A polarized image denoising method and system based on global collaborative transform BM3D filtering

Abstract

The application discloses a polarized image denoising method and system based on global collaborative transformation BM3D filtering, belongs to the technical field of image denoising, and solves the problem that traditional denoising methods based on Gaussian white noise assumption cannot effectively suppress spatial correlation noise and are prone to excessively smooth polarized features, thereby reducing the polarized navigation attitude measurement precision. The method comprises the following steps: S1, acquiring a polarized angle image and a polarized degree image corresponding to a to-be-processed polarized image; S2, determining the noise power spectral density of the to-be-processed polarized image; S3, based on the noise power spectral density, adopting a global collaborative transformation BM3D filtering algorithm to collaboratively denoise the polarized angle image and the polarized degree image respectively, and outputting a denoised polarized image. The application is suitable for the scene of bionic polarized light navigation and polarized vision attitude measurement under a complex atmospheric environment.

Description

Technical Field

[0001] This invention belongs to the field of image denoising technology, specifically relating to a polarization image denoising method based on global cooperative transformation BM3D filtering. Background Technology

[0002] Biomimetic polarized light navigation has become an important research direction in the field of autonomous navigation for unmanned vehicles, with broad application prospects in scenarios such as small unmanned vehicles and ground robots. The core of polarized visual attitude measurement lies in extracting key features such as the solar vector and neutral point from atmospheric polarized images, and then calculating the three-dimensional attitude of the vehicle; the quality of the polarized image directly determines the accuracy of feature extraction and the precision of attitude measurement.

[0003] In practical engineering applications, polarization images are inevitably subject to interference from both the external environment and the imaging system. On the one hand, under complex weather conditions such as fog, haze, and sandstorms, the concentration of atmospheric aerosols increases significantly. After sunlight is scattered multiple times by aerosol particles, significant spatially correlated noise is generated in the polarization image, causing the noise to no longer satisfy the independent and identically distributed characteristics, but exhibiting obvious spatial correlation and non-uniform distribution features. On the other hand, the photoelectric conversion, signal readout, and analog-to-digital conversion processes of the polarization camera itself introduce random noise such as shot noise, dark current noise, and reset noise, further reducing the image signal-to-noise ratio. The superposition of these two types of noise leads to increased deviation in solar vector estimation and decreased neutral point positioning accuracy, severely restricting the performance and environmental adaptability of polarization navigation systems under complex weather conditions. Therefore, effective denoising processing of polarization images to suppress aerosol scattering noise and sensor random noise, while preserving polarization feature structure information, is a key prerequisite for improving the accuracy of polarization navigation under complex weather conditions.

[0004] Traditional BM3D algorithms are typically based on additive white Gaussian noise models, assuming that the noises are independent of each other and using a uniform noise variance for modeling during collaborative filtering. However, this assumption is often difficult to hold in complex atmospheric environments, which significantly limits the denoising performance.

[0005] Data-driven deep learning denoising methods, with their powerful nonlinear fitting capabilities, have shown potential for restoration in complex noisy scenes. Related research has achieved polarization image restoration through architectures such as generative adversarial networks and improved U-Net. However, the performance of these methods is highly dependent on large-scale labeled polarization image datasets. The field of sky polarization navigation lacks publicly available high-quality labeled datasets, and self-built datasets suffer from limited scene coverage and a single data distribution, resulting in insufficient model generalization ability and significant performance degradation in unknown weather and environments. At the same time, deep learning models have weak physical interpretability and high computational complexity, making it difficult to meet the real-time requirements of embedded navigation platforms and hindering engineering implementation.

[0006] In addition, some studies have optimized noise post-processing at the attitude result level, using methods such as multi-scale decomposition, outlier detection, and adaptive filtering to smooth the heading angle sequence. However, these methods do not address the polarization information degradation problem at the image source; they can only alleviate the fluctuations in the output results to a certain extent and cannot fundamentally improve the accuracy of polarization feature extraction, thus offering limited improvement to attitude measurement performance.

