Physical prior and frequency domain attention mixed denoising method and system for ultraviolet imaging

Abstract

The invention provides an ultraviolet imaging-oriented physical prior and frequency domain attention hybrid denoising method and system, and the method comprises the steps: calculating a row offset item and a column offset item of an original ultraviolet image through a preprocessing module, and carrying out the gating inhibition over-compensation processing to obtain a first processing image; calculating two per-pixel noise parameter maps of the original ultraviolet image based on an estimator module, and calculating a de-noised map based on the two per-pixel noise parameter maps and the first processing map; performing layer normalization processing on the de-noised image to obtain a second processed image, processing the second processed image through a frequency domain gating block and an axial attention block in sequence to obtain a third processed image, removing isolated and clustered bright spots in the third processed image based on a false point suppression module to obtain a fourth processed image, and performing inverse transformation to obtain a fifth processed image; and calculating a transformation matrix based on the fifth processing image, the first processing image and the two per-pixel noise parameter images, and obtaining a final de-noised image based on the transformation matrix and the fifth processing image.

Description

Technical Field

[0001] This invention relates to the field of ultraviolet (UV) imaging technology, and in particular to a hybrid denoising method and system for ultraviolet imaging that combines physical priors and frequency domain attention. Background Technology

[0002] Ultraviolet (UV) imaging technology has important applications in many fields, including power equipment inspection, astronomical observation, biological imaging, security monitoring, and materials science. For example, in smart grids, solar-blind UV cameras are used to monitor corona discharge in high-voltage equipment, enabling the early detection of insulation defects. In astronomy, UV telescopes are used to observe the UV radiation of stars and galaxies. In life sciences, UV-excited fluorescence microscopy reveals cellular and molecular information. In safety and materials science, UV imaging is used to identify material properties. These "non-discharge" UV images typically lack strong discharge sources, have extremely low photon counts, and weak imaging signals, making them susceptible to various noise interferences. Random noise under low photon counts (also known as imaging noise) follows a Poisson statistical distribution; the weaker the signal, the stronger the relative noise. Simultaneously, UV detectors (such as ICCD / CMOS and image intensifiers) introduce readout noise during the readout process (usually approximated as a Gaussian distribution), forming a Poisson-Gaussian mixed noise model. Furthermore, some ultraviolet sensors exhibit specific noise patterns: for example, microchannel plate image intensifiers can generate spatially clustered shot noise, where a single photon event may excite locally adjacent pixels to emit light together, leading to noisy pixel clusters. If the image is compressed during storage or transmission, compression artifacts (such as blocky distortion and quantization noise) may also occur. These noises significantly reduce the signal-to-noise ratio of ultraviolet images, obscuring subtle ultraviolet feature signals and posing challenges to subsequent automated analysis and target detection.

[0003] To improve the quality of ultraviolet images and extract effective information, various image denoising (filtering) techniques have been developed across different fields. Traditional image processing methods include: median filtering (which effectively suppresses isolated impulse disturbances such as salt-and-pepper noise, but has limited effect on Gaussian / Poisson noise), bilateral filtering (which uses spatial distance and gray-level similarity weighted averaging to smooth noise while preserving some edge details, but can easily cause blurring or color cast when noise is strong), and wavelet transform denoising (which decomposes the image into multi-scale sub-bands and applies threshold shrinkage to wavelet coefficients to attenuate noise, effectively removing Gaussian noise and preserving edge texture; however, threshold selection is sensitive, as too large a threshold will result in loss of detail, while too small a threshold will leave residual noise). Nonlocal mean (NLM) and block matching 3D filtering (BM3D) are classic non-learning methods. NLM utilizes the self-similarity of the image to reduce noise by weighted averaging of similar small pixel blocks, while BM3D further performs 3D transform domain filtering on similar block groups, thus achieving extremely high removal efficiency for anisotropic random noise. However, such transform domain methods often assume that the noise is independent and identically distributed Gaussian noise. For noise that is strongly signal-dependent or clustered, it is not easy to distinguish the signal and noise characteristics in the transform domain. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a hybrid denoising method and system for ultraviolet imaging that combines physical priors and frequency domain attention, in order to eliminate or improve one or more defects existing in the prior art.

