CN-121998861-A - Acoustic image processing method and device based on sparse representation

CN121998861ACN 121998861 ACN121998861 ACN 121998861ACN-121998861-A

Abstract

The application belongs to the technical field of image processing, and relates to an acoustic image processing method and device based on sparse representation. The method comprises the steps of obtaining a noise image containing noise, constructing an overcomplete dictionary by taking a DCT (discrete cosine transform) base as an initial value, determining a sparse coefficient matrix by minimizing a reconstruction error and applying a sparse constraint, updating the overcomplete dictionary based on the sparse coefficient matrix, re-solving the sparse coefficient matrix until the overcomplete dictionary converges, performing sparse coding and reconstruction on the noise image by utilizing the converged overcomplete dictionary to obtain a preliminary denoising image, constructing a local linear association relation between the reference image and the preliminary denoising image by taking the initial noise image as a reference image, and performing edge enhancement on the reference image by a local linear mapping edge enhancement processing method guided by the reference image to obtain a final denoising image. The application can effectively inhibit Gaussian and speckle mixed noise in the sonar image, and remarkably reserve and strengthen edge structural characteristics.

Inventors

Jing Xinkang
TANG CHEN
LU JIE
ZHU JIAWEN
LUO LILONG
CHANG LIANG

Assignees

中国飞机强度研究所

Dates

Publication Date: 20260508
Application Date: 20260410

Claims (10)

1. An acoustic image processing method based on sparse representation, comprising: s1, acquiring a noise image containing noise; s2, constructing an overcomplete dictionary by taking the DCT base as an initial value; Step S3, determining a sparse coefficient matrix by minimizing reconstruction errors and applying sparse constraint on the basis of the current overcomplete dictionary according to each image block vector with direct current components removed in the noise image; S4, updating the overcomplete dictionary based on the sparse coefficient matrix, and re-solving the sparse coefficient matrix until the overcomplete dictionary converges; s5, performing sparse coding and reconstruction on the noise image by using the converged overcomplete dictionary to obtain a preliminary denoising image; And S6, constructing a local linear association relation between the reference image and the preliminary denoising image by taking the initial noise image as the reference image, and carrying out edge enhancement on the reference image by a local linear mapping edge enhancement processing method guided by the reference image to obtain a final denoising image.
2. The sparse representation based acoustic image processing method of claim 1, wherein step S2 further comprises: S21, generating a group of one-dimensional cosine base vectors, and carrying out mean value removal and unit norm normalization processing on the base vectors except the direct current component; step S22, constructing a two-dimensional cosine base matrix through tensor product of one-dimensional base vectors, and intercepting or zero filling the matrix according to the required atomic number K to form an initial overcomplete dictionary.
3. The sparse representation based acoustic image processing method of claim 1, wherein in step S3, coefficients of a sparse coefficient matrix are determined by the following optimization function : ; The constraint condition is that ; Wherein, the For the image block vector to be a vector of image blocks, In order to have an overcomplete dictionary, Is a sparsity control parameter.
4. The sparse representation based acoustic image processing method of claim 1, wherein step S6 further comprises: Step S61, in the local neighborhood with radius r In, calculate the local mean of the reference image Local mean of preliminary denoised image ; Step S62, calculating the cross-correlation parameters of the reference image and the preliminary denoising image according to the local mean value of the reference image and the local mean value of the preliminary denoising image Determining local variance of reference image according to local mean of reference image ; Step S63, solving the local linear weight coefficient according to the following formula And local gray scale offset : ; ; Wherein, the Is a regularization parameter; Step S64, calculating the local linear weight coefficient And local gray scale offset In its corresponding neighborhood The mean value is smoothed to obtain the optimized weight coefficient And offset amount ; And step S65, generating a final denoising image through linear function transformation according to the optimized weight coefficient and offset.
5. The sparse representation based acoustic image processing method of claim 4, wherein the radius of the local neighborhood r = 6, the regularization parameter 。
6. An acoustic image processing apparatus based on sparse representation for implementing the acoustic image processing method based on sparse representation according to claim 1, said apparatus comprising: the noise image acquisition module is used for acquiring a noise image containing noise; The over-complete dictionary initialization module is used for constructing an over-complete dictionary by taking the DCT base as an initial value; The sparse coefficient matrix determining module is used for determining a sparse coefficient matrix by minimizing a reconstruction error and applying a sparse constraint on the basis of the current overcomplete dictionary according to each image block vector with the direct current component removed in the noise image; The overcomplete dictionary generating module is used for updating the overcomplete dictionary based on the sparse coefficient matrix and re-solving the sparse coefficient matrix until the overcomplete dictionary converges; The preliminary denoising module is used for performing sparse coding and reconstruction on the noise image by utilizing the converged overcomplete dictionary to obtain a preliminary denoising image; The final denoising module is used for constructing a local linear association relation between the reference image and the preliminary denoising image by taking the initial noise image as the reference image, and performing edge enhancement on the reference image by a local linear mapping edge enhancement processing method guided by the reference image to obtain a final denoising image.
7. The sparse representation based acoustic image processing apparatus of claim 6, wherein the overcomplete dictionary initialization module comprises: The one-dimensional cosine base vector generation unit is used for generating a group of one-dimensional cosine base vectors and carrying out mean value removal and unit norm normalization processing on the base vectors except the direct current component; the two-dimensional cosine base matrix generation unit is used for constructing a two-dimensional cosine base matrix through tensor product of one-dimensional base vectors, and intercepting or zero filling operation is carried out on the matrix according to the required atomic number K so as to form an initial overcomplete dictionary.
8. The sparse representation based acoustic image processing apparatus of claim 6, wherein in the sparse coefficient matrix determination module, coefficients of a sparse coefficient matrix are determined by the following optimization function : ; The constraint condition is that ; Wherein, the For the image block vector to be a vector of image blocks, In order to have an overcomplete dictionary, Is a sparsity control parameter.
9. The sparse representation based acoustic image processing apparatus of claim 6, wherein the final denoising module comprises: A mean value calculation unit for calculating the mean value of the local neighborhood with radius r In, calculate the local mean of the reference image Local mean of preliminary denoised image ; A cross-correlation parameter and local variance calculation unit for calculating cross-correlation parameters of the reference image and the preliminary denoising image according to the local mean of the reference image and the local mean of the preliminary denoising image Determining local variance of reference image according to local mean of reference image ; A weight and offset calculation unit for solving local linear weight coefficient according to the following formula And local gray scale offset : ; ; Wherein, the Is a regularization parameter; a weight and offset optimizing unit for optimizing the local linear weight coefficient obtained by calculation And local gray scale offset In its corresponding neighborhood The mean value is smoothed to obtain the optimized weight coefficient And offset amount ; And the edge enhancement unit is used for generating a final denoising image through linear function transformation according to the optimized weight coefficient and offset.
10. The sparse representation based acoustic image processing apparatus of claim 9, wherein the radius of the local neighborhood r = 6, the regularization parameter 。