[0007] In summary, existing polarization image denoising techniques are insufficient to simultaneously meet the multiple requirements of suppressing spatially correlated noise, preserving polarization feature structures, low data dependence, and engineering feasibility under complex weather conditions. There is an urgent need for a high-precision denoising method that is tailored to the characteristics of aerosol scattering noise and adapted to the features of polarization images. Summary of the Invention

[0008] This invention provides a polarization image denoising method and system based on global cooperative transformation BM3D filtering. Its purpose is to solve the problem that the spatial correlation noise introduced by aerosol scattering under complex weather conditions is superimposed with the random noise of the imaging sensor, which leads to the degradation of polarization image features. Traditional denoising methods based on Gaussian white noise assumption are difficult to effectively suppress spatial correlation noise and are prone to over-smoothing polarization features, thereby reducing the accuracy of polarization navigation attitude measurement.

[0009] Firstly, the purpose of this invention is to provide a polarization image denoising method based on global cooperative transform BM3D filtering, comprising the following steps:

[0010] S1: Obtain the polarization angle image and polarization degree image corresponding to the polarization image to be processed;

[0011] S2: Determine the noise power spectral density of the polarization image to be processed;

[0012] S3: Based on the noise power spectral density, the global cooperative transformation BM3D filtering algorithm is used to perform cooperative denoising on the polarization angle image and the polarization degree image respectively, and output the denoised polarization image.

[0013] Furthermore, a preferred solution is provided: S3 specifically includes:

[0014] S31: The image is segmented into blocks, and the image block similarity evaluation is optimized based on the noise power spectral density. Similar image blocks are selected to construct an image block set.

[0015] S32: Perform a three-dimensional cooperative transformation on the image patch set to obtain the transform domain spectral coefficients;

[0016] S33: Attenuate and shrink the transform domain spectral coefficients according to the noise power spectral density;

[0017] S34: Perform inverse transformation and weighted aggregation on the attenuated spectral coefficients to obtain a denoised image.

[0018] Furthermore, a preferred solution is provided: In step S31, optimizing the image patch similarity evaluation based on the noise power spectral density specifically involves:

[0019] In the calculation of image block matching distance, a bias correction term caused by spatial correlation noise is introduced to eliminate the interference of correlation noise on similarity evaluation and obtain unbiased block matching distance.

[0020] Furthermore, a preferred embodiment is provided: in S33, the attenuation and shrinkage processing of the transform domain spectral coefficients includes a hard threshold filtering stage and a Wiener filtering stage;

[0021] In the hard threshold filtering stage, spectral coefficients below a preset threshold are set to zero.

[0022] In the Wiener filtering stage, the denoised image output from the hard threshold filtering stage is used as the prior. The Wiener filtering coefficients are calculated based on the prior signal spectrum and the noise power spectral density, and the spectral coefficients are weighted and attenuated.

[0023] Furthermore, a preferred embodiment is provided: In step S3, a global Fourier threshold re-filtering step is included before outputting the denoised polarization image, specifically as follows:

[0024] A global Fourier transform is performed on the residual image after preliminary denoising to obtain the residual spectrum;

[0025] The residual spectrum is compared with the noise power spectral density to identify the spectral components corresponding to the over-smoothed image details;

[0026] Based on the test results, a re-filtered input image is constructed, and the global collaborative transformation BM3D filtering process is performed again to preserve the polarization feature structure information.

[0027] Furthermore, a preferred embodiment is provided: the global Fourier threshold re-filtering step is performed once in the hard threshold filtering stage and once in the Wiener filtering stage;

[0028] After performing re-filtering in the hard thresholding stage, the output is used as the prior image for the Wiener filtering stage; after performing re-filtering in the Wiener filtering stage, the final denoised polarization image is output.

[0029] Furthermore, a preferred embodiment is provided: In step S2, determining the noise power spectral density of the polarization image to be processed specifically involves:

[0030] The spatially correlated noise caused by aerosol scattering is modeled as stationary correlated noise, and the power spectral density of the corresponding noise is obtained by Fourier transforming the spatial correlation convolution kernel of the noise.