[0005] One aspect of the present invention provides a hybrid denoising method for ultraviolet imaging combining physical priors and frequency domain attention, the method comprising the following steps: The preprocessing module calculates the row and column bias terms of the original ultraviolet image, and based on the row and column bias terms, gating suppression of overcompensation is applied to obtain the first processed image; The original ultraviolet image is calculated based on two per-pixel noise parameter maps of the original ultraviolet image using the estimator module, and a denoised map is calculated based on the two per-pixel noise parameter maps and the first processed map. The denoised image is subjected to layer normalization to obtain a second processed image. The second processed image is then processed sequentially through a frequency domain gating block and an axial attention block to obtain a third processed image. The frequency domain gating block is used to perform gating noise suppression on high-frequency components, and the axial attention block is equipped with width attention processing and height attention processing. Based on the pseudo-point suppression module, isolated and clustered bright spots in the third processing image are removed to obtain the fourth processing image, and the fourth processing image is inversely transformed to obtain the fifth processing image. The transformation matrix is ​​calculated based on the fifth processing map, the first processing map, and two per-pixel noise parameter maps. The final denoised map is obtained based on the transformation matrix and the fifth processing map.

[0006] By adopting the above scheme, this invention combines the advantages of both physical modeling and structural modeling to achieve efficient suppression of complex ultraviolet noise and preservation of detail. This is mainly reflected in: 1. Physical domain prior-friendly: The estimator module stabilizes the Poisson-Gaussian noise variance and is equipped with unbiased inverse; the residuals are constrained in the original noise domain in a weighted form through inverse function transformation, significantly reducing the risk of low-count mismatch; 2. Device noise directional suppression: Through preprocessing, row / column bias and stripes are explicitly suppressed, and the pseudo-point suppression module can robustly suppress hot pixels / clustered speckles; 3. Efficient global modeling: A backbone network composed of frequency-domain gating blocks and axial attention blocks is used to obtain long-range dependencies and weaken structural artifacts at low cost through frequency-domain gating + axial attention, avoiding the memory / bandwidth overhead of traditional U-shaped structures. It does not require dependent or clustered noise, and signal and noise features can be distinguished in the transform domain.

[0007] In some embodiments of the present invention, in the step of calculating the row bias and column bias of the original ultraviolet image using a preprocessing module, the row bias and column bias are determined using the following formula: in, and These represent row bias terms and column bias terms, respectively. One-dimensional convolution processing representing the row direction, This represents one-dimensional convolution processing along the column direction. Represents the original ultraviolet image. This represents a one-dimensional pooling operation along the column direction. A one-dimensional pooling operation representing the row direction.

[0008] In some embodiments of the present invention, in the step of using gating to suppress overcompensation based on the row bias term and column bias term to obtain a first processed graph, the first processed graph is determined using the following formula: ); in, This represents the first processed image. Represents the original ultraviolet image. This indicates a learnable scalar graph. and These represent row bias terms and column bias terms, respectively.

[0009] In some embodiments of the present invention, in the step of calculating two per-pixel noise parameter maps of the original ultraviolet image based on the estimator module, the estimator module is provided with two convolutional layers, and the two convolutional layers process the original ultraviolet image respectively to obtain two per-pixel noise parameter maps.

[0010] In the specific implementation process, according to the observation model defined in this scheme, this scheme needs to estimate the row / column bias, perform the first step of denoising, and then input it into the model for post-processing. This scheme uses two one-dimensional convolutions and the mean kernel to estimate the row / column bias respectively.

[0011] In some embodiments of the present invention, in the step of calculating the denoised image based on the two per-pixel noise parameter maps and the first processing map, the denoised image is calculated using the following formula: ; in, This represents the denoised image. and Each represents a per-pixel noise parameter map. This represents the first processed image.

[0012] In some embodiments of the present invention, in the step of processing the second processed image sequentially through a frequency domain gating block and an axial attention block to obtain a third processed image, the frequency domain gating block is configured with two channels to process the second processed image respectively. The first channel is configured with an MLP layer and a layer normalization layer arranged sequentially to output the first channel result data. The second channel is configured with a frequency domain transformation layer to obtain low-frequency data and high-frequency data. The high-frequency data is processed sequentially through a two-dimensional convolution, a sigmid layer and a dot product layer to obtain high-frequency result data. The high-frequency result data and the low-frequency data are concatenated and inverse frequency domain transformation is performed to obtain the second channel result data. The first channel result data and the second channel result data are residually concatenated to obtain the third processed image.

