Sonar image denoising method of adaptive threshold and improved guide filter

Abstract

The invention discloses a sonar image denoising method of an adaptive threshold and an improved guide filter. According to the method, a noise-containing sonar image is decomposed into a high-frequency coefficient and a low-frequency coefficient through non-subsampled shearlet transform (NSST). Aiming at the noise characteristics in the high-frequency sub-band, adopting a self-adaptive threshold algorithm to complete de-noising; for a low-frequency component, a guide filter introducing peak perception weight and scale constraint is adopted, so that edge blur is avoided while background noise is smoothed. And finally, reconstructing high and low frequency coefficients through NSST inverse transformation to obtain a de-noised image. According to the method, the visual effect of the image is improved, and a good foundation is laid for detection and recognition of the image target in the next step.

Description

Technical Field

[0001] The invention belongs to the technical field of sonar image processing, and in particular relates to a sonar image denoising method using an adaptive threshold and an improved guided filter. Background Art

[0002] Sonar imaging has important applications in ocean exploration and underwater target identification. However, due to the complexity of the ocean environment, the imaging process is susceptible to interference from seafloor reverberation, ambient noise, and equipment self-noise. Multiplicative speckle noise (speckle noise) is the most significant. Speckle noise appears as randomly distributed granular interference in the image, severely reducing image contrast, blurring details and distorting target outlines. This directly impacts subsequent image segmentation, feature extraction, and target recognition.

[0003] Traditional sonar image denoising methods mainly include spatial domain filtering and frequency domain transformation. Spatial domain filtering (such as Lee filter, Frost filter, mean filter, etc.) achieves noise suppression by adjusting the window size, but it can easily lead to blurred edges in non-uniform areas, making it difficult to balance the requirements of denoising and detail preservation. Frequency domain methods (such as wavelet transform) improve denoising performance through multi-scale decomposition, but their basis functions lack anisotropy and are insufficient in representing linear singularities in the image (such as edges and textures), making it difficult to effectively preserve key structural information in sonar images. In recent years, although deep learning-based denoising methods have performed well in image processing, they rely on large-scale training data and high-performance computing resources, and cannot meet the strict real-time, low-power, and embedded deployment requirements of underwater equipment (such as autonomous underwater vehicles (AUVs)).

[0004] In summary, the existing technology has the following defects:

[0005] 1. The contradiction between noise suppression and edge preservation: Traditional filtering algorithms have good denoising effects in uniform areas, but they are prone to causing edge blur and loss of texture details in complex scenes;

[0006] 2. Limitations of frequency domain methods: Frequency domain techniques such as wavelet transform are not adaptable enough to multi-dimensional features (such as direction and scale), making it difficult to balance noise removal and structural integrity.

[0007] 3. Insufficient applicability of deep learning: It relies on large amounts of labeled data and high computing power, making it difficult to achieve efficient processing in resource-constrained underwater environments;

[0008] 4. Inaccurate speckle noise modeling: Existing methods simplify the modeling of the statistical characteristics of sonar image speckle noise (such as gamma distribution), resulting in insufficient robustness of the denoising algorithm.

[0009] Therefore, there is an urgent need for a sonar image processing method that can efficiently denoise in complex noise environments while retaining image edges and details to meet the actual needs of underwater detection and target recognition. Summary of the Invention

[0010] In order to overcome the problems in the background technology, the present invention provides a sonar image denoising method with an adaptive threshold and an improved guided filter.

[0011] To achieve the above object, the present invention is implemented through the following technical solutions:

[0012] A sonar image denoising method using an adaptive threshold and an improved guided filter comprises the following steps:

[0013] S1: Generate multiplicative speckle noise that obeys the gamma distribution according to the noise distribution characteristics of the sonar image and superimpose it on the original sonar image;

[0014] S2: Use non-subsampled shearlet transform (NSST) to perform multi-scale decomposition on the noisy sonar image to obtain multiple high-frequency sub-band images and one low-frequency sub-band image;

[0015] S3: performing noise reduction processing on the high-frequency sub-band image based on a directional weight factor and an adaptive threshold function, wherein the adaptive threshold function dynamically adjusts the threshold by combining the noise standard deviation and the signal standard deviation;

[0016] S4: performing smoothing on the low-frequency sub-band image by using an improved guided filter that introduces peak perception weight and multi-scale constraints;

[0017] S5: The processed high-frequency sub-band and low-frequency sub-band are reconstructed through the inverse NSST transform to obtain the denoised sonar image.