[0031] Secondly, the purpose of this invention is to propose a polarization image denoising system based on global cooperative transform BM3D filtering, used to implement the polarization image denoising method based on global cooperative transform BM3D filtering described in any one or more of the above-mentioned schemes, including:

[0032] Image acquisition module: used to acquire the polarization angle image and polarization degree image corresponding to the polarization image to be processed;

[0033] Noise modeling module: used to determine the noise power spectral density of the polarization image to be processed, wherein the noise includes spatially correlated noise caused by aerosol scattering and random noise from the imaging sensor;

[0034] Collaborative denoising module: Based on the noise power spectral density, it uses the global collaborative transformation BM3D filtering algorithm to perform collaborative denoising on the polarization angle image and polarization degree image respectively, and outputs the denoised polarization image.

[0035] Thirdly, the present invention aims to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes the polarization image denoising method based on global cooperative transformation BM3D filtering according to any one or more of the above-described schemes.

[0036] Fourthly, the present invention aims to provide a computer-readable storage medium for storing a computer program that executes the polarization image denoising method based on global cooperative transformation BM3D filtering described in any one or more of the above-described schemes.

[0037] Compared with the prior art, the advantages of the present invention are:

[0038] This invention addresses the shortcomings of traditional BM3D methods in handling spatially correlated noise by proposing a BM3D polarization image denoising method based on global cooperative transformation. It achieves sparse representation through non-local similar block grouping and 3D cooperative transformation, and combines noise modeling and frequency domain threshold adjustment strategies to effectively suppress two types of noise while mitigating over-smoothing. Experimental results show that under hazy weather conditions, the proposed method improves the average PSNR of polarization angle images by 17.32% and 10.96% compared to no denoising and mean filtering, respectively; under artificially constructed local dust environments, the average PSNR is improved by 33.21% and 13.06%, respectively, verifying the effectiveness of the proposed method under complex noise conditions and providing reliable data support for subsequent solar vector estimation and attitude measurement.

[0039] This invention is applicable to biomimetic polarized light navigation and polarized visual attitude measurement scenarios in complex atmospheric environments. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0041] Figure 1 The flowchart shows a polarization image denoising method based on global cooperative transform BM3D filtering.

[0042] Figure 2 The flowchart shows the polarization image denoising preprocessing process based on global cooperative transform BM3D.

[0043] Figure 3 A comparison of image denoising methods with different polarization angles under actual hazy weather conditions;

[0044] Figure 4 Comparison of denoising methods for images with different polarization angles in a man-made local dust storm environment;

[0045] Figure 5 A comparison of PSNR of different polarization image denoising methods under actual hazy weather conditions;

[0046] Figure 6 A comparison of PSNR for different polarization image denoising methods in a man-made local dusty environment. Detailed Implementation

[0047] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0048] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0050] Implementation Method 1

[0051] This embodiment proposes a polarization image denoising method based on global cooperative transform BM3D filtering, including the following steps:

[0052] Obtain the noisy polarization image as actually observed: At this point, the image is a noise-free image. Smoothed additive noise The results of pollution are expressed as:

[0053] , (1)

[0054] in, express 3D image domain Coordinates in the noise Including spatially correlated noise introduced by aerosol scattering and zero-mean Gaussian random noise generated by the imaging sensor, it can be expressed as:

[0055] , (2)

[0056] in, This represents Gaussian noise with a zero mean and unit variance distribution. Convolution kernels representing spatial correlation of noise, This represents the Kronecker product operation.

[0057] because The variance is 1, therefore the noise The variance can be represented .

[0058] Meanwhile, the relevant noise can also be detected by the power spectral density. express:

[0059] , (3)

[0060] in, This represents a two-dimensional Fourier transform.

[0061] The goal of polarization image preprocessing is to extract polarization from images affected by noise. Estimating a noise-free image To suppress the impact of sensor random noise and aerosol scattering noise under complex weather conditions on the estimation results of the sun's position and neutral point position, this method employs a global cooperative transform-based BM3D denoising strategy to denoise the polarization angle and degree of polarization images. The power spectral density of the relevant noise is effectively estimated through global cooperative transform, and combined with global Fourier transform and re-filtering strategies, noise is suppressed while reducing image detail loss caused by over-smoothing. The overall algorithm flow is as follows: Figure 1 As shown.

