Sonar image segmentation method based on anisotropic residual noise reduction and sparse fuzzy clustering
Through the anisotropic residual noise reduction and sparse fuzzy clustering methods, the problem of low segmentation accuracy caused by noise interference in sonar images is solved, high-precision underwater target segmentation is achieved, and segmentation speed and robustness are improved.
Patent Information
- Application Number
- CN202411056662.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-12-06
AI Technical Summary
The poor imaging effect of sonar image is caused by blurring the target edge information and low segmentation accuracy. Especially in the target segmentation underwater sinking, it is difficult to achieve high-precision segmentation by commonly used image segmentation methods.
A sonar image segmentation method for anisotropic residual noise reduction and sparse fuzzy clustering is proposed. The noise is initially reduced through morphological reconstruction and residual noise reduction structure, and the KL divergence constraint and sparse membership function are introduced to extract sonar images and speed up the convergence of the objective function.
It significantly alleviates impulse noise and spot noise in sonar images, improves segmentation accuracy and speed, enhances the robustness and anti-interference ability of the algorithm, and is suitable for underwater target segmentation.
Smart Images

Figure CN119091137A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical fields of signal processing, digital image processing and machine learning, and in particular relates to a sonar image segmentation method of anisotropic residual denoising and sparse fuzzy clustering. Background Art
[0002] Sonar image target segmentation is one of the hot issues in the field of underwater acoustics and an important step in achieving target detection. In recent years, imaging sonar has become an important tool for obtaining underwater target information. It can provide underwater scene information in water environments with low visibility and intuitively express underwater conditions through two-dimensional images or three-dimensional point clouds. Therefore, it is widely used in underwater search and rescue, seabed detection and other detection activities.
[0003] Due to the influence of seabed reverberation and noise interference, the segmentation of bottom targets in sonar images has always been a difficult challenge. For sonar image segmentation, the commonly used methods mainly include clustering methods, active contour models, Markov random field segmentation, etc. Among them, fuzzy C-means (FCM), as a classic unsupervised machine learning algorithm, has been widely used in the fields of sonar images and optical images due to its simplicity and low complexity. The improvement of the FCM method focuses on the introduction of spatial information and the change of kernel distance. Considering the robustness and computational efficiency of the FCM method, many scholars have introduced signal processing or mathematical operations into the improvement of FCM in recent years to achieve a balance between segmentation accuracy and computational efficiency. Through investigation, the research on FCM methods in optical image noise environments has made certain progress. However, when acting on sonar images, the classic FCM algorithm is not robust to noise, especially when there are typical impulse noise and speckle noise in sonar images. Therefore, the use of sonar images is often accompanied by the study of preprocessing methods to reduce the impact of noise to a certain extent. Considering that sonar images often have the characteristics of low contrast and blurred target boundaries, traditional noise reduction methods will aggravate edge blur and have a certain impact on segmentation accuracy.
[0004] Conventional multi-beam forming methods use the output amplitude of the multi-element array when generating sonar images, and its size represents the size of the corresponding pixel value of the generated sonar image. At the same time, the beam image will have sidelobe interference. In many cases, the sonar image has low contrast and blurred boundaries. Especially when using imaging sonar to scan sinking targets, the targets are often submerged in the complex seabed background. Therefore, when applying the commonly used image segmentation methods to complex sonar image segmentation, the accuracy of segmentation is an urgent problem to be solved. Summary of the invention
[0005] The purpose of the present invention is to solve the existing problem of low segmentation accuracy caused by poor sonar image imaging effect and blurred target edge information, to achieve high-precision segmentation of underwater sinking targets, and to propose a sonar image segmentation method of anisotropic residual denoising and sparse fuzzy clustering. The method firstly reduces the Gaussian and pulse interference in the image through morphological reconstruction, and uses anisotropic denoising method to construct a residual denoising structure, which increases the amount of target information retained while reducing speckle noise; extracts the self-structural information of the processed sonar image, extracts features through two filters, and obtains two sets of sonar image feature sets; then, introduces KL divergence constraints to ensure that the current pixel value intensity is close to the average intensity of its neighborhood pixel values; finally, in order to speed up the convergence of the objective function, introduces a sparse membership function to increase the convergence speed, and improves the segmentation speed while ensuring the segmentation accuracy.