[0018] Furthermore, the step S1 specifically includes:

[0019] (1) Generate a complex gamma random field, whose imaginary and real parts are independent gamma distributed random variables with the same mean and standard deviation;

[0020] (2) filtering the complex gamma random field by a 3×3 window low-pass averaging method to eliminate the correlation between adjacent noise fields;

[0021] (3) Multiply the filtered noise field with the original sonar image to generate an image containing gamma-distributed speckle noise.

[0022] Furthermore, the specific steps of NSST decomposition in step S2 are:

[0023] (1) Using non-subsampling pyramid (NSP) to decompose the noisy image into high-frequency components and low-frequency components;

[0024] (2) Constructing a Meyer window for the high-frequency component to perform directional localization and obtain multi-directional sub-bands;

[0025] (3) Perform inverse Fourier transform on each directional subband to obtain the non-subsampled shearlet coefficients.

[0026] Furthermore, the adaptive threshold function in step S3 is defined as:

[0027]

[0028] Among them, λ is a custom parameter, σ is the standard deviation of the noise, and σ x is the standard deviation of the signal.

[0029]

[0030] The direction weight factor α is further introduced and expressed as:

[0031]

[0032] The corrected threshold function is:

[0033]

[0034] Furthermore, the loss function of the improved guided filter in step S4 is:

[0035]

[0036] in, is σ k 2 The mean of ε0 is a positive constant.

[0037] Calculate a according to the new loss function k and b k , we can get:

[0038]

[0039] Furthermore, the specific process of the NSST inverse transformation in step S5 is as follows:

[0040] (1) Performing directional localized inverse transformation on the high-frequency sub-band after noise reduction;

[0041] (2) Fusion of high-frequency and low-frequency subbands via non-subsampled inverse pyramid transform;

[0042] (3) Output the reconstructed noise-reduced image.

[0043] Furthermore, the mean and variance of the gamma distribution are adjustable to simulate speckle noise of different intensities.

[0044] Beneficial effects of the present invention:

[0045] 1. Efficient noise suppression and edge detail preservation:

[0046] The sonar image is decomposed into multiple scales by using the non-subsampled shearlet transform (NSST). By utilizing its multi-directional and multi-resolution characteristics, the high-frequency details (such as edges and textures) and low-frequency structural information in the image are effectively captured, avoiding the detail loss problem caused by the isotropy of the basis function in traditional wavelet transform.

[0047] For high-frequency sub-band noise, an adaptive threshold function is used Combining the noise standard deviation (σ) and the signal standard deviation (σ x ) dynamically adjusts the threshold and introduces a directional weight factor (α) to optimize the threshold correction according to the mean difference of sub-band coefficients, significantly improving the targeted noise suppression while reducing the false deletion of valid signals.

[0048] 2. High noise modeling accuracy:

[0049] By generating multiplicative speckle noise that obeys the gamma distribution, the statistical characteristics of speckle noise in sonar images (such as long-tail distribution) are simulated, and a 3×3 window low-pass averaging method is used to eliminate the correlation of the noise field, making the noise model closer to the actual scene, thereby improving the robustness and adaptability of the denoising algorithm.

[0050] The mean and variance of the gamma distribution are adjustable, supporting the simulation and denoising verification of noises of different intensities, thus enhancing the versatility of the method.

[0051] 3. Computational efficiency and practicality:

[0052] The NSST decomposition and reconstruction framework is adopted to avoid the information loss caused by traditional downsampling operations. At the same time, the calculation process is optimized through direction localization and inverse Fourier transform, which significantly reduces the algorithm complexity.