[0062] The original noisy polarization image and noise power spectral density As input, BM3D hard thresholding is first performed to generate a denoised image. After that, Perform global Fourier thresholding to generate a new noisy image. and the corresponding power spectral density .right Perform BM3D hard thresholding filtering again to generate new denoised image estimates. This estimated value is used as the prior value for subsequent BM3D Wiener filtering. After the first BM3D Wiener filtering, the estimated value is output. . After a second Fourier transform thresholding process and a second BM3D Wiener filter, the final denoised polarization image can be output. .

[0063] Figure 2 A polarization image denoising process based on global cooperative transformation BM3D is presented. The cooperative filtering process processes a set of local image patches with similar structural features extracted from the original image.

[0064] set up Indicates from image Extracted from A set of noisy two-dimensional image patches, with coordinates as follows: Each image patch has the same shape, and the total number of pixels is denoted as . The coordinates are represented by the coordinates of the first element. Therefore, for a two-dimensional image, the coordinates of the image block are the coordinates of the top-left pixel. Represents a two-dimensional domain transformation operation, and uses... Indicates the first Image patches Transform spectral coefficients for The transformed first basis functions This represents the matrix inner product. This process corresponds to... Figure 2The arrows pointing from the original image block to each two-dimensional transform coefficient block indicate that a two-dimensional transform is performed on each local image block to achieve the mapping from the spatial domain to the transform domain.

[0065] Then, the length of the transformation coefficients described above will be... sequence Stack along the third dimension and perform a one-dimensional transformation. To obtain a set of Spectrum This process corresponds to Figure 2 The curved arrow in the middle indicates the mapping from the two-dimensional transformation result to the three-dimensional co-transformation, that is, joint processing of similar image patches in the non-local dimension. and These are used to index the spectral components in the local and global dimensions, respectively. The core of the image denoising algorithm based on global cooperative transform BM3D filtering lies in the calculation of... Spectrum variance , recorded as By applying threshold shrinkage and weighted filtering to the three-dimensional spectral coefficients, effective suppression of noise components can be achieved in the transform domain.

[0066] The noise variance of the spectral coefficients is expressed as:

[0067] , (4)

[0068] in, express Regarding the antisymmetric mapping of the origin Represents the square of the L2 norm. This represents the Kronecker product operation. Describes the norm, Representing the Fourier transform, the rightmost equation in the above formula can be obtained from the Parseval isometry and equation (3). After obtaining the two-dimensional spectral coefficients, The spectral coefficients can be directly obtained through and The tensor product of the transformation is calculated and expressed as:

[0069] (5)

[0070] in, Represents a one-dimensional nonlocal transformation No. The first basis function One element, This represents the convolution operation. Therefore, the aforementioned three-dimensional transform coefficients are essentially a comprehensive representation of a set of one-dimensional spectral coefficients weighted by basis functions in the non-local dimension, allowing for the further derivation of its noise variance expression:

[0071] , (6)

[0072] in, This indicates that within the global cooperative transformation domain, the first... The two-dimensional transformation basis function and the first The noise variance of the three-dimensional spectral coefficients corresponding to the one-dimensional global transformation basis functions is used to characterize the energy distribution characteristics of the correlated noise in the three-dimensional transform domain. (Symbols) Represents the spatial correlation kernel of noise Regarding the antisymmetric mapping of the origin in two-dimensional space This represents the Kronecker product operation, used to transform two-dimensional basis functions. With one-dimensional global transformation basis Combining in tensor space, where for The reverse arrangement form. Operator Represents a two-dimensional Fourier transform. The power spectral density function representing the relevant noise. This represents the total number of pixels within the image domain.