[0006] The present invention is implemented by the following technical scheme. The present invention proposes a sonar image segmentation method of anisotropic residual denoising and sparse fuzzy clustering, and the method comprises:
[0007] Step 1: For the collected sonar image data, denoising is first performed through morphological reconstruction and residual denoising structure to achieve preprocessing;
[0008] Step 2: Perform different filter transformations on the preprocessed sonar image to obtain two sub-images of the same size as the original image after transformation, and use the two sub-images as input;
[0009] Step 3: Introduce the KL divergence constraint in FCM to ensure that the intensity of the current pixel value is close to the average intensity of its neighboring pixel values, and ensure that the membership matrices of the two sub-images are as consistent as possible. , At the same time, in order to speed up the convergence of the objective function, a sparse membership function is introduced to increase the convergence speed.
[0010] Furthermore, in step 1, there are a lot of impulse noise and speckle noise in the sonar image. Due to its complex noise characteristics, the general spatial filtering method cannot take into account a variety of noise distributions. First, the noise in the sonar image is preliminarily processed through morphological reconstruction (MR), and the specific form is as follows:
[0011]
[0012]
[0013]
[0014] Defining Morphological Dilation Reconstruction f represents the original input image, g represents the marker image, g≤f, δ is the dilation operation, and ∧ represents the point-by-point minimum.
[0015]
[0016]
[0017]
[0018] Defining Morphological Corrosion Reconstruction g≥f, ε is the corrosion operation, and ∨ represents the point-by-point maximum value.
[0019] Morphological closed reconstruction is used as the initial impulse noise denoising method and the subsequent label reconstruction method. g = ε(f) and g = δ(f) are selected as the label images for dilation reconstruction and erosion reconstruction respectively. Based on the above operator, morphological closed reconstruction R C (f) is defined as:
[0020]
[0021] Furthermore, in step 1, the residual denoising structure is continued to be used to complete the denoising; first, the SRAD denoising method is used to denoise the sonar image, and the specific form is as follows:
[0022]
[0023] I(t=0)=I0
[0024] Among them, div represents the divergence operator, represents the gradient operator, represents the diffusion equation, and ||·|| represents the amplitude. t represents the introduced time operator, which can reflect the correlation between the denoising process and the diffusion time. According to the gradient value and the diffusion coefficient, the main form of the diffusion equation is:
[0025]
[0026]
[0027] Where k represents the diffusion threshold. is much larger than k, the diffusion equation tends to 0, that is, diffusion is inhibited; when If it is much smaller than k, the diffusion equation tends to 1, which means that diffusion is enhanced.
[0028] Discretize the diffusion equation and further change the edge coefficients. The specific form is as follows:
[0029]
[0030]
[0031] Replacing the gradient operator with q(x,y,t) is called the instantaneous coefficient of variation (ICOV), where q0(t) represents the diffusion threshold at that moment and c(x,y,t) is the diffusion coefficient.
[0032] Furthermore, the noise residual image of the input image is subtracted from the denoised image to obtain the noise residual image of the input image. At this time, the noise residual image still contains certain signal information. The noise residual information is once again anisotropically filtered to obtain residual information, and the noise residual image and residual information are added to obtain the final filtering result.
[0033] Furthermore, in step 2, the preprocessed sonar image is transformed with different filters to obtain two sub-images of the same size as the original image after transformation. The two sub-images are used as input and the feature sub-images are first obtained by the Gamma enhancement method. Figure 1 , the specific steps are as follows:
[0034] X=(I+c) γ
[0035] Among them, X represents the processed sub-image, I represents the input gray value, and c is a constant coefficient used to control the slope of the curve, which is generally 1. γ determines the relationship between input and output. If γ>1, the area with higher gray levels will be stretched and the area with lower gray levels will be compressed, making the overall image darker; otherwise, the overall image will be brighter.
[0036] Furthermore, in step 2, the preprocessed sonar image is transformed with different filters to obtain two sub-images of the same size as the original image after transformation. The two sub-images are used as input and mean filtering is introduced to construct feature sub-images. The specific steps are as follows:
[0037]
[0038] Among them, X is processed using mean filtering to obtain Where ave(·) represents mean filtering.