[0013] In some embodiments of the present invention, in the step of processing the second processing graph sequentially through a frequency domain gating block and an axial attention block to obtain a third processing graph, the axial attention block obtains a query matrix, a key matrix, and a value matrix based on a query projection matrix, a key projection matrix, and a value projection matrix, respectively. The query matrix, key matrix, and value matrix are all flattened in both width and height, and in both width and height. The result of the width and height flattening is subjected to one-dimensional horizontal attention processing, and the result of the height and width flattening is subjected to one-dimensional vertical attention processing. The processing results of the one-dimensional horizontal and vertical attention processing are spliced ​​together and processed using an MLP layer to obtain the third processing graph.

[0014] In some embodiments of the present invention, in the step of removing isolated and clustered bright spots in the third processing image based on the pseudo-point suppression module to obtain the fourth processing image, the pseudo-point suppression module is provided with an average pooling convolutional layer and a softclip layer; in the step of performing an inverse transformation on the fourth processing image to obtain the fifth processing image, the fourth processing image is subjected to an inverse transformation of a function that calculates the denoised image to obtain the fifth processing image.

[0015] In some embodiments of the present invention, in the step of calculating the transformation matrix based on the fifth processing map, the first processing map, and the two per-pixel noise parameter maps, the transformation matrix is ​​calculated using the following formula: in, This represents the transformation matrix, where i represents the position i in the fifth processing image, the first processing image, and the two per-pixel noise parameter images. This represents the value of position i in the first processing graph. and These represent the values ​​at position i in the two per-pixel noise parameter maps, respectively. This represents the value of position i in the fifth processing diagram. and All of these are preset calculation parameters.

[0016] In some embodiments of the present invention, in the step of obtaining the final denoised image based on the transformation matrix and the fifth processing image, the final denoised image is determined using the following formula: ; in, This represents the final denoised image. Represents each position of the transformation matrix Calculate the gradient. This represents the fifth processing diagram. These are the preset calculation parameters. This means setting values ​​less than 0 to 0.

[0017] A second aspect of the present invention also provides a hybrid denoising system for physical priors and frequency domain attention for ultraviolet imaging. The system includes a computer device, the computer device including a processor and a memory, the memory storing computer instructions, and the processor executing the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method described above.

[0018] A third aspect of the invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned hybrid denoising method for ultraviolet imaging combining physical priors and frequency domain attention.

[0019] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.

[0020] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0021] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0022] Figure 1 This is a schematic diagram of one implementation of the physical prior and frequency domain attention hybrid denoising method for ultraviolet imaging in this scheme; Figure 2 This is a schematic diagram of the frequency domain gating block in this scheme; Figure 3 This is a schematic diagram of the axial attention block in this scheme; Figure 4 This is a schematic diagram of the pseudopoint suppression module in this scheme; Figure 5 This is a schematic diagram of the estimator module in this scheme; Figure 6 The denoising performance of this scheme on synthetic noise data in one experimental group; Figure 7 This is the denoising performance of this scheme on the synthetic noise data of another experimental group. Detailed Implementation

[0023] 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 embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0024] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0025] In existing technologies, when noise exhibits spatial correlation or clustered distribution, traditional wavelet or BM3D methods struggle to identify and remove this structured noise in a timely manner.

[0026] Statistical modeling methods theoretically seek optimal estimates by constructing probabilistic models of noise and images. For example, Bayesian estimation methods assume that the prior noise-free image follows a certain probability distribution (such as a smoothness prior, a sparse prior, etc.), and derive the maximum a posteriori (MAP) solution by combining the likelihood model of the noise (Poisson or Gaussian distribution). Classical Bayesian denoising includes Wiener filtering (which makes the optimal linear estimate based on the power spectra of the signal and noise in the frequency domain) and BayesShrink based on wavelet coefficients (adaptively adjusting the threshold). These methods can achieve a higher signal-to-noise ratio than empirical filtering for specific noise models, but they require accurate estimation of noise statistical parameters, and linear filtering is difficult to handle nonlinear distortion. Markov random field (MRF) models characterize the image prior by assigning an interaction potential function to neighboring pixels (such as encouraging local smoothing and edge preservation), and then construct an energy function by combining it with the noise model and using iterative optimization (such as Gibbs sampling or belief propagation) to recover the image. MRF can incorporate spatial context information and has been studied for the restoration of Poisson noise images. However, it has high computational complexity, is prone to getting trapped in local extrema, and improper selection of prior models may lead to oversmoothing.