[0073] Global image patch matching is achieved by evaluating reference blocks. The similarity between the image patch and its surrounding image patches is implemented. An evaluation function is used for each image patch. Process the data and select the best and most overlapping options. An image patch set is constructed from image patches. The reference image patch and the noisy image patch are compared. Norm as an evaluation function:

[0074] , (7)

[0075] The above equation can be further expanded as follows:

[0076] , (8)

[0077] in, , and It is formed by any orthogonal transformation The generated spectrum. Let be a non-centrally distributed chi-square random variable with 1 degree of freedom, and its mean and variance be:

[0078] , (9)

[0079] , (10)

[0080] For the corresponding and The transformation can be calculated using equation (6). ,express and A special form. According to... From equation (9), the expectation of equation (8) can be expressed as:

[0081] , (11)

[0082] The above method can quantify the correlation noise bias in the image patch matching process and show that this bias is caused by the correlation noise. This is caused by the relative positions between image patches, therefore the expected value will vary. Changes with changes. By from minus The impact can be obtained An unbiased estimate is expressed as:

[0083] , (12)

[0084] Thus, the final image patch similarity evaluation function is constructed:

[0085] (13)

[0086] The first step in BM3D filtering is image patch matching, in order to generate... A set of similar image patches is used to compare the reference image patch with all image patches in the surrounding finite region. The evaluation function is shown in Equation (14).

[0087] , (14)

[0088] Among them, when This implementation can incorporate noise that differs from the reference image patch into the evaluation function. Set the experience value to 3.

[0089] Then for the image patch set Spectrum attenuation processing:

[0090] , (15)

[0091] in, The attenuation factor takes values ​​from... The decision is determined by factors such as noise statistics. Typically, hard thresholding is used in the first stage of BM3D processing, and empirical Wiener filtering is employed in the second stage.

[0092] In hard thresholding, by passing values ​​less than the threshold... Attenuation is achieved by setting the spectral coefficients to zero, since these coefficients are mainly generated by noise. The attenuation factor is expressed as:

[0093] , (16)

[0094] in, As a fixed constant, this implementation is set to an empirical value of 2.7.

[0095] In the Wiener filtering stage, the attenuation coefficient of the transfer function is calculated based on the prior value and the variance of the noise spectral coefficient, as shown in the following equation:

[0096] , (17)

[0097] in, This is the noise-free image estimate obtained during the hard thresholding stage. After calculating the attenuation factor, it can be applied... In the spectrum, thus obtaining the estimated values ​​of each image patch:

[0098] , (18)

[0099] in, for inverse transform, for The inverse transform Each basis function is used to compute a set of image patches. The residual noise variance can be used to calculate the weights of each image patch:

[0100] , (19)

[0101] in, Represents image blocks The aggregated weights will give greater weight to image patches with less residual noise, thereby reducing the impact of image artifacts.

[0102] After mapping an image to the frequency domain, an ideal noise-free image can usually be effectively characterized by a small number of high-amplitude spectral coefficients, while noise components are widely distributed among a large number of low-amplitude spectral coefficients. However, in reality, when noise energy is concentrated or compressed to a certain extent in the transform domain, even if its variance can be estimated relatively accurately, it is difficult to achieve the ideal denoising effect, and it may even cause excessive smoothing of image details.

[0103] To address the aforementioned problems, this implementation introduces a re-filtering strategy based on global Fourier threshold control. First, it analyzes the image residual... The distribution characteristics in the global frequency domain are compared with the noise power spectral density to quantitatively assess the potential loss of detail during BM3D denoising.

[0104] , (20)

[0105] in, This indicates that a Fast Fourier Transform is performed on the image residuals. Inverse transformation, This represents a morphological filter that can reduce the impact of noise through erosion-dilation operations. express The test is shown in the following formula:

[0106] , (twenty one)

[0107] Define a power spectral density that satisfies Noisy images for:

[0108] , (twenty two)

[0109] Ideally, when the collaborative filtering process can effectively suppress noise without introducing obvious over-smoothing, Approximately zero, therefore we have .at this time, Furthermore, the noise power spectral density is also zero, therefore the estimate obtained from re-filtering is different from the initial denoising result. There were no significant differences between them. However, when the collaborative filter produces over-smoothing, the image residuals... If there exists relative frequency in the spectrum A larger coefficient corresponds to that frequency. Set to 1, thus It can preserve image details and related noise information that have been over-smoothed. After global Fourier thresholding, the generated... and By re-performing the BM3D hard threshold filter and the BM3D Wiener filter, respectively, all the re-filtering steps can be completed.

[0110] The pseudocode of the BM3D polarization image denoising algorithm based on global cooperative transformation is shown in Table 1.

