Image denoising method and device based on statistical analysis

Abstract

The invention provides an image denoising method and device based on statistical analysis, and relates to the technical field of image processing, and the method comprises the following steps: carrying out the processing and analysis of a laser phase unwrapping image, and building a training image data set; constructing an image denoising model, and training the image denoising model by using the training image data set; acquiring an actual laser image, and denoising by using the trained image denoising model; according to the method, a noise model of laser imaging is established through theoretical analysis and experimental verification, a noiseless phase unwrapped image is simulated to construct a corresponding data set, the established training image data set is adopted to train an image denoising model, the image denoising model is subsequently utilized to denoise the actual phase unwrapped image, and the noise-free phase unwrapped image is obtained. Noise suppression of a real phase unwrapped image can be effectively realized, and the laser imaging denoising performance is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more particularly to an image denoising method and device based on statistical analysis. Background Art

[0002] At present, laser imaging technology forms an observable stripe pattern on the surface of an object through a laser projector, and then these laser stripe images are captured by an imaging device, and a point cloud is generated using reconstruction technology. However, in this process, the noise caused by coherence will introduce significant errors in the laser stripe image, phase wrapping image and phase unwrapping image, thereby reducing the clarity of the image. The ideal phase unwrapping image should be continuous in space and have a predictable trend, but the interference of noise may cause discontinuous jumps and abnormal phase changes, which in turn affect the final three-dimensional data. Therefore, denoising of the phase unwrapping image is particularly important to reduce the impact of noise on the data.

[0003] Researchers have proposed a variety of filtering techniques to address the above problems, which can be roughly divided into three categories: spatial domain filtering, transform domain filtering, and deep learning filtering. Spatial domain filtering is mainly based on the statistical characteristics of the image and processes the entire image, including mean filtering, median filtering, Gaussian filtering, bilateral filtering, and Kuan filtering. Transform domain filtering methods involve domain transformation of the image, such as Fourier transform and wavelet transform, as well as model-based denoising methods such as BM3D and WNNM, which can flexibly handle different noise levels. With the rapid development of deep learning technology, neural network-based filtering models, such as DnCNN, FFDNet, CBDNet, RIDNet, and PMRID, have become a research hotspot in this field. Although deep learning filtering algorithms continue to emerge, their performance depends to a certain extent on high-quality training data, and most of their data sets come from image data sets such as the Berkeley segmentation data set. However, the existing BSD data sets are usually not suitable for specific filtering tasks, and creating data sets suitable for specific filtering tasks is both complex and time-consuming.

[0004] Therefore, how to provide an image denoising method based on statistical analysis that can solve the above problems is an urgent problem that those skilled in the art need to solve. Summary of the invention

[0005] In view of this, the present invention provides an image denoising method and device based on statistical analysis. A noise model of laser imaging is established through theoretical analysis and experimental verification, and a noise-free phase unwrapping image is simulated to construct a corresponding data set. The denoising model is trained through the data set, and the denoising model is subsequently used for denoising, which can effectively achieve noise suppression of the real phase unwrapping image and significantly improve the performance of laser imaging denoising.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] An image denoising method based on statistical analysis comprises the following steps:

[0008] Acquire laser images and phase unwrapping images generated in laser imaging for processing and analysis, and construct a training image data set;

[0009] Constructing an image denoising model, and training the image denoising model using the training image dataset;

[0010] An actual phase unwrapping image is obtained, and denoised using the trained image denoising model.

[0011] Preferably, the specific process of constructing the training image data set includes:

[0012] Processing the laser image to obtain a corresponding echo signal;

[0013] Processing the echo signal to obtain a corresponding noise model;

[0014] The phase unwrapping image is analyzed in combination with the noise model, and a training image data set is constructed according to the analysis result.

[0015] Preferably, the specific process of processing the echo signal includes:

[0016] Analyze the echo signal to obtain a probability density function corresponding to the amplitude of the echo signal;

[0017] Determine, according to the probability density function, a probability density function of an additive noise signal included in the echo signal;

[0018] The phase unwrapping image is fitted, and the fitting result, the laser image and the probability density function of the additive noise signal are logarithmically transformed and subtracted point by point to determine the noise model.

[0019] Preferably, the specific process of processing the echo signal further includes:

[0020] The noise model is verified.

[0021] Preferably, the specific process of determining the probability density function of the additive noise signal contained in the echo signal includes:

[0022] Determining a probability density function of pixel intensities included in the phase unwrapped image according to the probability density function of the amplitude;

[0023] The probability density function of the original noise is determined, and the probability density function of the signal including the additive noise is determined based on the additive noise model.