[0027] In recent years, deep learning methods have achieved breakthroughs in image denoising. Convolutional Neural Networks (CNNs) can learn nonlinear functions that map noisy images to noise-free images from a large number of samples, and have been shown to outperform traditional algorithms under various noise conditions, including Gaussian and Poisson noise. In particular, deep CNNs can combine multi-scale convolution and residual learning to extract details, powerfully denoising while preserving edges. In addition, self-supervised and unsupervised denoising strategies have been proposed in recent years. For example, Noise2Noise uses pairs of independent noisy images as labels to train the denoiser, and the Noise2Void / Blindspot method even trains the model using only a single noisy image by occluding some pixels. However, when the noise does not satisfy the ideal independent and identically distributed assumption (such as the presence of spatially correlated complex noise), the simplified noise assumptions used by many self-supervised methods can lead to performance degradation.

[0028] In summary, existing ultraviolet image denoising methods still suffer from four common shortcomings: 1. Statistical mismatch – many methods default to additive independent Gaussian noise, making it difficult to characterize Poisson-Gaussian signal-related noise under low photon counts and device-specific noise such as spatial clustering and shot caused by MCP image intensifiers; 2. Domain mismatch – most methods operate in the image domain after sRGB / ISP, losing sensor parameters such as black level and gain, making it difficult to constrain data consistency in the RAW physical domain; 3. Insufficient structural modeling – traditional transform domains and local convolutions have limited ability to distinguish non-stationary, long-range correlated noise such as stripes / FPN, hot pixels, and compression artifacts, easily resulting in over-smoothing of details or artifact residue.

[0029] To address this, this scheme addresses both the physical and structural aspects simultaneously: in the RAW domain, it first employs row / column bias suppression (RAW-Prep) and adaptive generalized Anscombe variance stabilization (PSL) to align with Poisson-Gaussian statistics; on the representation side, it uses a non-UNet serial backbone, combining FourierGate frequency-domain gating with axial token-mixer to model long-range dependencies at low cost and suppress stripe / block artifacts; then, HPS is used to apply robust MAD soft clamping to isolated / clustered bright spots; finally, PGC performs a one-step physical consistency update in the original noise domain to ensure that the residuals are protected. Weighted constraints. This design is compatible with self-supervised paradigms such as Noise2Noise / Noise2Void to alleviate the scarcity of net images, and can also be optionally connected to generative refinement modules such as posterior sampling to achieve more robust noise reduction and detail fidelity under ultraviolet imaging conditions with low counts, strong device noise, and structural noise.

[0030] like Figure 1 As shown, in step S100, the preprocessing module calculates the row bias term and column bias term of the original ultraviolet image, and based on the row bias term and column bias term, gating suppression of overcompensation is performed to obtain the first processed image; Step S200: Calculate two per-pixel noise parameter maps of the original ultraviolet image based on the estimator module, and calculate a denoised map based on the two per-pixel noise parameter maps and the first processed map; In the specific implementation process, such as Figure 5 As shown, a lightweight 3x3 Conv estimator is used to estimate the noise parameter map per pixel. Then, after removing the bias... Apply the generalized Anscomble transform.

[0031] Specifically, the estimator module is a small model consisting of stacked 2D conv + relu + 2D conv + relu + 2D conv, and finally connected to an nn.softplus activation function.

[0032] Step S300: The denoised image is subjected to layer normalization to obtain a second processed image. The second processed image is then processed sequentially through a frequency domain gating block and an axial attention block to obtain a third processed image. The frequency domain gating block is used to perform gating noise suppression on high-frequency components. The axial attention block is configured with width attention processing and height attention processing. In the specific implementation process, the backbone network of this scheme adopts alternating stacked frequency domain gating blocks and axial attention blocks. In the step of processing the second processing map sequentially through the frequency domain gating blocks and axial attention blocks to obtain the third processing map, this scheme sets up sequentially arranged frequency domain gating blocks and axial attention blocks in each processing block. The third processing map is obtained by processing with at least one processing block. If there are multiple processing blocks, the multiple processing blocks are connected in series.

[0033] Step S400: Based on the pseudo-point suppression module, isolated and clustered bright spots in the third processing image are removed to obtain the fourth processing image. The fourth processing image is then inversely transformed to obtain the fifth processing image. Step S500: Calculate the transformation matrix based on the fifth processing map, the first processing map, and the two per-pixel noise parameter maps, and obtain the final denoised map based on the transformation matrix and the fifth processing map.

