Near-field electromagnetic wave imaging multi-modal noise cooperative suppression method

Abstract

The application discloses a near-field electromagnetic wave imaging multi-modal noise cooperative suppression method, relates to the technical field of electromagnetic wave imaging, and aims to solve the problems of poor suppression effect of spot noise, stripe noise and Gaussian mixed noise and easy loss of target structure in the prior art. The method adopts a three-stage cooperative scheme of "complex domain spot suppression-multi-direction stripe separation-cross-channel Gaussian suppression", first removes the spot noise by adaptively processing the amplitude and phase of the complex domain image through a threshold value; then eliminates the stripe noise by performing horizontal / vertical ADOM filtering on the real part and the imaginary part of the complex domain, respectively; finally, combines BM3D filtering and three-dimensional transform domain optimization to suppress Gaussian noise, and improves the robustness through multi-frequency point fusion. Experiments show that the SSIM of the method is improved by 21.7% compared with the traditional method, the GSSIM and PSNR are optimal, the target structure features can be reserved in a strong noise environment, and the method is suitable for defect detection and other scenes of near-field electromagnetic wave synthetic aperture imaging.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic wave imaging technology, specifically to a method for collaborative suppression of multimodal noise in near-field electromagnetic wave imaging. Background Technology

[0002] Electromagnetic wave imaging, due to its unique physical properties, can penetrate various media and provide high-resolution imaging results. However, in practical applications, electromagnetic wave imaging systems are often affected by various noises. These noises not only severely interfere with the clarity of the imaging images but also affect subsequent tasks such as image processing, target detection, and recognition. Especially in near-field electromagnetic wave imaging, the types of noise caused by system and environmental factors are complex, and the presence of noise limits the accuracy of the imaging results. Therefore, how to effectively remove noise while maintaining the structural features of the image is an important task in the research of near-field electromagnetic wave imaging technology.

[0003] Existing near-field electromagnetic wave imaging denoising techniques mainly employ traditional filtering methods (such as mean filtering, median filtering, and wavelet thresholding) and end-to-end image enhancement models based on deep learning (such as convolutional neural networks and generative adversarial networks). These methods can suppress Gaussian noise, salt-and-pepper noise, and system thermal noise to some extent, but they generally have limitations: traditional methods tend to over-smooth details, leading to the loss of edge and weak target features; while deep learning models rely on a large amount of labeled data, which limits their generalization ability in real-world scenarios, and they are not effective in suppressing non-uniform background interference, multipath effects, and phase noise introduced by complex environments; in addition, most algorithms do not fully consider the physical characteristics of electromagnetic wave imaging, making it difficult to maintain a balance between structural fidelity and signal-to-noise ratio improvement under strong noise, thus restricting imaging accuracy and subsequent automatic recognition performance. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for collaborative suppression of multimodal noise in near-field electromagnetic wave imaging, which solves the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides a method for collaborative suppression of multimodal noise in near-field electromagnetic wave imaging, comprising the following steps:

[0006] S1. Complex domain speckle noise suppression, as detailed below:

[0007] S101. Acquire the raw near-field electromagnetic wave imaging image containing noise. ,in Represents the field of complex numbers. Separate the original near-field electromagnetic wave imaging image according to the image pixel dimension. amplitude components With phase components ;

[0008] S102, Based on the amplitude component global mean Compared with global standard deviation The amplitude adaptive threshold is constructed as follows:

[0009]

[0010] Based on the phase components global mean Compared with global standard deviation The phase-adaptive threshold is constructed as follows:

[0011]

[0012] S103. Adjust the amplitude component according to the adaptive threshold. Phase components Nonlinear suppression processing was performed separately to obtain the amplitude of the speckle noise after suppression. With phase ;

[0013] S104, based on and Reconstructing complex domain images ,in The imaginary unit;

[0014] S2. Complex domain fringe noise suppression, as detailed below:

[0015] S201, to real part Normalization to Perform horizontal ADOM filtering to obtain Vertical ADOM filtering is obtained The ADOM filter satisfies:

[0016]

[0017]

[0018] in For directional differential operators, , These are the gradients in the horizontal and vertical directions, respectively. For regularization parameters;

[0019] S202, to imaginary part Normalized to [0, 255], the horizontal imaginary part filtering result is obtained using the ADOM filtering method described above. Filtering results of the imaginary part in the vertical direction ;

