A harbor active sonar echo image background suppression method based on a non-parametric Bayesian model

Abstract

The application relates to a sonar echo image background suppression method based on a non-parametric Bayesian model, and mainly aims at the background suppression problem of sonar echo image target detection in a port environment. Machine learning is applied to the sonar echo image background suppression problem, a non-parametric Bayesian model is used for background modeling of the sonar echo image, and statistical characteristics and a threshold method are used for suppressing the background reverberation of the image. Compared with a parameter method, the non-parametric Bayesian model can automatically select the optimal component number required for establishing a mixed model according to the actual distribution of data samples, and solves the overfitting or underfitting problem caused by unknown mixed components of a traditional parameter model. Meanwhile, since the non-parametric Bayesian method can automatically infer the statistics required for establishing a mixed model from data, the dependence on prior information of actual data before modeling is further reduced, and the interference of background reverberation in the sonar echo image on target detection is significantly reduced in the suppression effect.

Description

Technical Field

[0001] The invention belongs to the field of signal processing and relates to a sonar echo image background suppression method based on a non-parametric Bayesian model. Background Art

[0002] Detecting and tracking divers or unmanned underwater vehicles in port environments remains one of the most challenging problems in ocean monitoring. Currently, the most effective detection method is to obtain target echoes from high-frequency active sonar and display the sonar echoes as echo maps reflecting the echo intensity at different ranges and azimuths through beamforming, matched filtering, and normalization methods. These echoes are usually displayed as several pixel areas with different color intensities. Due to the multipath effect, active sonar in coastal environments will cause strong reverberations due to scattering caused by rough bottom structures, bubbles, and fish schools. These reverberations are often confused with the target on the echo map, which has a serious impact on target detection and tracking. Traditional methods mainly rely on background modeling and use parametric methods. They have a strong reliance on prior information of the data, but prior information is difficult to obtain in real environments. Summary of the Invention

[0003] Technical problems to be solved

[0004] To overcome the shortcomings of existing technologies, the present invention proposes a sonar echo image background suppression method based on a nonparametric Bayesian model. This method applies machine learning to the sonar echo image background suppression problem, using the Gaussian-Wishart distribution to establish a background model for the echo image pixel sequence, thereby achieving target and reverberation separation. This method does not rely on prior data information, is simpler to operate, easier to master, and has a wider range of applications.

[0005] Technical Solution

[0006] A sonar echo image background suppression method based on a non-parametric Bayesian model is characterized by the following steps:

[0007] Step 1: Extract pixel sequence: For the image sequence to be processed, extract all data according to the pixel position to obtain the data sequence at the pixel position;

[0008] Step 2: Model building: For each pixel sequence, a Gaussian mixed background model is built using the Gaussian-Wishart distribution as a priori. The variance and mean of each component in the mixed model obey the Gaussian distribution of the Gaussian-Wishart prior.

[0009] Step 3. Train the model: Set parameters and use the Gibbs algorithm to update the model. The number of iterations is set to n. Each iteration updates the probability P(x|k, bg) that the data belongs to different components, indicating the probability that each data belongs to component k. After the iteration is completed, a background model P(x|bg) of the pixel sequence with respect to the data is obtained.

[0010] Step 4: Separate the target and the reverberation background: For the trained background model, select the corresponding mean and standard deviation, and calculate the separation boundary between the target and the reverberation background using the threshold method to achieve separation of the target and the reverberation background.

[0011] When extracting the pixel sequence, the value of a pixel {x, y} over time is regarded as a "pixel sequence", that is, a time series of a scalar grayscale value or a vector of color pixel values; in the pixel {x, y}, the sequence of all pixel values ​​from the first frame to the maximum frame arranged in time order is the pixel sequence, and each pixel corresponds to a pixel sequence.

[0012] The Dirichlet process is introduced into the hybrid model to participate in modeling, and the corresponding weights of the optimal number of components for hybrid modeling are automatically inferred based on the distribution of the data. When a component is given, the probability P(x|k, bg) that a certain data x belongs to the current component k of the background model (bg) is determined by the mean, variance and weight obtained above. If the current number of components is insufficient to fit the data, a new component will be created by the Dirichlet process and the weights of each component will be updated. The tightness of the new component is determined by the Dirichlet process strength parameter. Finally, all the components are mixed together to obtain the background model P(x|bg). At this point, a background model of a pixel sequence is obtained.