[0034] like Figure 6 and 7 As shown, by adopting the above scheme, this invention combines the advantages of both physical modeling and structural modeling to achieve efficient suppression and detail preservation of complex ultraviolet noise. This is mainly reflected in: 1. Physical domain prior-friendly: The estimator module stabilizes the Poisson-Gaussian noise variance and is equipped with unbiased inverse; the residuals are constrained in the original noise domain in a weighted form through inverse function transformation, significantly reducing the risk of low-count mismatch; 2. Device noise directional suppression: Through preprocessing, row / column bias and stripes are explicitly suppressed, and the pseudo-point suppression module can robustly suppress hot pixels / clustered speckles; 3. Efficient global modeling: A backbone network composed of frequency-domain gating blocks and axial attention blocks is used to obtain long-range dependencies and weaken structural artifacts at low cost with frequency-domain gating + axial attention, avoiding the memory / bandwidth overhead of traditional U-shaped structures. It does not require dependent or clustered noise, and signal and noise features can be distinguished in the transform domain.

[0035] In some embodiments of the present invention, in the step of calculating the row bias and column bias of the original ultraviolet image using a preprocessing module, the row bias and column bias are determined using the following formula: in, and These represent row bias terms and column bias terms, respectively. One-dimensional convolution processing representing the row direction, This represents one-dimensional convolution processing along the column direction. Represents the original ultraviolet image. This represents a one-dimensional pooling operation along the column direction. A one-dimensional pooling operation representing the row direction.

[0036] In some embodiments of the present invention, before the step of calculating the row bias term and column bias term of the original ultraviolet image using the preprocessing module, the method further includes determining whether a black level exists, and if so, removing the black level from the original ultraviolet image.

[0037] In some embodiments of the present invention, in the step of using gating to suppress overcompensation based on the row bias term and column bias term to obtain a first processed graph, the first processed graph is determined using the following formula: ); in, This represents the first processed image. Represents the original ultraviolet image. This indicates a learnable scalar graph. and These represent row bias terms and column bias terms, respectively.

[0038] In some embodiments of the present invention, in the step of calculating two per-pixel noise parameter maps of the original ultraviolet image based on the estimator module, the estimator module is provided with two convolutional layers, and the two convolutional layers process the original ultraviolet image respectively to obtain two per-pixel noise parameter maps.

[0039] In the specific implementation process, the two per-pixel noise parameter maps are respectively the per-pixel noise parameter maps corresponding to the Gaussian readout noise variance in the generalized Anscombe. Plot of per-pixel noise parameters corresponding to gain .

[0040] In the specific implementation process, according to the observation model defined in this scheme, this scheme needs to estimate the row / column bias, perform the first step of denoising, and then input it into the model for post-processing. This scheme uses two one-dimensional convolutions and the mean kernel to estimate the row / column bias respectively.

[0041] In some embodiments of the present invention, in the step of calculating the denoised image based on the two per-pixel noise parameter maps and the first processing map, the denoised image is calculated using the following formula: ; in, This represents the denoised image. and Each represents a per-pixel noise parameter map. This represents the first processed image.

[0042] Specifically, a lightweight 3x3 Conv estimator is first used to estimate the noise parameter map per pixel. Then, after removing the bias... Apply the generalized Anscomble transform; Stabilizing the Poisson-Gaussian noise variance to an approximate constant facilitates subsequent learning of the network within an approximate AWGN framework.

[0043] like Figure 2 As shown, in some embodiments of the present invention, in the step of processing the second processed image sequentially through a frequency domain gating block and an axial attention block to obtain a third processed image, the frequency domain gating block has a structure with two channels to process the second processed image respectively. The first channel has an MLP layer and a layer normalization layer arranged sequentially to output the first channel result data. The second channel has a frequency domain transformation layer to obtain low-frequency data and high-frequency data. The high-frequency data is processed sequentially through a two-dimensional convolution, a sigmid layer and a dot product layer to obtain high-frequency result data. The high-frequency result data and the low-frequency data are concatenated and inverse frequency domain transformation is performed to obtain the second channel result data. The first channel result data and the second channel result data are residually concatenated to obtain the third processed image.

[0044] In the specific implementation process Figure 2 The 2D convolutions are depthwise separable convolutions with layer normalization. This suppresses high-frequency artifacts / stripes while preserving edge fidelity under low UV SNR.

[0045] like Figure 3 As shown, in some embodiments of the present invention, in the step of processing the second processing graph sequentially through a frequency domain gating block and an axial attention block to obtain a third processing graph, the axial attention block obtains a query matrix, a key matrix, and a value matrix based on a query projection matrix, a key projection matrix, and a value projection matrix, respectively. The query matrix, key matrix, and value matrix are all flattened in both width and height, and in both width and height. The result of the width and height flattening is subjected to one-dimensional attention processing in the horizontal direction, and the result of the height and width flattening is subjected to one-dimensional attention processing in the vertical direction. The processing results of the horizontal and vertical one-dimensional attention processing are spliced ​​together and processed using an MLP layer to obtain the third processing graph.