[0039] Furthermore, in step three, the KL divergence constraint is introduced into the FCM, that is, the KL divergence constraint term is added to the objective function. The specific steps are as follows:
[0040]
[0041]
[0042] in, u ij and is the membership function of the element corresponding to the cluster center in the graph, α is the regular term coefficient, which controls the influence of KL divergence on the objective function.
[0043] According to the Lagrange multiplier method, the iterative method for the cluster center v and the membership function u is obtained:
[0044]
[0045]
[0046] Furthermore, in step three, in order to speed up the convergence of the objective function, a sparse membership function is introduced to increase the convergence speed. The specific steps are as follows:
[0047]
[0048] Introducing sparse membership u ij_s , replacing the original membership function, where ε is a minimum value. The larger its value is, the sparser the membership matrix is and the faster the iteration speed is.
[0049] The beneficial effects of the present invention are:
[0050] Compared with other classic sonar image target segmentation methods, the method of the present invention greatly alleviates the pulse noise and speckle noise that exist in large quantities in sonar images through morphological reconstruction and residual filtering, and improves the segmentation speed to a certain extent. The method of the present invention introduces a self-structure information extraction method, obtains richer input feature quantities by transforming the input sample itself, and uses the iterative update process of KL divergence constraint FCM to greatly improve the robustness of the image segmentation algorithm. The method of the present invention introduces the concept of sparse membership function and solves the problem of low efficiency of FCM applied to sonar image target segmentation by sparse membership. Through the segmentation test of sonar images, it is known that the method of the present invention has better robustness and stronger anti-interference ability, and can take into account the accuracy and efficiency of image segmentation. The method of the present invention is more suitable for processing the problem of underwater target segmentation of sonar images. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 It is a flow chart of the sonar image segmentation method based on anisotropic residual denoising and sparse KL fuzzy clustering according to the present invention;
[0052] Figure 2 It is a denoising flow chart combining morphological reconstruction and residual denoising;
[0053] Figure 3 This is the segmentation result after noise reduction in the synthetic image;
[0054] Figure 4 This is the result of forward-looking sonar image segmentation;
[0055] Figure 5 The following is a curve showing the effect of adding sparse membership on the accuracy change when using the segmentation method of the present invention with different regularization term coefficient settings. DETAILED DESCRIPTION
[0056] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.
[0057] In active sonar target recognition, the transducer array gathers the acoustic signal to form a beam. The generated acoustic signal propagates underwater and generates an echo signal when encountering a target. The sonar image is processed by the transducer receiving array elements. Compared with traditional optical images, sonar images are greatly affected by noise. Sonar images show low contrast, blurred target edges, and severe sidelobe interference. Using sonar images directly for target segmentation tasks without processing will encounter many problems.
[0058] Combination Figure 1-Figure 5 The present invention proposes a sonar image segmentation method based on anisotropic residual denoising and sparse fuzzy clustering, the method comprising the following steps:
[0059] S1: For the collected sonar image data, the data is first enhanced and denoised through morphological reconstruction and residual denoising structure to achieve preprocessing;
[0060] S2: Perform different filter transformations on the preprocessed sonar image to obtain two sub-images of the same size as the original image after transformation, and use the two sub-images as input;
[0061] S3: Introduce KL divergence constraints in FCM to ensure that the intensity of the current pixel value is close to the average intensity of its neighboring pixel values, and ensure that the membership matrices of the two sub-graphs are as consistent as possible. At the same time, in order to speed up the convergence of the objective function, introduce sparse membership function to increase the convergence speed.
[0062] Combination Figure 2 The step S1 pre-processes the sonar image, specifically comprising the following steps:
[0063] S1.1: There are a lot of impulse noise and speckle noise in the sonar image. First, the noise in the sonar image is preliminarily processed through morphological reconstruction (MR). The specific form is as follows:
[0064]
[0065]
[0066]
[0067] Defining Morphological Dilation Reconstruction f represents the original input image, g represents the marker image, g≤f, δ is the dilation operation, and ∧ represents the point-by-point minimum.