[0020] S203, Reconstructed complex domain image after stripe noise suppression ;

[0021] S3. Cross-channel Gaussian noise suppression, as detailed below:

[0022] S301, to real part implement Filtering ,right imaginary part Perform BM3D filtering to obtain ,in This indicates a block matching 3D filtering operation;

[0023] S302, to and Perform three-dimensional wavelet packet transform on the union of the sets. Through adaptive threshold operator After processing, it undergoes inverse three-dimensional wavelet packet transform. Obtain the final output image ,Right now ;

[0024] S4. Multi-frequency point fusion optimization, as detailed below:

[0025] For images processed by steps S1 to S3 at different frequency points, phase alignment and amplitude superposition of coherent fusion and pixel-level weighted averaging of incoherent fusion are performed respectively to achieve comprehensive suppression of multimodal noise.

[0026] Optionally, the specific formula for the nonlinear suppression process in step S1 is as follows:

[0027]

[0028]

[0029] in These are the pixel coordinates of the image.

[0030] Optionally, the specific formula for the ADOM filtering of the imaginary part in step S2 is as follows:

[0031]

[0032]

[0033] in , These represent the imaginary parts of the image in the horizontal and vertical directions, respectively. For directional differential operators, and These represent the gradients in the horizontal and vertical directions, respectively. This is the regularization parameter.

[0034] Optionally, the BM3D filtering process in step S3 includes: performing block matching and grouping on the real and imaginary parts respectively, performing 3D transformation and thresholding on each group of blocks, and then obtaining the result through inverse 3D transformation and block recombination. and The adaptive threshold operator threshold Adaptively adjusts based on image noise intensity; the higher the noise intensity, the more... The larger the value, the better.

[0035] Optionally, the specific method for multi-frequency point fusion optimization in step S4 is as follows: for high-frequency images such as 30GHz and 40GHz, the least squares phase alignment algorithm is used for coherent fusion, and a weighting coefficient based on image sharpness is used for incoherent fusion, with higher-sharp pixels having greater weights, and the weighting coefficient range being [range missing]. .

[0036] This invention provides a method for collaborative suppression of multimodal noise in near-field electromagnetic wave imaging, which has the following beneficial effects:

[0037] This near-field electromagnetic wave imaging multimodal noise collaborative suppression method, through a three-stage divide-and-conquer strategy, designs dedicated suppression modules for multiplicative speckle noise, structural stripe noise and additive Gaussian noise respectively, which can simultaneously solve the problem of coexistence of multiple types of noise in near-field electromagnetic wave imaging and fill the technical gap in mixed noise collaborative processing.

[0038] Furthermore, this method, through complex domain component processing, ADOM directional filtering, and three-dimensional transform domain optimization, strictly preserves the target scattering information while denoising: for stainless steel sheet spot noise processing, SSIM reaches 0.6046; for quartz block stripe noise processing, GSSIM reaches 0.5370, significantly better than Fourier filtering (0.4556) and TV method (0.3620), ensuring the accuracy of subsequent target detection and recognition;

[0039] Meanwhile, this method does not require labeled data and constructs a cross-band noise coupling model through multi-frequency point coherent / incoherent fusion: in the 40GHz high-frequency scenario, the GSSIM retention rate reaches 90.4%, which is higher than the 87.9% of Fourier filtering; for mixed noise processing, the MSE is only 0.0028, which is significantly lower than Fourier filtering (0.0047), and can effectively adapt to the complex noise environment of electromagnetic wave high-frequency imaging;

[0040] This method outperforms traditional methods in all key quantitative indicators: in speckle noise processing, although the PSNR is 13.21dB, it achieves a significant improvement in structural fidelity at a slight cost; in stripe noise processing, the PSNR reaches 26.77dB; in mixed noise processing, the PSNR reaches 25.52dB, which is much higher than the TV method (8.60dB) and the LRHP method (9.09dB). Its stable performance can be directly applied to engineering scenarios such as industrial non-destructive testing and security imaging, providing key support for the practical application of near-field electromagnetic wave imaging technology. Attached Figure Description

[0041] Figure 1 The image shows the noise reduction results for the stainless steel sheet in this invention; (a) median filtering result, SSIM=0.4963, PSNR=13.84, MSE=0.0413; (b) Frost filtering result, SSIM=0.4053, PSNR=14.50, MSE=0.0355; (c) NLMeans filtering result, SSIM=0.4704, PSNR=13.58, MSE=0.0439; (d) multimodal noise collaborative filtering result, SSIM=0.6046, PSNR=13.21, MSE=0.0478.