[0024] Preferably, the image denoising model is a DnCNN model, comprising an input layer, a hidden layer and an output layer connected in sequence.

[0025] The present invention also provides an image denoising device based on statistical analysis, comprising:

[0026] A data set construction module is used to process and analyze the laser images and phase unwrapping images generated in laser imaging, and to construct a training image data set;

[0027] A model building module, used to build an image denoising model and train the image denoising model using the training image data set;

[0028] The processing module is used to obtain an actual laser image and perform denoising using the trained image denoising model.

[0029] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and device for image denoising based on statistical analysis, which effectively filters out the noise by efficiently suppressing the specific Gumbel noise in the image. The algorithm is specifically optimized for phase unwrapping images, and significantly improves the performance of laser imaging denoising through precise statistical analysis and deep learning technology.

[0030] There are currently a variety of image filtering models based on deep learning, such as the DnCNN model, FFDNet model, CBDNet model, RIDNet model, PMRID network model, etc. Although these deep learning filtering algorithms continue to emerge, their performance depends largely on high-quality training data, and most of these data sets come from the Berkeley segmentation data set. Creating data sets specific to filtering tasks is usually complex and time-consuming, which increases the difficulty of implementation.

[0031] Through in-depth theoretical derivation, the present invention reveals that the noise model in laser imaging follows the Gumbel distribution, and develops a targeted phase unwrapping image dataset. Combined with the classic DnCNN network model, the purpose of three-dimensional point cloud filtering is achieved through phase unwrapping image denoising. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0033] Figure 1 An overall flow chart of an image denoising method based on statistical analysis provided by the present invention;

[0034] Figure 2 A schematic diagram of the principle of three-dimensional laser imaging provided by an embodiment of the present invention;

[0035] Figure 3 A schematic diagram of the noise distribution of a target provided by an embodiment of the present invention, wherein a represents the original data of the plane and the three-dimensional display of the fitted plane, and b represents the noise intensity distribution of the plane;

[0036] Figure 4 A schematic diagram of the DnCNN network structure provided by an embodiment of the present invention;

[0037] Figure 5 A block diagram of the structural principles of an image denoising device based on statistical analysis provided by the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.

[0039] See also Figure 1 As shown, the embodiment of the present invention discloses an image denoising method based on statistical analysis, comprising the following steps:

[0040] Acquire laser images and phase unwrapping images generated in laser imaging for processing and analysis, and construct a training image data set;

[0041] Build an image denoising model and use the training image dataset to train the image denoising model;

[0042] Get the actual phase unwrapped image and denoise it using the trained image denoising model.

[0043] In a specific embodiment, the specific process of constructing the training image dataset includes:

[0044] Process the laser image to obtain the corresponding echo signal;

[0045] Process the echo signal to obtain the corresponding noise model;

[0046] The phase unwrapping image is analyzed in combination with the noise model, and a training image dataset is constructed based on the analysis results.

[0047] In a specific embodiment, the specific process of processing the echo signal includes:

[0048] Analyze the echo signal to obtain the probability density function corresponding to the amplitude of the echo signal;

[0049] Determine the probability density function of the additive noise signal contained in the echo signal according to the probability density function;

[0050] The phase unwrapping image is fitted, and the fitting result, the laser image and the probability density function of the additive noise signal are logarithmically transformed and subtracted point by point to determine the noise model.

[0051] In a specific embodiment, the specific process of processing the echo signal further includes:

[0052] Verify the noise model.

[0053] In a specific embodiment, the specific process of determining the probability density function of the additive noise signal contained in the echo signal includes:

[0054] Determining a probability density function of pixel intensities included in the phase unwrapping image according to the probability density function of the amplitude;

[0055] The probability density function of the original noise is determined, and the probability density function of the signal including the additive noise is determined based on the additive noise model.