[0046] In practical implementation, the axial attention block of this scheme uses a mixer of axial self-attention and channel MLP, which has a lower computational cost than global attention.

[0047] Using the above scheme, the entire backbone network is interspersed with multi-scale downsampling = average pooling (rather than encoding and decoding) several times, and attention pooling is used to aggregate multi-scale features.

[0048] like Figure 4As shown, in some embodiments of the present invention, in the step of removing isolated and clustered bright spots in the third processing image based on the pseudo-point suppression module to obtain the fourth processing image, the pseudo-point suppression module is provided with an average pooling convolutional layer and a softclip layer; in the step of performing an inverse transformation on the fourth processing image to obtain the fifth processing image, the fourth processing image is subjected to an inverse transformation of a function for calculating the denoised image to obtain the fifth processing image.

[0049] The softclip layer is used to determine the matrix output by the mean pooling convolutional layer. It achieves a smooth clipping effect through gradient transparency to avoid harsh edges. First, it splits the matrix into a part greater than or equal to 0 and a part less than 0. For the part less than 0, it multiplies the part less than 0 by 0.5 and subtracts it from the part greater than or equal to 0 and the hyperparameter 1 before outputting the result.

[0050] In the step of applying the inverse transform of the function that calculates the denoised graph to the fourth processed graph, the stable domain result of the network output is... After optimal inverse transformation Returning to signal estimation To better reflect the original noise statistics, a one-step update approximating maximum likelihood is performed.

[0051] Using the above scheme, after passing through the backbone network, the output... Constructing a differential intensity map in the image domain According to the robustness threshold Soft clamping of over-threshold pixel channels Apply soft clamping to remove isolated / clustered bright spots without over-smoothing.

[0052] In some embodiments of the present invention, in the step of calculating the transformation matrix based on the fifth processing map, the first processing map, and the two per-pixel noise parameter maps, the transformation matrix is ​​calculated using the following formula: in, This represents the transformation matrix, where i represents the position i in the fifth processing image, the first processing image, and the two per-pixel noise parameter images. This represents the value of position i in the first processing graph. and These represent the values ​​at position i in the two per-pixel noise parameter maps, respectively. This represents the value of position i in the fifth processing diagram. and All of these are preset calculation parameters.

[0053] In some embodiments of the present invention, in the step of obtaining the final denoised image based on the transformation matrix and the fifth processing image, the final denoised image is determined using the following formula: ; in, This represents the final denoised image. Represents each position of the transformation matrix Calculate the gradient. This represents the fifth processing diagram. These are the preset calculation parameters. This means setting values ​​less than 0 to 0.

[0054] Specifically, the residuals in the original noise domain are forced to be weighted by the variance of physical noise.

[0055] This solution addresses the complex interferences present in non-discharge ultraviolet images under low photon count conditions, including Poisson shot noise, readout Gaussian noise, fixed pattern noise (row / column offset, stripes), hot pixels, and potentially superimposed compression artifacts. It provides a novel architecture-based noise reduction method that works in the RAW physical domain and balances noise suppression with detail preservation.

[0056] In summary, this scheme prioritizes organizing the dataset in the RAW / monochrome domain during the dataset construction and calibration phase to ensure that subsequent algorithms can fully utilize the separability of linear response and noise statistics. First, a stable acquisition environment is established for the target ultraviolet camera, and an ultraviolet bandpass filter with clear cutoff characteristics is configured in the optical path to reduce systematic bias caused by long-wavelength crosstalk. To ensure parameter identifiability, dark-field frames and flat-field frames are acquired simultaneously under the same exposure system: the dark-field frames are used to estimate the black level and readout noise, while the flat-field frames are used to obtain the system gain and possible column / row bias trends. For dark scenes, full-image pixel statistics determine the black level using quantiles or robust means. Based on this, the standard deviation is calculated per pixel or block to obtain an initial value for readout noise. For flat scenes, the mean-variance relationship of pixel intensity is measured under different illumination / exposure combinations. The slope of its linear segment provides an initial value for gain. Simultaneously, a one-dimensional fit is performed on the flat scenes column-wise (or row-wise) to obtain a slowly varying column-wise (or row-wise) bias template, providing a reasonable prior range for the learnable gating of the RAW-Prep module. These parameters are stored as "sample-level metadata" along with each original image in the project, allowing subsequent PSL variance stabilization and PGC consistency updates to be directly referenced or used in the online estimator. Based on this, corrections were made.