[0068]
[0069]
[0070]
[0071] Defining Morphological Corrosion Reconstruction g≥f, ε is the corrosion operation, and ∨ represents the point-by-point maximum value.
[0072] Morphological closed reconstruction is used as the initial impulse noise denoising method and the subsequent label reconstruction method. g = ε(f) and g = δ(f) are selected as the label images for dilation reconstruction and erosion reconstruction respectively. Based on the above operator, morphological closed reconstruction R C (f) is defined as:
[0073]
[0074] S1.2: For the morphologically reconstructed sonar image, the residual denoising structure is used to complete the denoising. First, the sonar image is denoised by the anisotropic diffusion (SRAD) denoising method. The specific form is as follows:
[0075]
[0076] I(t=0)=I0
[0077] Among them, div represents the divergence operator, represents the gradient operator, represents the diffusion equation, and ||·|| represents the amplitude. t represents the introduced time operator, which can reflect the correlation between the denoising process and the diffusion time. According to the gradient value and the diffusion coefficient, the main form of the diffusion equation is:
[0078]
[0079]
[0080] Where k represents the diffusion threshold. When ||▽I|| is much larger than k, the diffusion equation tends to 0, which means that diffusion is suppressed; when ||▽I|| is much smaller than k, the diffusion equation tends to 1, which means that diffusion is enhanced.
[0081] Discretize the diffusion equation and further change the edge coefficients. The specific form is as follows:
[0082]
[0083]
[0084] Replacing the gradient operator with q(x,y,t) is called the instantaneous coefficient of variation (ICOV), where q0(t) represents the diffusion threshold at that moment and c(x,y,t) is the diffusion coefficient.
[0085] S1.3: Subtract the denoised image from the input image to obtain the noise residual image of the input image. At this time, the noise residual image still contains certain signal information. Anisotropic filtering is applied to the noise residual information again to obtain residual information. The noise residual image and the residual information are added to obtain the final filtering result.
[0086] The step S2 obtains two sub-images of the same size as the original image after transformation, and specifically includes the following steps:
[0087] S2.1: The preprocessed sonar image is transformed with different filters to obtain two sub-images of the same size as the original image after transformation. First, the characteristic sub-image is obtained by the Gamma enhancement method. Figure 1 , the specific steps are as follows:
[0088] X=(I+c) γ
[0089] Among them, X represents the processed sub-image, I represents the input gray value, and c is a constant coefficient used to control the slope of the curve, which is generally 1. γ determines the relationship between input and output. If γ>1, the area with higher gray levels will be stretched and the area with lower gray levels will be compressed, making the overall image darker; otherwise, the overall image will be brighter.
[0090] S2.2: The preprocessed sonar image is transformed with different filters to obtain two sub-images of the same size as the original image after transformation. The mean filter is introduced to construct the feature sub-image. The specific steps are as follows:
[0091]
[0092] Among them, X is processed using mean filtering to obtain Where ave(·) represents mean filtering.
[0093] S2.3: Stretch the two groups of processed sonar image sub-images obtained in S2.1 and S2.2 into one-dimensional vectors as input data sets for subsequent algorithms.
[0094] Combination Figure 1 The step S3 introduces KL divergence constraint and sparse membership function in FCM to increase the convergence speed, and specifically includes the following steps:
[0095] S3.1: Introduce KL divergence constraint in FCM, that is, add KL divergence constraint term in the objective function. The specific steps are as follows:
[0096]
[0097]
[0098] in, u ij and is the membership function of the element corresponding to the cluster center in the graph, α is the regular term coefficient, which controls the influence of KL divergence on the objective function.
[0099] According to the Lagrange multiplier method, the iterative method for the cluster center v and the membership function u is obtained:
[0100]
[0101]
[0102] S3.2: In order to speed up the convergence of the objective function, a sparse membership function is introduced to increase the convergence speed. The specific steps are as follows:
[0103]
[0104] Introducing sparse membership u ij_s , replace the original membership function, where ε is a minimum value. The larger its value is, the sparser the membership matrix is and the faster the iteration speed is. The value of ε can be selected as 10 -2 .
[0105] S3.3: Reconstruct the processed one-dimensional array according to the length and width of the original input sonar image to obtain a processed sonar image.