[0053] Improve the loss function design of the guided filter (f(a k ,b k ))Through the parameters ε and γ k 2 Dynamic adjustment reduces the number of iterations and improves real-time processing capabilities, making it suitable for deployment of underwater equipment with limited resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0055] Figure 1 is a flow chart of the present invention;

[0056] Figure 2 It is a schematic diagram of NSST decomposition of the present invention;

[0057] Figure 3 is the original sonar image of the present invention;

[0058] Figure 4 is a noisy sonar image with a variance of 0.8 added according to the present invention;

[0059] Figure 5 This is the denoised image of the present invention. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0061] Example 1

[0062] Combine Figure 1 This embodiment describes a sonar image denoising method using an adaptive threshold and an improved guided filter, and the method specifically includes the following steps:

[0063] S1: Generate multiplicative speckle noise that obeys the gamma distribution according to the noise distribution characteristics of the sonar image and superimpose it on the original sonar image;

[0064] S2: Use non-subsampled shearlet transform (NSST) to perform multi-scale decomposition on the noisy sonar image to obtain multiple high-frequency sub-band images and one low-frequency sub-band image;

[0065] S3: performing noise reduction processing on the high-frequency sub-band image based on a directional weight factor and an adaptive threshold function, wherein the adaptive threshold function dynamically adjusts the threshold by combining the noise standard deviation and the signal standard deviation;

[0066] S4: performing smoothing on the low-frequency sub-band image by using an improved guided filter that introduces peak perception weight and multi-scale constraints;

[0067] S5: The processed high-frequency sub-band and low-frequency sub-band are reconstructed through the inverse NSST transform to obtain the denoised sonar image.

[0068] In this embodiment, step S1 specifically includes:

[0069] According to the noise distribution characteristics of the sonar image, speckle noise that obeys the gamma distribution is added to the sonar image. The specific operation is as follows:

[0070] (1) Generate a complex gamma random field. The imaginary and real parts of the random field are independent random variables. They all obey the gamma distribution with the same mean and standard deviation. By changing the mean and variance values, different degrees of speckle noise can be obtained.

[0071] (2) Since there is a certain correlation between adjacent fields of speckle noise, the random field is filtered using the low-pass averaging method. The selected window size is 3×3, and the output amplitude is obtained after filtering.

[0072] (3) Multiply the generated noise with the original sonar image to output an image with speckle noise that follows a gamma distribution.

[0073] In this embodiment, step S2 specifically includes:

[0074] NSST is broken down into the following steps:

[0075] (1) Multi-scale decomposition: Non-subsampled pyramid (NSP) is used to decompose the noisy image into high-frequency and low-frequency components;

[0076] (2) Directional localization: Construct a Meyer window for the high-frequency image and perform directional localization to obtain subbands in different directions;

[0077] (3) Perform inverse Fourier transform on the subband in each direction to obtain the non-subsampled shearlet coefficients.

[0078] When the dimension n = 2, the affine system with synthetic expansion can be expressed as:

[0079] M AB (ψ)={ψ j,l,k (x)=|det A| j / 2 ψ(B l A j (xk)):j∈Z,k∈Z 2}

[0080] Where ψ∈L 2 (R 2), A and B are both two-dimensional reversible matrices, and |det B| = 1; j and l are scale parameters and shear parameters respectively, and k is the translation parameter. If for any f∈L 2 (R 2 ), M AB (ψ) form a tight support frame, that is,

[0081] ∑ j,l,k | <f,ψ j,l,k >| 2 =||f|| 2

[0082] In L 2 (R 2 ) defines a special synthetic wavelet, namely the shearlet transform.

[0083] In this embodiment, step S3 specifically includes:

[0084] For the high-frequency part after NSST decomposition, an adaptive threshold is used to reduce the noise. First, the basic threshold is defined as follows:

[0085]

[0086] Among them, λ is a custom parameter used to adjust the threshold value. In the above formula, σ is the standard deviation of the noise, σ x is the standard deviation of the signal.

[0087]

[0088] On this basis, the direction weight factor α is introduced, which is expressed as:

[0089]

[0090] Among them, μ(D i,j ) is the mean of the high-pass coefficients of different numbers of different layers after NSST decomposition: μ(D i ) is the mean of the high-pass coefficients of different numbers on the same layer.

[0091] Finally, the improved threshold function formula adopted by the present invention is as follows:

[0092]

[0093] In this embodiment, step S3 specifically includes:

[0094] For the low-frequency image after NSST decomposition, a guided filter with peak weight and multi-scale constraint is used to process it. After introducing peak perception weight and multi-scale constraint, a new loss function is obtained, which is expressed as:

[0095]

[0096] in, is σ k 2 The mean of ε0 is a positive constant.

[0097] Calculate a according to the new loss function k and b k , we can get:

[0098]

[0099] In this embodiment, step S5 specifically includes: performing an inverse NSST transform on the processed high-frequency sub-band and low-frequency sub-band to obtain a final noise-reduced image.