Description

Acoustic image processing method and device based on sparse representation Technical Field The application belongs to the technical field of image processing, and particularly relates to an acoustic image processing method and device based on sparse representation. Background The sonar imaging technology is used as a core means of underwater environment sensing and target detection, is widely applied to the fields of ocean engineering, underwater security, resource exploration and the like, and the output sonar image quality directly determines the performance upper limit of downstream computer vision tasks (such as target detection, semantic segmentation and terrain matching). The problems of image gray dynamic range compression, detail information distortion and the like are caused by the coupling influence of underwater propagation medium characteristics (water scattering and acoustic attenuation) and hardware system noise (transducer thermal noise and signal acquisition circuit interference), mixed pollution of multiplicative speckle noise (caused by the coherence of acoustic multipath reflection) and additive Gaussian noise is commonly existed in a sonar image, the spatial distribution of the two types of noise has non-uniformity and correlation, the edge contour (such as an underwater structure boundary and a submarine micro-topography texture) of an underwater target is easy to mask, and the environmental perception precision and the target recognition robustness of the sonar system are seriously restricted. In the traditional denoising method, bilateral filtering is used as a typical airspace edge preservation filtering method, the core of the technical route is to construct a joint weight function of a spatial domain and a gray domain, a local filtering window is defined for each pixel in an image, the spatial distance weight of the pixel in the window and a central pixel is calculated by a Gaussian function, meanwhile, gray similarity weight is constructed based on the gray difference value between the pixels, after the two types of weights are multiplied to obtain the joint weight, the gray value of the pixel in the window is weighted and averaged, so that smooth suppression of noise is realized, and meanwhile, the gray abrupt edge region is protected through the gray similarity weight, and the edge blurring is avoided. The guide filtering is based on a local linear model design technical route, a specific image is firstly selected as a guide image, the filter output image and the guide image are assumed to meet the linear mapping relation in a local window, the linear coefficient and the offset in each window are solved by minimizing the error of the input noise-containing image and the output of the linear model and introducing regularization term constraint model parameters, all window parameters containing the same pixel are averaged, and finally a globally smooth denoising image following the edge structure of the guide image is generated, so that the edge detail is ensured to be reserved while the noise is smoothed. The technical route of wavelet transformation is developed around multi-scale frequency domain decomposition and reconstruction, firstly, a proper wavelet basis function is selected, a noisy sonar image is decomposed into a low-frequency approximate component reflecting the whole structure and a high-frequency detail component containing edges and noise, the high-frequency component is subjected to threshold processing (such as hard threshold elimination of small coefficients and soft threshold shrinkage coefficients) to inhibit the noise, effective edge information is reserved, and then the processed low-frequency component and the high-frequency component are reconstructed through inverse wavelet transformation to obtain a denoised image, so that the separation of the noise and signals in the frequency domain is realized. The deep learning denoising method adopts an end-to-end technical route, a deep neural network (such as a convolutional neural network and a transducer) is firstly constructed, a large number of noise-containing-clean image pairs are used as training data, the network learns the mapping relation from the noise-containing image to the clean image through forward propagation, the difference between the network output and the clean label is measured through a loss function (such as mean square error and perception loss) in the training process, the network parameters are optimized through backward propagation, the network has the capability of adaptively extracting noise characteristics and image structural characteristics, after training is completed, a sonar image to be denoised is input into the network, the denoising result is directly output through nonlinear mapping of the network, the artificial design filtering rule is not needed, and the complex noise suppression is realized by means of data driving. In the research of sonar image denoising, although th