[0013] The specific parameters of the modeling process are:

[0014] Input image height 500 Input image width 360 Input image frame number 1393 Maximum pixel intensity 255 Dirichlet process intensity parameter 0.01 Maximum number of Gibbs iterations 1000 Gaussian-Wishart distribution degrees of freedom 5

[0015] In step 4: the weighted mean mu and weighted standard deviation sigma of each pixel sequence are calculated, and a threshold thr is set; then, the data falling within the interval [mu-thr*sigma, mu+thr*sigma] is classified as background data, and the data falling outside the interval is classified as target data, thereby obtaining a target and background separation result; finally, the background data in the separation result is suppressed or eliminated, and the final result is presented in the form of an image sequence.

[0016] Beneficial effects

[0017] This paper proposes a sonar echo image background suppression method based on a nonparametric Bayesian model, specifically targeting the background suppression problem in sonar echo image target detection in port environments. By applying machine learning to the sonar echo image background suppression problem, the paper uses a nonparametric Bayesian model to model the background of the sonar echo image and utilizes statistical properties and a thresholding method to suppress the image's background reverberation. This method significantly reduces the interference of background reverberation compared to existing sonar image processing methods.

[0018] The technical effect of the present invention is that it proposes a new non-parametric Bayesian background model for sonar echo images. Since it adopts a non-parametric method, compared with the parametric method, it can automatically select the optimal number of components required to establish a mixed model according to the actual distribution of data samples, thereby solving the overfitting or underfitting problem of the traditional parametric model caused by the unknown mixed components. At the same time, since the non-parametric Bayesian method can automatically infer the statistics required to establish a mixed model from the data, the dependence on the prior information of the actual data before modeling is further reduced, and the interference of the background reverberation in the sonar echo image on the target detection is significantly reduced in terms of the suppression effect. Figure 3 shown. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Background suppression flow chart for sonar echo images.

[0020] Figure 2 Schematic diagram of the pixel sequence of sonar echo image.

[0021] Figure 3 This is the background suppression effect diagram introduced in this article, where (a) is the cumulative image of multiple frames before processing. The trajectory within the red box is the trajectory of the moving target, and the trajectory outside the box is interference; (b) is the background suppression image after processing, in which the interference is significantly reduced and the target trajectory is more prominent. DETAILED DESCRIPTION

[0022] The present invention will now be further described with reference to the embodiments and accompanying drawings:

[0023] (1) Extract pixel sequence. The value of a specific pixel over time is considered as a "pixel sequence", that is, a time series of a scalar grayscale value or a vector of color pixel values. For a specific pixel {x, y}, the known information at any moment is its history. For example, in pixel {x, y}, the sequence of all pixel values ​​from the first frame to the maximum frame arranged in time order is the pixel sequence, and each pixel corresponds to a pixel sequence, such as Figure 1 The value of each pixel represents a measure of the radiation in the sensor direction of the first object that intersects the pixel's acoustic wave.

[0024] (2) Model building. This step operates on the pixel sequence and builds a model for each pixel sequence. Specifically, we build a Gaussian mixture model for each pixel sequence, but unlike traditional methods, we use a non-parametric model method so that the model can autonomously learn the number of components of the mixture model from the data. In specific implementation, we initialize a mixture model for each pixel sequence, where the variance and mean of each component obey the Gaussian-Wishart distribution, rather than being directly calculated from the data, which reduces the model's dependence on prior information of the data. In order to enable the model to autonomously learn the number of components of the mixture model from the data, we introduce the Dirichlet process to participate in the modeling, which can automatically infer the corresponding weights of the optimal number of components of the mixture model based on the distribution of the data. When a component is given, the probability P(x|k, bg) that a certain data x belongs to the current component k of the background model (bg) is determined by the mean, variance and weight obtained above. If the current number of components is not enough to fit the data, a new component will be created by the Dirichlet process and the weights of each component will be updated. The tightness of the creation of the new component is determined by the Dirichlet process strength parameter. Finally, all the components are mixed together to obtain the background model P(x|bg). At this point, a background model of a pixel sequence is obtained.