[0042] Figure 2 The images show the quartz cube denoising noise removal results in this invention; where (a) Fourier transform filtering results, GSSIM=0.4556, PSNR=23.46, MSE=0.0045; (b) Total variation denoising filtering results, GSSIM=0.3620, PSNR=10.71, MSE=0.0849; (c) Low-rank matrix results, GSSIM=0.2924, PSNR=9.66, MSE=0.1082; (d) Multimodal noise collaborative filtering results, GSSIM=0.5370, PSNR=26.77, MSE=0.0021.

[0043] Figure 3 The results show the combined denoising effect of quartz block speckle noise and stripe noise in this invention; where (a) Fourier transform filtering result, GSSIM=0.4005, PSNR=23.24, MSE=0.0047; (b) Total variation denoising filtering result, GSSIM=0.2586, PSNR=8.60, MSE=0.1380; (c) Low-rank matrix filtering result, GSSIM=0.2566, PSNR=9.09, MSE=0.1233; (d) Multimodal noise collaborative filtering result, GSSIM=0.4856, PSNR=25.52, MSE=0.0028. Detailed Implementation

[0044] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0045] This invention provides a technical solution: a method for collaborative suppression of multimodal noise in near-field electromagnetic wave imaging, comprising the following steps:

[0046] S1. Complex domain speckle noise suppression, as detailed below:

[0047] S101. Acquire the raw near-field electromagnetic wave imaging image containing noise. ,in Represents the field of complex numbers. Separate the original near-field electromagnetic wave imaging image according to the image pixel dimension. amplitude components With phase components ;

[0048] S102, Based on the amplitude component global mean Compared with global standard deviation The amplitude adaptive threshold is constructed as follows:

[0049]

[0050] Based on the phase components global mean Compared with global standard deviation The phase-adaptive threshold is constructed as follows:

[0051]

[0052] S103. Adjust the amplitude component according to the adaptive threshold. Phase components Nonlinear suppression processing was performed separately to obtain the amplitude of the speckle noise after suppression. With phase ;

[0053] S104, based on and Reconstructing complex domain images ,in The imaginary unit;

[0054] The specific formula for nonlinear suppression is as follows:

[0055]

[0056]

[0057] in These are the pixel coordinates of the image.

[0058] S2. Complex domain fringe noise suppression, as detailed below:

[0059] S201, to real part Normalization to Perform horizontal ADOM filtering to obtain Vertical ADOM filtering is obtained The ADOM filter satisfies:

[0060]

[0061]

[0062] in For directional differential operators, , These are the gradients in the horizontal and vertical directions, respectively. For regularization parameters;

[0063] S202, to imaginary part Normalized to [0, 255], the horizontal imaginary part filtering result is obtained using the ADOM filtering method described above. Filtering results of the imaginary part in the vertical direction ;

[0064] S203, Reconstructed complex domain image after stripe noise suppression ;

[0065] The specific formula for the imaginary part ADOM filtering is:

[0066]

[0067]

[0068] in , These represent the imaginary parts of the image in the horizontal and vertical directions, respectively. For directional differential operators, and These represent the gradients in the horizontal and vertical directions, respectively. For regularization parameters;

[0069] S3. Cross-channel Gaussian noise suppression, as detailed below:

[0070] S301, to real part implement Filtering ,right imaginary part Perform BM3D filtering to obtain ,in This indicates a block matching 3D filtering operation;

[0071] S302, to and Perform three-dimensional wavelet packet transform on the union of the sets. Through adaptive threshold operator After processing, it undergoes inverse three-dimensional wavelet packet transform. Obtain the final output image ,Right now ;

[0072] The execution process of BM3D filtering includes: performing block matching and grouping on the real and imaginary parts respectively, performing 3D transformation and thresholding on each group of blocks, and then obtaining the final product through inverse 3D transformation and block recombination. and The adaptive threshold operator threshold Adaptively adjusts based on image noise intensity; the higher the noise intensity, the more... The larger the value;

[0073] S4. Multi-frequency point fusion optimization, as detailed below:

[0074] For images processed by steps S1 to S3 at different frequency points, phase alignment and amplitude superposition of coherent fusion and pixel-level weighted averaging of incoherent fusion are performed respectively to achieve comprehensive suppression of multimodal noise.