[0057] In the absence of rigorously cleaned images, this scheme employs a multi-frame registration-robust fusion method to construct "pseudo-cleaned images" as training references: for short exposure sequences of the same scene, sub-pixel-level registration is first performed, followed by fusion using a median or extreme averaging strategy, thereby significantly improving the signal-to-noise ratio and suppressing occasional heated pixels. The resulting high SNR reference and corresponding observations constitute training pairs; when sufficient references are still unavailable, self-supervised training can be conducted using only real observations and calibration parameters. To improve generalization, samples are uniformly linearized and de-darkened before being imported into training, and Bayer RAW is split into four-channel inputs as needed; spatial augmentation (random cropping, rotation, flipping) and dynamic range normalization are performed without altering statistical relationships. Data partitioning follows the principle of "hierarchical randomization based on scene and exposure conditions," ensuring consistent distribution of target type, brightness, and noise intensity across training, validation, and testing subsets, thus guaranteeing the comparability and stability of evaluation results.

[0058] This scheme ensures training feasibility using a self-supervised paradigm. Its core approach involves generating blind spot masks in the stable or RAW domain, predicting the masked pixels using only neighborhood information, and calculating the reconstruction error between the prediction and the original observation at the mask location. Since blind spots eliminate self-copying, the network is forced to learn conditional mean estimation under noisy conditions, thus statistically tending towards unbiasedness. The training loss objective is defined as follows: in is the DWT or DCT operator, and ∇ is the Sobel / Prewitt gradient. Recommended weights: The recommended gradient descent configuration is as follows: optimizer AdamW; learning rate... Weight decay Linear annealing (1–5k steps).

[0059] Optional implementation methods and variations: Without departing from the spirit of the present invention, the frequency domain operator can be replaced by DWT, DCT, learnable wavelet or FFT subband filtering; in extremely low SNR scenarios, a small number of steps of energy reduction or sampling refinement can be introduced after PGC to recover weak details.

[0060] Implementation Considerations: In practical applications, priority should be given to ensuring that the linear relationship in the RAW domain is not disrupted, and irreversible nonlinear processing should be avoided in the ISP; updates should be made regularly with the latest dark / flat field data. The estimation is to prevent statistical drift caused by ambient temperature and device aging; for Bayer-type sensors, four-channel RAW can be directly input into the network, or single-channel processing can be performed after high-quality demosaicing, and the choice between the two should be based on minimizing the overall system error; in extremely low photon scenarios, appropriately increasing the lower bound constants related to variance in PSL and PGC helps to maintain numerical stability and avoid gradient explosion.

[0061] The noise reduction performance of each module in this invention is shown in Table 1 below: Table 1 This invention also provides a hybrid denoising system for physical priors and frequency domain attention for ultraviolet imaging. The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method described above.

[0062] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned hybrid denoising method for ultraviolet imaging combining physical priors and frequency domain attention. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0063] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0064] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0065] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A hybrid denoising method combining physical priors and frequency domain attention for ultraviolet imaging, characterized in that, The steps of this method include: The preprocessing module calculates the row and column bias terms of the original ultraviolet image, and based on the row and column bias terms, gating suppression of overcompensation is applied to obtain the first processed image; The original ultraviolet image is calculated based on two per-pixel noise parameter maps of the original ultraviolet image using the estimator module, and a denoised map is calculated based on the two per-pixel noise parameter maps and the first processed map. The denoised image is subjected to layer normalization to obtain a second processed image. The second processed image is then processed sequentially through a frequency domain gating block and an axial attention block to obtain a third processed image. The frequency domain gating block is used to perform gating noise suppression on high-frequency components, and the axial attention block is equipped with width attention processing and height attention processing. Based on the pseudo-point suppression module, isolated and clustered bright spots in the third processing image are removed to obtain the fourth processing image, and the fourth processing image is inversely transformed to obtain the fifth processing image. The transformation matrix is ​​calculated based on the fifth processing map, the first processing map, and two per-pixel noise parameter maps. The final denoised map is obtained based on the transformation matrix and the fifth processing map.

2. The hybrid denoising method for ultraviolet imaging based on physical priors and frequency domain attention as described in claim 1, characterized in that, In the step of calculating the row and column bias terms of the original ultraviolet image using the preprocessing module, the row and column bias terms are determined using the following formula: in, and These represent row bias terms and column bias terms, respectively. One-dimensional convolution processing representing the row direction, This represents one-dimensional convolution processing along the column direction. Represents the original ultraviolet image. This represents a one-dimensional pooling operation along the column direction. A one-dimensional pooling operation representing the row direction.