[0056] Specifically, Figure 2 This is a diagram of the theoretical analysis of three-dimensional laser imaging. When the light pulse emitted by the light source propagates in multiple directions and encounters irregular scatterers and uneven interfaces, scattered light waves with different propagation directions are generated. Part of the scattered light waves will be returned in the form of backscattering and captured by the receiver. Since the returned scattered light waves come from the same light source, they follow specific statistical laws. The echo generated by each scatterer can be expressed as a complex number. The total received echo signal is regarded as the superposition of multiple echoes. The specific expression is:

[0057]

[0058] In the formula, represents the echo of the kth scattering point, A k and θ k are amplitude and phase respectively, and n is the total number of scattering points. The total echo signal can be expressed as a complex number Z = A r +jA i =Acosθ+jAsinθ. When n is large, the real part A r and the imaginary part A iIt is approximately in accordance with the Gaussian distribution with a mean of 0 and a standard deviation of σ. The amplitude A of the echo signal is calculated by formula (2):

[0059]

[0060] In the formula, A r 2 and A i 2 It will obey the chi-square distribution with a degree of freedom of 1. Therefore, the probability density function of the echo signal amplitude will conform to the Rayleigh distribution, and the specific expression is:

[0061]

[0062] Where σ is the standard deviation. The pixel intensity I is calculated by I = A 2 Calculate, the specific expression of the probability density function of I is:

[0063]

[0064] The coherence between lasers generates a large amount of multiplicative noise model:

[0065] G(x,y)=F(x,y)N(x,y) (5)

[0066] In the formula, G(x,y) is the noisy laser image, F(x,y) is the noise-free desired image, and N(x,y) is the noise component. The multiplicative noise is converted into additive noise using logarithmic transformation to simplify the noise model. The specific expression of the model is:

[0067] G′(x,y)=F′(x,y)+N′(x,y) (6)

[0068] Where G'(x,y), F'(x,y) and N'(x,y) are the logarithmic transformed laser noise image, the logarithmic transformed noise-free desired image and the logarithmic transformed noise component, respectively. The specific expression of N' is:

[0069] N′=aln(N)+b (7)

[0070] Where a and b are coefficients. Then the probability density function f of N' is N′ (n′) is obtained by transforming the probability density function of the original noise N:

[0071]

[0072] In the imaging process, noise usually comes from the superposition of a large number of scattered light waves generated by the same light source, so they follow the same statistical laws as the echo signal. The specific expression of the probability density function of the original noise is:

[0073]

[0074] The specific expression of the probability density function of additive noise is obtained as follows:

[0075]

[0076] The original noise distribution is logarithmically transformed and its probability density function is obtained, which conforms to the Gumbel model equation.

[0077] In order to verify the noise model, a flat plate was selected as the target object, and its simple geometric characteristics and controllable surface reflection properties provided ideal test conditions for the experiment. This design can accurately evaluate and verify the statistical characteristics of image noise, providing a solid foundation for the subsequent development and optimization of noise suppression algorithms.

[0078] First, it is reasonable to assume that the phase unwrapped image mainly contains noise components and useful phase change signals. The least squares optimization algorithm is used to fit the image data with a polynomial equation, and the fitting plane is considered to be a useful signal without noise:

[0079] Z=α0+α1X+α2Y+α3X 2 +α4XY+α5Y 2 +α6X 3 +α7X 2 Y+α8XY 2 +α9Y 3 (11)

[0080] In the formula, Z is the dependent variable, X and Y are independent variables, and α k (k=1,2,...,9) are model parameters. Figure 3 (a) is a three-dimensional display of the original data and the fitted data. The original data and the fitted data are logarithmically transformed and subtracted point by point to extract the noise component after logarithmic transformation. The specific expression is:

[0081] log[N(x,y)]=log[G(x,y)]-log[F(x,y)] (12)

[0082] After cleaning the data, the maximum likelihood estimation (MLE) algorithm was used to perform statistical analysis on the noise data, and Gaussian distribution and Gumbel distribution were fitted. By analyzing the fitting curve and histogram of the model, Figure 3 As shown in (b), the Gaussian distribution presents a classic symmetrical bell-shaped curve, while the Gumbel distribution shows a significant skewness and its tail decays slowly. This verifies the consistency between the logarithmic-transformed laser noise data and the Gumbel distribution.

[0083] When constructing the training image dataset later, the training image dataset contains 400 high-depth (16-bit) phase unwrapping images. These high-precision data help improve the algorithm's sensitivity to image noise, thereby improving the accuracy of reconstruction. The image size of the training image dataset is 180×180 pixels. Based on the analysis of the actual laser phase unwrapping image, the pixel value increases from left to right in steps of 30-50, and the pixel step range of the simulated phase unwrapping image is set to Step∈(30,50) to simulate the increasing trend of pixels. At the same time, considering the depth difference of the target object, there will be pixel jumps in the image, and its range is set to Jump∈(-30000, 30000) to reflect the pixel mutation caused by the different target depths. In addition, due to the non-parallelism of the actual target relative to the reference surface, a random number is added to the original phase tilt, and its range is number∈(-100, 100) to simulate the uncertainty of the target phase change. The present invention sets a noise scale parameter range σ∈[0.001, 0.01] to cover various noise levels that may be encountered in actual imaging. The model is used to suppress the Gumbel noise of the phase unwrapped image to achieve more accurate data reconstruction in three-dimensional imaging.