[0075] The specific method for multi-frequency fusion optimization is as follows: For images in high-frequency bands such as 30GHz and 40GHz, the least squares phase alignment algorithm is used for coherent fusion, and a weighting coefficient based on image sharpness is used for incoherent fusion. Pixels with higher sharpness have greater weights, and the range of the weighting coefficients is as follows: .

[0076] Example

[0077] Noise reduction experiments were conducted on stainless steel test pieces with varying amounts of speckle noise. Images processed by the algorithm without speckle noise were used as clean reference images to facilitate the calculation of image quality evaluation metrics after denoising. Noise reduction processing was performed on images of stainless steel pieces with dense speckle noise distribution, and the results are as follows: Figure 1As shown, while median filtering achieves a PSNR of 13.84 dB, superior to other traditional methods, its SSIM value of 0.4963 reveals severe information loss, particularly the generation of vertical stripe noise not present in the original image. Frost filtering, while achieving relatively optimal PSNR (14.50 dB) and MSE (0.0355), has the lowest SSIM value (0.4053) of all methods, reflecting the large-area artifacts produced by this algorithm under strong noise interference and its inability to completely eliminate speckle noise. The NLMeans method, in extreme... The performance of the method deteriorated significantly under high noise conditions, with both PSNR (13.58dB) and MSE (0.0439) being worse than the baseline method, highlighting the failure of nonlocal similarity measurement in high noise scenarios. Although the method of this invention makes a slight compromise on MSE (0.0478) and PSNR (13.21dB), the SSIM value (0.6046) is 21.7% higher than the suboptimal method, which proves that the multimodal cooperative mechanism has significant advantages in maintaining the structure of the target region and removing speckle noise, and is more suitable for the defect detection needs in high noise environments.

[0078] To verify that the multimodal noise collaborative suppression method of this invention can effectively remove fringe noise in near-field electromagnetic wave synthetic aperture imaging experiments, this section conducts a denoising experiment on a quartz cube fringe noise image. Since fringe noise is prevalent in all frequency bands of electromagnetic waves, there is no clean reference image free of fringe noise. Here, an image quality evaluation index—Gradient Structural Similarity Index Measure (GSSIM)—is used, which is an improvement on SSIM based on the characteristics of fringe noise. The value of GSSIM is between 0 and 1, and the closer the value is to 1, the better the denoising effect and the more detail information is preserved. It should be noted that if the denoising effect on the fringe noise image is poor, the value of GSSIM may also be close to 0.

[0079] Figure 2This paper compares the restoration effects of different denoising algorithms on quartz cube structures in near-field electromagnetic wave imaging. Quantitative analysis shows that the multimodal collaborative filtering method significantly outperforms traditional methods in both GSSIM (0.5370) and PSNR (26.77dB), with its MSE (0.0021) being only 46.7% of that of the Fourier filter, indicating that this method effectively preserves the target structural features while suppressing fringe noise. While the Fourier filter performs second best in PSNR (23.46dB), its GSSIM (0.4556) reveals that spectral filtering may lead to loss of texture details, especially limiting its generalization ability in non-periodic noise scenarios. The low GSSIM (0.3620) and high MSE (0.0849) of the TV method reflect its over-smoothing characteristics, resulting in severe degradation of high-frequency information, verifying the limitations of variational methods in processing complex fringe noise. Although the LRHP method removes some fringe noise, the background noise is too obvious.

[0080] Figure 3 The denoising performance of mixed noise (speckle noise + stripe noise) under electromagnetic near-field imaging is compared. Multimodal collaborative filtering maintains the best GSSIM (0.4856) and PSNR (25.52 dB), with its MSE (0.0028) being 40.4% lower than that of Fourier filtering, confirming the robustness of the algorithm in higher frequency noise-coupled scenarios. The GSSIM (0.4005) of Fourier filtering decreases by 12.1% compared to the 30 GHz experiment, revealing its inherent limitations in suppressing aperiodic speckle noise. The GSSIM (0.2586 / 0.2566) of the TV and LRHP methods are significantly degraded. The PSNR being below 10dB and MSE exceeding 0.12 indicates that traditional variational methods are prone to oversmoothing and artifact accumulation under mixed noise. Notably, the multimodal method achieves a GSSIM retention rate of 90.4% at 40GHz (relative to the 30GHz benchmark), while Fourier filtering only maintains 87.9%, highlighting the generalization ability of its constructed cross-band noise coupling model. Experiments demonstrate that as the noise complexity increases with frequency, the multimodal method can effectively decouple the mutual interference of noises from different mechanisms by jointly optimizing frequency domain attenuation and spatial sparsity constraints, providing a reliable denoising solution for high-frequency electromagnetic wave imaging.