[0111] Table 1. Pseudocode of polarization image denoising algorithm based on global cooperative transform BM3D filtering

[0112]

[0113] Implementation Method 2

[0114] To evaluate the feasibility of the proposed polarization image denoising method based on global cooperative transform BM3D filtering, this embodiment conducted relevant experimental verification and analysis. Polarization images collected during actual hazy weather and in artificially constructed localized dust storm environments were selected as the research objects. First, the attitude information provided by the fiber optic inertial navigation and GNSS integrated navigation system was used as a reference benchmark. An ideal polarization angle image was constructed using an atmospheric polarization model, serving as a noise-free reference image. Subsequently, the actually collected polarization angle images were treated as noisy observation data, and comparative analyses were performed with the reference image before and after denoising. Finally, the peak signal-to-noise ratio (PSNR) was used as a quantitative evaluation index to assess the denoising effect, thereby verifying the proposed method's ability and effectiveness in suppressing spatially correlated noise caused by increased aerosol concentration and random noise generated by the camera itself under complex weather conditions.

[0115] Under hazy weather conditions, the Air Quality Index (AQI) is 184, indicating moderate pollution. The concentration of aerosols in the atmosphere increases significantly, thus affecting the sky polarization pattern. Furthermore, a localized dust environment was constructed using dust emission methods, creating a high aerosol concentration scene within the camera's field of view. The aerosol content near the camera increases significantly, leading to polarization information degradation and noise enhancement. This provides an effective experimental scenario for evaluating the performance of polarization image denoising algorithms under dusty weather conditions. It should be noted that this localized dust environment primarily characterizes the interference characteristics under high aerosol concentration conditions; its spatial distribution and scattering characteristics still differ from real large-scale dusty weather, but it can reflect the basic laws of polarization image degradation.

[0116] like Figure 3 and Figure 4As shown, the comparison results of polarization angle images within the ROI region are presented, including those without denoising, those with mean filtering, and those processed by BM3D based on global cooperative transformation. It can be seen that under complex weather conditions, when the concentration of aerosols in the atmosphere increases significantly, the original polarization angle image is affected by both aerosol scattering noise and sensor random noise, exhibiting obvious random disturbances and local discontinuities. After mean filtering, the random noise in the image is suppressed to some extent, but relatively obvious low-frequency noise still remains in local areas, indicating that its ability to suppress spatially correlated noise caused by aerosol scattering under complex weather conditions is limited. In contrast, the BM3D denoising method based on global cooperative transformation can effectively suppress spatially correlated noise caused by aerosol scattering and random noise generated by the sensor itself, while better preserving the structural information and solar meridian characteristics of the polarization angle image, making the polarization angle distribution more continuous and stable, and significantly reducing local abnormal disturbances.

[0117] Figure 5 and Figure 6 The image shows a comparison of the peak signal-to-noise ratio (PSNR) of the polarization angle images before and after denoising in the aforementioned smoggy weather and artificially constructed local dust storm environments (created through dust dispersion) with the baseline image. The calculation formula is as follows:

[0118] , (twenty three)

[0119] in, This represents the mean square error. Indicates a noise-free image in The pixel value of the location, Indicates the image after denoising. The pixel value of the location, This represents the maximum possible value of an image pixel. To intuitively describe the effect of the method described in this invention, this embodiment uses the peak signal-to-noise ratio of a pseudo-color image with polarization angle as an evaluation metric. .

[0120] As can be seen, under actual hazy weather conditions, compared with the undenoised polarization angle image and the polarization angle image after mean denoising, the average PSNR of the global cooperative transformation BM3D denoising algorithm proposed in this invention is improved by 17.32% and 10.96%, respectively. In the artificially constructed local sandstorm environment, the average PSNR is improved by 33.21% and 13.06%, respectively, verifying the feasibility of the method proposed in this invention.

[0121] Table 2 Comparison of different polarization image denoising methods under hazy weather conditions

[0122]

[0123] Table 3. Comparison of denoising methods for images with different polarizations in artificially constructed local dust storm environments.