3. The hybrid denoising method for ultraviolet imaging based on physical priors and frequency domain attention as described in claim 1, characterized in that, In the step of using gating to suppress overcompensation based on the row and column bias terms to obtain the first processed graph, the first processed graph is determined using the following formula: ）； in, This represents the first processed image. Represents the original ultraviolet image. This indicates a learnable scalar graph. and These represent row offset terms and column offset terms, respectively.

4. The hybrid denoising method for ultraviolet imaging based on physical priors and frequency domain attention as described in claim 1, characterized in that, In the step of calculating two per-pixel noise parameter maps of the original ultraviolet image based on the estimator module, the estimator module is equipped with two convolutional layers, which process the original ultraviolet image respectively to obtain two per-pixel noise parameter maps.

5. The hybrid denoising method for ultraviolet imaging based on physical priors and frequency domain attention as described in claim 1, characterized in that, In the step of calculating the denoised image based on the two per-pixel noise parameter maps and the first processed image, the denoised image is calculated using the following formula: ； in, This represents the denoised image. and Each represents a per-pixel noise parameter map. This represents the first processed image.

6. The hybrid denoising method for ultraviolet imaging based on physical priors and frequency domain attention as described in claim 1, characterized in that, In the step of processing the second processed image sequentially through a frequency domain gating block and an axial attention block to obtain the third processed image, the frequency domain gating block has two channels to process the second processed image. The first channel has an MLP layer and a layer normalization layer set sequentially to output the first channel result data. The second channel has a frequency domain transformation layer to obtain low-frequency data and high-frequency data. The high-frequency data is processed sequentially through a two-dimensional convolution, a sigmid layer, and a dot product layer to obtain high-frequency result data. The high-frequency result data and the low-frequency data are concatenated and inverse frequency domain transformation is performed to obtain the second channel result data. The first channel result data and the second channel result data are residually concatenated to obtain the third processed image.

7. The hybrid denoising method for ultraviolet imaging based on physical priors and frequency domain attention as described in claim 1, characterized in that, In the step of processing the second processed image sequentially through a frequency domain gating block and an axial attention block to obtain the third processed image, the axial attention block obtains a query matrix, a key matrix, and a value matrix based on the query projection matrix, the key projection matrix, and the value projection matrix, respectively. The query matrix, the key matrix, and the value matrix are all flattened in both width and height, and in both width and height. The result of the width and height flattening is subjected to one-dimensional horizontal attention processing, and the result of the height and width flattening is subjected to one-dimensional vertical attention processing. The processing results of the one-dimensional horizontal and vertical attention processing are spliced ​​together and processed using an MLP layer to obtain the third processed image.

8. The hybrid denoising method for ultraviolet imaging based on physical priors and frequency domain attention as described in claim 5, characterized in that, In the step of removing isolated and clustered bright spots in the third processing image based on the pseudo-point suppression module to obtain the fourth processing image, the pseudo-point suppression module is equipped with an average pooling convolutional layer; in the step of performing an inverse transformation on the fourth processing image to obtain the fifth processing image, the fourth processing image is subjected to an inverse transformation of a function that calculates the denoised image to obtain the fifth processing image.

9. The hybrid denoising method for ultraviolet imaging based on physical priors and frequency domain attention according to any one of claims 1 to 8, characterized in that, In the step of calculating the transformation matrix based on the fifth processing map, the first processing map, and the two per-pixel noise parameter maps, the transformation matrix is ​​calculated using the following formula: in, This represents the transformation matrix, where i represents the position i in the fifth processing image, the first processing image, and the two per-pixel noise parameter images. This represents the value of position i in the first processing graph. and These represent the values ​​at position i in the two per-pixel noise parameter maps, respectively. This represents the value of position i in the fifth processing diagram. and All of these are preset calculation parameters.

10. The hybrid denoising method for ultraviolet imaging based on physical priors and frequency domain attention as described in claim 9, characterized in that, In the step of obtaining the final denoised image based on the transformation matrix and the fifth processing image, the final denoised image is determined using the following formula: ； in, This represents the final denoised image. Represents each position of the transformation matrix Calculate the gradient. This represents the fifth processing diagram. These are the preset calculation parameters. This means setting values ​​less than 0 to 0.

Images