[0100] Example 2

[0101] See Figures 1 to 5 As shown, a sonar image denoising method using an adaptive threshold and an improved guided filter in this embodiment first decomposes the noisy sonar image into high-frequency and low-frequency components using a non-subsampled shearlet transform. The high-frequency component is processed using an adaptive threshold; the low-frequency component is processed using an improved guided filter. Finally, the denoised sonar image is obtained using an inverse NSST transform.

[0102] by Figure 3 Take the denoising experiment as an example, the specific process is as follows:

[0103] S1: Yes Figure 3 Adding speckle noise with a variance of 0.8 gives Figure 4 ;

[0104] S2: Using NSST Figure 4 Decompose to obtain high-frequency and low-frequency parts;

[0105] S3: Adopt the adaptive threshold method to reduce the noise of the high frequency part;

[0106] S4: Use the improved guided filter to process the decomposed low-frequency part;

[0107] S5: Perform NSST inverse transformation on the image processed in the above two steps to obtain the final denoised image, that is, Figure 5 .

[0108] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A sonar image denoising method using an adaptive threshold and an improved guided filter, characterized in that: The following steps are involved: S1: Generate multiplicative speckle noise that obeys the gamma distribution according to the noise distribution characteristics of the sonar image and superimpose it on the original sonar image; S2: Use non-subsampled shearlet transform (NSST) to perform multi-scale decomposition on the noisy sonar image to obtain multiple high-frequency sub-band images and one low-frequency sub-band image; S3: performing noise reduction processing on the high-frequency sub-band image based on a directional weight factor and an adaptive threshold function, wherein the adaptive threshold function dynamically adjusts the threshold by combining the noise standard deviation and the signal standard deviation; S4: performing smoothing on the low-frequency sub-band image by using an improved guided filter that introduces peak perception weight and multi-scale constraints; S5: The processed high-frequency sub-band and low-frequency sub-band are reconstructed through the inverse NSST transform to obtain the denoised sonar image.

2. The sonar image denoising method using an adaptive threshold and an improved guided filter according to claim 1, characterized in that: The step S1 specifically includes: (1) Generate a complex gamma random field, whose imaginary and real parts are independent gamma distributed random variables with the same mean and standard deviation; (2) filtering the complex gamma random field by a 3×3 window low-pass averaging method to eliminate the correlation between adjacent noise fields; (3) Multiply the filtered noise field with the original sonar image to generate an image containing gamma-distributed speckle noise.

3. The sonar image denoising method using an adaptive threshold and an improved guided filter according to claim 1, characterized in that: The specific steps of NSST decomposition in step S2 are: (1) Using non-subsampling pyramid (NSP) to decompose the noisy image into high-frequency components and low-frequency components; (2) Constructing a Meyer window for the high-frequency component to perform directional localization and obtain multi-directional sub-bands; (3) Perform inverse Fourier transform on each directional subband to obtain the non-subsampled shearlet coefficients.

4. The sonar image denoising method using an adaptive threshold and an improved guided filter according to claim 1, characterized in that: The definition of the adaptive threshold function in step S3 is: Among them, λ is a custom parameter, σ is the standard deviation of the noise, and σ x is the standard deviation of the signal. The direction weight factor α is further introduced and expressed as: Among them, μ(D i,j ) is the mean of the high-pass coefficients of different numbers of different layers after NSST decomposition: μ(D i ) is the mean of the high-pass coefficients of different numbers on the same layer, and the modified threshold function is:

5. The sonar image denoising method using an adaptive threshold and an improved guided filter according to claim 1, characterized in that: The loss function of the improved guided filter in step S4 is: in, is σ k 2 The mean of ε0 is a positive constant. Calculate a according to the new loss function k and b k , we can get:

6. The sonar image denoising method using an adaptive threshold and an improved guided filter according to claim 1, characterized in that: The specific process of the NSST inverse transformation in step S5 is as follows: (1) Performing directional localized inverse transformation on the high-frequency sub-band after noise reduction; (2) Fusion of high-frequency and low-frequency subbands via non-subsampled inverse pyramid transform; (3) Output the reconstructed noise-reduced image.

7. The sonar image denoising method using an adaptive threshold and an improved guided filter according to any one of claims 1 to 6, characterized in that: The mean and variance of the gamma distribution are adjustable and are used to simulate speckle noise of different intensities.

Images