[0106] The evaluation process of the present invention is described below:
[0107] In this embodiment, the segmentation accuracy of a synthetic image with added noise is used to evaluate the advantages and disadvantages of the noise reduction method after using different noise reduction methods. The evaluation indicators include accuracy Acc, sensitivity Sen. and specificity Spe. The specific calculation formulas of the above evaluation indicators are as follows:
[0108] Acc=(TP+TN) / (FN+FP+TP+TN)
[0109] Sen.=TP / (FN+TP)
[0110] Spe.=TN / (FP+TN)
[0111] Among them, TP represents the number of samples whose types and discrimination results are both positive samples, FP represents the number of samples that are negative samples but are mistakenly judged as positive samples, TN represents the number of samples whose types and discrimination results are both negative samples, and FN represents the number of samples that are positive samples but are mistakenly judged as negative samples.
[0112] The test used a classic noise reduction method and compared it with the method described in the present invention, and the results shown in Table 1 were obtained.
[0113] Table 1
[0114]
[0115] Combining Table 1 and Figure 3 After denoising, the method proposed in the present invention can significantly improve the image segmentation performance, and the effect is relatively ideal.
[0116] In this embodiment, the evaluation index of segmentation accuracy includes the partition entropy V pe , division coefficient V pc 、XB index V XB . Partition entropy V pe and the partition coefficient V pc Reflects the fuzziness of the segmentation matrix, V pc The larger the value, the smaller the fuzziness of the segmentation matrix, and the better the segmentation performance; V pe The smaller it is, the more accurate the pixel classification is and the better the segmentation effect is. XB Find a balance point between intra-class compactness and inter-class separation. The smaller the result, the better the clustering effect. The specific calculation formula of the above evaluation indicators is as follows:
[0117]
[0118]
[0119]
[0120] Where n is the number of sample points, K is the number of categories, and u ki is the corresponding degree of membership.
[0121] Test forward-looking sonar images using classic FCM with Figure 4 The results shown in Table 2 were obtained by comparison with the method described in the present invention shown in Table 2.
[0122] Table 2
[0123]
[0124] In Table 2, the applicability of the proposed method for target segmentation in forward-looking sonar images is verified based on numerical evaluation.
[0125] The beneficial effects of the present invention are:
[0126] (1) The use of morphological reconstruction and residual denoising greatly alleviates the large amount of impulse noise and speckle noise in sonar images and improves the segmentation speed to a certain extent;
[0127] (2) Using the feature extraction method of self-structural information, a richer input feature quantity is obtained by transforming the input sample itself;
[0128] (3) The KL divergence is used to constrain the iterative update process of FCM, which greatly increases the noise resistance and robustness of the algorithm. Sparse membership is introduced to improve the convergence speed, such as Figure 5 As shown, this method achieves a balance between segmentation efficiency and segmentation accuracy.
[0129] The above is a detailed introduction to the sonar image segmentation method of anisotropic residual denoising and sparse fuzzy clustering proposed in the present invention. In this article, specific examples are used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A sonar image segmentation method based on anisotropic residual denoising and sparse fuzzy clustering, characterized by: The method comprises the following steps: Step 1: For the collected sonar image data, denoising is first performed through morphological reconstruction and residual denoising structure to achieve preprocessing; Step 2: Perform different filter transformations on the preprocessed sonar image to obtain two sub-images of the same size as the original image after transformation, and use the two sub-images as input; Step 3: Introduce KL divergence constraints in FCM to ensure that the intensity of the current pixel value is close to the average intensity of its neighboring pixel values, and ensure that the membership matrices of the two sub-graphs are as consistent as possible; at the same time, in order to speed up the convergence of the objective function, introduce sparse membership function to increase the convergence speed.
