the complex noise image into an approximate Gaussian noise image; Step 4: Using the pre-trained diffusion model, dynamically determine the number of inverse diffusion embedding steps for each band after noise conversion, and perform denoising processing for each band. Step 5: Sequentially stitch together the denoised bands and perform high-dimensional reconstruction using the pseudo-inverse matrix of the coefficient matrix.

2. The method for constructing a hyperspectral image denoising diffusion model according to claim 1, characterized in that, In step 1, the noisy low-resolution hyperspectral image is obtained by cropping and normalizing a real clean hyperspectral image in the spatial dimension, and then adding Gaussian noise and stripe noise to make the hyperspectral image a noisy low-resolution hyperspectral image with complex noise.

3. The method for constructing a hyperspectral image denoising diffusion model according to claim 1, characterized in that, In step 3, the noise conversion method is as follows: the network architecture consisting of the decoder and encoder is trained through the inner loop, and the auxiliary variables are optimized through the Gaussian denoiser trained through the outer loop. The inner and outer loops are iteratively optimized so that the decoder outputs an approximate Gaussian noise image to restore the complex noise image.

4. The method for constructing a hyperspectral image denoising diffusion model according to claim 1, characterized in that, In step 4, the method for dynamically determining the number of inverse diffusion embedding steps for each band after noise conversion using the pre-trained diffusion model is as follows: calculate the noise intensity of the approximate Gaussian noise image, and at the same time calculate the noise intensity of the noise-added image at step t of the pre-trained diffusion model. Select the diffusion time step with the ratio of the diffusion model to the Gaussian approximate noise intensity closest to 1 as the number of inverse diffusion embedding steps.

5. The method for constructing a hyperspectral image denoising diffusion model according to claim 4, characterized in that, The formula for calculating the number of inverse diffusion embedding steps is as follows: in, This represents the noise intensity of an image with approximate Gaussian noise. This represents the noise intensity of the image added at step t of the diffusion model.

6. A hyperspectral image denoising diffusion model construction system that implements the hyperspectral image denoising diffusion model construction method as described in any one of claims 1-5, characterized in that, include: The data acquisition module generates noisy, low-resolution hyperspectral images for testing from real, clean hyperspectral images. The data processing module uses the elbow method to perform dynamic spectral compression on noisy low-resolution hyperspectral images. After noise conversion, the inverse diffusion embedding step is dynamically determined so that the inverse process of the diffusion model starts from a starting point consistent with the real noise intensity, maximizing the denoising accuracy. Finally, high-dimensional reconstruction is performed to obtain the denoised hyperspectral image.

7. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed in a computer, causes the computer to perform a hyperspectral image denoising diffusion model construction method as described in any one of claims 1-5.

8. A computing device, characterized in that, It includes a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements a hyperspectral image denoising diffusion model construction method as described in any one of claims 1-5.

Images