[0124]

[0125] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A polarization image denoising method based on global cooperative transform BM3D filtering, characterized in that, Includes the following steps: S1: Obtain the polarization angle image and polarization degree image corresponding to the polarization image to be processed; S2: Determine the noise power spectral density of the polarization image to be processed; S3: Based on the noise power spectral density, the global cooperative transformation BM3D filtering algorithm is used to perform cooperative denoising on the polarization angle image and the polarization degree image respectively, and output the denoised polarization image.

2. The polarization image denoising method based on global cooperative transform BM3D filtering according to claim 1, characterized in that, S3 specifically includes: S31: The image is segmented into blocks, and the image block similarity evaluation is optimized based on the noise power spectral density. Similar image blocks are selected to construct an image block set. S32: Perform a three-dimensional cooperative transformation on the image patch set to obtain the transform domain spectral coefficients; S33: Attenuate and shrink the transform domain spectral coefficients according to the noise power spectral density; S34: Perform inverse transformation and weighted aggregation on the attenuated spectral coefficients to obtain a denoised image.

3. The polarization image denoising method based on global cooperative transform BM3D filtering according to claim 2, characterized in that, In step S31, optimizing the image patch similarity assessment based on the noise power spectral density specifically involves: In the calculation of image block matching distance, a bias correction term caused by spatial correlation noise is introduced to eliminate the interference of correlation noise on similarity evaluation and obtain unbiased block matching distance.

4. The polarization image denoising method based on global cooperative transform BM3D filtering according to claim 2, characterized in that, In S33, the attenuation and shrinkage processing of the transform domain spectral coefficients includes a hard threshold filtering stage and a Wiener filtering stage. In the hard threshold filtering stage, spectral coefficients below a preset threshold are set to zero. In the Wiener filtering stage, the denoised image output from the hard threshold filtering stage is used as the prior. The Wiener filtering coefficients are calculated based on the prior signal spectrum and the noise power spectral density, and the spectral coefficients are weighted and attenuated.

5. The polarization image denoising method based on global cooperative transform BM3D filtering according to claim 4, characterized in that, In step S3, a global Fourier threshold re-filtering step is included before outputting the denoised polarization image. Specifically: A global Fourier transform is performed on the residual image after preliminary denoising to obtain the residual spectrum; The residual spectrum is compared with the noise power spectral density to identify the spectral components corresponding to the over-smoothed image details; Based on the test results, a re-filtered input image is constructed, and the global collaborative transformation BM3D filtering process is performed again to preserve the polarization feature structure information.

6. The polarization image denoising method based on global cooperative transform BM3D filtering according to claim 5, characterized in that, The global Fourier threshold re-filtering step is executed once in the hard threshold filtering stage and once in the Wiener filtering stage. After performing re-filtering in the hard thresholding stage, the output is used as the prior image for the Wiener filtering stage; after performing re-filtering in the Wiener filtering stage, the final denoised polarization image is output.

7. The polarization image denoising method based on global cooperative transform BM3D filtering according to claim 1, characterized in that, In step S2, determining the noise power spectral density of the polarization image to be processed specifically involves: The spatially correlated noise caused by aerosol scattering is modeled as stationary correlated noise, and the power spectral density of the corresponding noise is obtained by Fourier transforming the spatial correlation convolution kernel of the noise.

8. A polarization image denoising system based on global cooperative transform BM3D filtering, characterized in that, The system is used to implement the polarization image denoising method based on global cooperative transform BM3D filtering as described in any one of claims 1-7, including: Image acquisition module: used to acquire the polarization angle image and polarization degree image corresponding to the polarization image to be processed; Noise modeling module: used to determine the noise power spectral density of the polarization image to be processed, wherein the noise includes spatially correlated noise caused by aerosol scattering and random noise from the imaging sensor; Collaborative denoising module: Based on the noise power spectral density, it uses the global collaborative transformation BM3D filtering algorithm to perform collaborative denoising on the polarization angle image and polarization degree image respectively, and outputs the denoised polarization image.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the polarization image denoising method based on global cooperative transformation BM3D filtering according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes the polarization image denoising method based on global cooperative transformation BM3D filtering as described in any one of claims 1-7.

Images