[0081] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for collaborative suppression of multimodal noise in near-field electromagnetic wave imaging, characterized in that, Includes the following steps: S1. Complex domain speckle noise suppression, as detailed below: S101. Acquire the raw near-field electromagnetic wave imaging image containing noise. ,in Represents the field of complex numbers. Separate the original near-field electromagnetic wave imaging image according to the image pixel dimension. amplitude components With phase components ; S102, Based on the amplitude component global mean Compared with global standard deviation The amplitude adaptive threshold is constructed as follows: ； Based on the phase components global mean Compared with global standard deviation The phase-adaptive threshold is constructed as follows: ； S103. Adjust the amplitude component according to the adaptive threshold. Phase components Nonlinear suppression processing was performed separately to obtain the amplitude of the speckle noise after suppression. With phase ; S104, based on and Reconstructing complex domain images ,in The imaginary unit; S2. Complex domain fringe noise suppression, as detailed below: S201, to real part Normalization to Perform horizontal ADOM filtering to obtain Vertical ADOM filtering is obtained The ADOM filter satisfies: ； ； in For directional differential operators, , These are the gradients in the horizontal and vertical directions, respectively. For regularization parameters; S202, to imaginary part Normalized to [0, 255], the horizontal imaginary part filtering result is obtained using the ADOM filtering method described above. Filtering results of the imaginary part in the vertical direction ; S203, Reconstructed complex domain image after stripe noise suppression ; S3. Cross-channel Gaussian noise suppression, as detailed below: S301, to real part implement Filtering ,right imaginary part Performing BM3D filtering yields ,in This indicates a block matching 3D filtering operation; S302, to and Perform three-dimensional wavelet packet transform on the union of the sets. Through adaptive threshold operator After processing, it undergoes inverse three-dimensional wavelet packet transform. Obtain the final output image ,Right now ; S4. Multi-frequency point fusion optimization, as detailed below: For images processed by steps S1 to S3 at different frequency points, phase alignment and amplitude superposition of coherent fusion and pixel-level weighted averaging of incoherent fusion are performed respectively to achieve comprehensive suppression of multimodal noise.

2. The method for coordinated suppression of multimodal noise in near-field electromagnetic wave imaging according to claim 1, characterized in that, The specific formula for the nonlinear suppression process in step S1 is as follows: ； ； in These are the pixel coordinates of the image.

3. The method for collaborative suppression of multimodal noise in near-field electromagnetic wave imaging according to claim 1, characterized in that, The specific formula for ADOM filtering of the imaginary part in step S2 is as follows: ； ； in , These represent the imaginary parts of the image in the horizontal and vertical directions, respectively. For directional differential operators, and These represent the gradients in the horizontal and vertical directions, respectively. This is the regularization parameter.

4. The method for coordinated suppression of multimodal noise in near-field electromagnetic wave imaging according to claim 1, characterized in that, The execution process of BM3D filtering in step S3 includes: performing block matching and grouping on the real and imaginary parts respectively, performing 3D transformation and thresholding on each group of blocks, and then obtaining the result through inverse 3D transformation and block recombination. and The adaptive threshold operator threshold Adaptively adjusts based on image noise intensity; the higher the noise intensity, the more... The larger the value, the better.

5. The method for coordinated suppression of multimodal noise in near-field electromagnetic wave imaging according to claim 1, characterized in that, The specific method for multi-frequency point fusion optimization in step S4 is as follows: For high-frequency images, the least squares phase alignment algorithm is used for coherent fusion, and a weighting coefficient based on image sharpness is used for incoherent fusion. Pixels with higher sharpness have greater weights, and the range of the weighting coefficients is as follows: .

Images