[0084] In a specific embodiment, the image denoising model is a DnCNN model, which includes an input layer, a hidden layer, and an output layer connected in sequence.

[0085] For details, see Figure 4 As shown, the embodiment of the present invention adopts a deep convolutional neural network (Denoising Convolutional Neural Network, DnCNN) to effectively suppress the noise in the phase unwrapping image. The network depth is set to 17, aiming to obtain rich spatial features. The network structure consists of three main parts: the input layer uses a convolution layer and a ReLU activation function to preliminarily extract image features, and configures 64 convolution kernels to generate 64 feature channels. The hidden layer is composed of multiple convolution layers, batch normalization layers and ReLU activation functions to further extract and fuse image features, and all layers use 3×3 convolution kernels. The output layer linearly combines 64 feature channels through the convolution layer to reconstruct the filtered image. Each convolution layer is followed by a ReLU activation function, which introduces nonlinearity to enhance the model's ability to learn complex features. The batch normalization process in the hidden layer effectively reduces the internal covariate shift and accelerates the convergence speed of the model.

[0086] The goal of network training is to learn a residual mapping R(y) (y represents a noisy image) so that the residual image R(y) predicted by the network is close to the real residual image v (v = yt, t is the real image), and finally the ideal image x is obtained through x = yR(y). During the training process, the mean square error (MSE) is used as the loss function to optimize the residual mapping. The specific expression is:

[0087]

[0088] Where M is the number of training samples; v i is the true residual of the i-th phase unwrapped image; R i (y i ) is the residual result predicted by the network.

[0089] See also Figure 5 As shown, an embodiment of the present invention provides a device for using an image denoising method based on statistical analysis as described in any one of the above embodiments, including:

[0090] A data set construction module is used to process and analyze the laser images and phase unwrapping images generated in laser imaging, and to construct a training image data set;

[0091] A model building module, used to build an image denoising model and train the image denoising model using a training image dataset;

[0092] The processing module is used to obtain the actual laser image and perform denoising using the trained image denoising model.

[0093] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0094] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An image denoising method based on statistical analysis, characterized in that: The following steps are involved: Acquire the laser image generated in laser imaging and the corresponding phase unwrapping image for processing and analysis, and construct a training image data set; Constructing an image denoising model, and training the image denoising model using the training image dataset; An actual phase unwrapping image is obtained, and denoised using the trained image denoising model.

2. The image denoising method based on statistical analysis according to claim 1, characterized in that: The specific process of constructing a training image dataset includes: Processing the laser image to obtain a corresponding echo signal; Processing the echo signal to obtain a corresponding noise model; The phase unwrapping image is analyzed in combination with the noise model, and a training image data set is constructed according to the analysis result.

3. The image denoising method based on statistical analysis according to claim 2, characterized in that: The specific process of processing the echo signal includes: Analyze the echo signal to obtain a probability density function corresponding to the amplitude of the echo signal; Determine, according to the probability density function, a probability density function of an additive noise signal included in the echo signal; The phase unwrapping image is fitted, and the fitting result, the laser image and the probability density function of the additive noise signal are logarithmically transformed and subtracted point by point to determine the noise model.

4. The image denoising method based on statistical analysis according to claim 3, characterized in that: The specific process of processing the echo signal also includes: The noise model is verified.

5. The image denoising method based on statistical analysis according to claim 3, characterized in that: The specific process of determining the probability density function of the additive noise signal contained in the echo signal includes: Determining a probability density function of pixel intensities included in the phase unwrapped image according to the probability density function of the amplitude; The probability density function of the original noise is determined, and the probability density function of the signal including the additive noise is determined based on the additive noise model.

6. The image denoising method based on statistical analysis according to claim 1, characterized in that: The image denoising model is a DnCNN model, which includes an input layer, a hidden layer and an output layer connected in sequence.

7. A device using the image denoising method based on statistical analysis according to any one of claims 1 to 6, characterized in that: include: A data set construction module is used to process and analyze the laser images and phase unwrapping images generated in laser imaging, and to construct a training image data set; A model building module, used to build an image denoising model and train the image denoising model using the training image data set; The processing module is used to obtain an actual laser image and perform denoising using the trained image denoising model.

Images