[0025] Table 1: Specific parameters of the modeling process

[0026] Input image height 500 Input image width 360 Input image frame number 1393 Maximum pixel intensity 255 Dirichlet process intensity parameter 0.01 Maximum number of Gibbs iterations 1000 Gaussian-Wishart distribution degrees of freedom 5

[0027] (3) Model training. After the initial model is established, the Gibbs algorithm is used to update the model. The number of iterations is set to n = 1000. Each iteration will produce a new probability P(x|k, bg), which represents the probability that each data point belongs to component k. After the iteration, a background model P(x|bg) of the pixel sequence with respect to the data is obtained. The above operation is performed for each pixel sequence, and there are as many background models as there are pixel sequences.

[0028] (4) Separate the target and reverberation. For the background model obtained, separate the target and reverberation according to its statistical characteristics. After the background model is obtained, separate the target data and background data of each pixel sequence. First, calculate the weighted mean mu and weighted standard deviation sigma of each pixel sequence, and set the threshold thr; then, the data falling within the interval [mu-thr*sigma, mu+thr*sigma] is classified as background data, and the data falling outside the interval is classified as target data, and the target and background separation results are obtained; finally, the background data in the separation result is suppressed or eliminated, and the final result is presented in the form of an image sequence.

Claims

1. A sonar echo image background suppression method based on a nonparametric Bayesian model, characterized in that Here are the steps: Step 1: Extract pixel sequence: For the image sequence to be processed, extract all data according to the pixel position to obtain the data sequence at the pixel position; Step 2: Model building: For each pixel sequence, a Gaussian mixed background model is built using the Gaussian-Wishart distribution as a priori. The variance and mean of each component in the mixed model obey the Gaussian distribution of the Gaussian-Wishart prior. Step 3: Train the model: Set parameters and use Gibbs algorithm to update the model. The number of iterations is times, and each iteration updates the probability that the data belongs to different components , indicating that each data belongs to a component The probability of; after the iteration is completed, a background model of the pixel sequence about the data is obtained ; Step 4: Separate the target and the reverberation background: For the trained background model, select the corresponding mean and standard deviation, and calculate the separation boundary between the target and the reverberation background using the threshold method to achieve separation of the target and the reverberation background.

2. The sonar echo image background suppression method based on the nonparametric Bayesian model according to claim 1, characterized in that: When extracting pixel sequences, a pixel is extracted over time. The values ​​of are considered as "pixel sequences", that is, time series of scalar grayscale values ​​or vectors of color pixel values; In the image, the sequence of all pixel values ​​from the first frame to the maximum frame arranged in time order is the pixel sequence, and each pixel corresponds to a pixel sequence.

3. The sonar echo image background suppression method based on the nonparametric Bayesian model according to claim 1, characterized in that: The Dirichlet process is introduced into the mixture model to participate in the modeling, and the corresponding weight of the optimal number of components for the mixture modeling is automatically inferred according to the distribution of the data. When a component is given, a certain data Belong to the background model ( ) Current component Probability It is determined by the mean, variance and weight obtained above; if the current number of components is not enough to fit the data, a new component will be created by the Dirichlet process and the weight of each component will be updated. The tightness of the new component is determined by the Dirichlet process intensity parameter; finally, all the components are mixed together to obtain the background model , so far, a background model of a pixel sequence is obtained.

4. The sonar echo image background suppression method based on the non-parametric Bayesian model according to claim 3 is characterized by: The modeling process parameters are as follows: setting the input image height to 500, setting the input image width to 360, setting the input image frame number to 1393, setting the maximum pixel intensity to 255, setting the Dirichlet process intensity parameter to 0.01, setting the maximum number of Gibbs iterations to 1000, and setting the Gaussian-Wishart distribution degrees of freedom to 5.

5. The sonar echo image background suppression method based on the non-parametric Bayesian model according to claim 1, characterized in that: In step 4: calculate the weighted mean of each pixel sequence and weighted standard deviation , set the threshold ; Then, it falls in the interval [ ] is classified as background data, and the data outside the interval is classified as target data, and the target and background separation results are obtained; finally, the background data in the separation results are suppressed or eliminated, and the final result is presented in the form of an image sequence.

Images