2. The sonar image segmentation method based on anisotropic residual denoising and sparse fuzzy clustering according to claim 1, characterized in that: In step 1, the noise in the sonar image is initially processed by morphologically reconstructing MR, which is specifically in the following form: Defining Morphological Dilation Reconstruction f represents the original input image, g represents the marker image markerimage, g≤f, δ is the dilation operation, and ∧ represents the point-by-point minimum value; Defining Morphological Corrosion Reconstruction g≥f, ε is the corrosion operation, and ∨ represents the point-by-point maximum value; Morphological closed reconstruction is used as the initial impulse noise denoising method and the subsequent label reconstruction method. g = ε(f) and g = δ(f) are selected as the label images for dilation reconstruction and erosion reconstruction respectively. Based on the above operator, morphological closed reconstruction R C (f) is defined as:
3. The sonar image segmentation method based on anisotropic residual denoising and sparse fuzzy clustering according to claim 2, characterized in that: In step 1, the residual denoising structure is continued to be used to complete the denoising; first, the anisotropic diffusion SRAD denoising method is used to denoise the sonar image, and the specific form is as follows: Among them, div represents the divergence operator, represents the gradient operator, represents the diffusion equation, ||·|| represents the amplitude; t represents the introduced time operator, which can reflect the correlation between the denoising process and the diffusion time. According to the gradient value and the diffusion coefficient, the main form of the diffusion equation is: Where k represents the diffusion threshold; when is much larger than k, the diffusion equation tends to 0, that is, diffusion is inhibited; when If it is much smaller than k, the diffusion equation approaches 1, which means that diffusion is enhanced; Discretize the diffusion equation and further change the edge coefficients. The specific form is as follows: Replacing the gradient operator with q(x,y,t) is called the instantaneous coefficient of variation (ICOV), where q0(t) represents the diffusion threshold at that moment and c(x,y,t) is the diffusion coefficient.
4. The sonar image segmentation method based on anisotropic residual denoising and sparse fuzzy clustering according to claim 1, characterized in that: In step 1, the noise residual image of the input image is obtained by subtracting the denoised image from the input image. At this time, the noise residual image still contains certain useful signal information; Anisotropic filtering is applied to the noise residual information again to obtain residual information, and the noise residual image and the residual information are added to obtain the final filtering result.
5. The sonar image segmentation method based on anisotropic residual denoising and sparse fuzzy clustering according to claim 1, characterized in that: In step 2, the preprocessed sonar image is transformed with different filters to obtain two sub-images of the same size as the original image after transformation, and the two sub-images are used as input.
6. The sonar image segmentation method based on anisotropic residual denoising and sparse fuzzy clustering according to claim 5, characterized in that: The characteristic sub-image 1 and the characteristic sub-image 2 are obtained by the Gamma enhancement method and the mean filtering respectively. The specific steps are as follows: X=(I+c) γ Among them, X represents the processed feature sub-image 1, I represents the input gray value, c is a constant coefficient used to control the slope of the curve, and its value is 1; γ determines the relationship between input and output. If γ>1, the area with higher gray level will be stretched and the area with lower gray level will be compressed, making the overall image darker; otherwise, the overall image will be brighter; Among them, X is processed using mean filtering to obtain represents feature sub-graph 2, where ave(·) represents mean filtering.
7. The sonar image segmentation method of anisotropic residual denoising and sparse fuzzy clustering according to claim 1, characterized in that: In step 3, the KL divergence constraint is introduced in FCM, that is, the KL divergence constraint term is added to the objective function. The specific steps are as follows: in, u ij and is the membership function of the element corresponding to the cluster center in the graph, α is the regular term coefficient, which controls the influence of KL divergence on the objective function. According to the Lagrange multiplier method, the iterative method for the cluster center v and the membership function u is obtained:
8. The sonar image segmentation method of anisotropic residual denoising and sparse fuzzy clustering according to claim 1, characterized in that: In step three, a sparse membership function is introduced to increase the convergence speed. The specific steps are as follows: Introducing sparse membership u ij_s , replacing the original membership function, where ε is a minimum value. The larger its value is, the sparser the membership matrix is and the faster the iteration speed is.
Citation Information
Patent Citations
A sonar image equalization method based on SLIC and adaptive filtering
CN109712099A
An image segmentation method based on wavelet energy and a fuzzy C-means
CN109712149A
Forward-looking sonar image segmentation method based on empirical mode decomposition
CN114494280A
Image segmentation method based on residual diffusion model
CN116993770A
Method and apparatus for image processing
US20100092102A1
Cited By
On-line detection method and device for ball grinding and polishing machining track
CN121893152A
