Background Threshold Based on Performance Evaluation and Method for Separating Target and Reverberation Background

Abstract

The present invention relates to a background threshold and separation method for targets and reverberation backgrounds based on performance evaluation, mainly aiming at the problem of threshold selection in background suppression. The statistical idea is applied to the problem of image background suppression. The distribution model and integral means are used to calculate the probabilities of performance indicators FN and FP, and the loss function is designed by using the correlation between FN and FP and the threshold. The parameters of the loss function can control the preference for the expected performance indicators, and the threshold meeting the expectations is obtained. Each pixel sequence has a separate threshold, and a more refined background suppression result can be obtained. The present invention provides a basis and reference for threshold selection and improves the efficiency of image background suppression. The performance indicators are linked to the threshold selection to construct the loss function.

Description

Technical Field

[0001] The present invention belongs to the field of signal processing, and relates to a background threshold based on performance evaluation and a method for separating targets and reverberation backgrounds. It is a method for selecting a background suppression threshold of a port active sonar based on performance evaluation, and a method for separating targets and reverberation backgrounds. Background Art

[0002] In the field of moving target detection and background elimination of images, a widely used processing method is to first obtain the pixel sequence of the image, then establish a background model for each pixel sequence, and finally select the background threshold based on the experience of researchers to achieve background elimination. When the background model is obtained, the selection of the threshold has a great impact on the performance of background elimination. The current method has great subjectivity, resulting in low universality. When facing different data, it is often necessary to reselect the threshold to achieve the expected performance of researchers. On the one hand, it increases the labor cost, and on the other hand, it reduces the processing efficiency. For such disadvantages, there is a need for an image background threshold selection method based on performance evaluation to provide a reference for researchers to select the threshold. Summary of the Invention

[0003] Technical Problems to be Solved

[0004] In order to avoid the deficiencies of the prior art, the present invention proposes a method for separating background thresholds and targets and reverberation backgrounds based on performance evaluation

[0005] Technical Solution

[0006] A method for selecting a background suppression threshold of a port active sonar based on performance evaluation, characterized by the following steps:

[0007] Step 1, distribution fitting: perform statistical fitting on each pixel sequence of an image sequence respectively to obtain the corresponding fitting parameters, including the shape parameter a and the scale parameter b;

[0008] Step 2, calculate the occurrence probability of performance indicators: for the fitting distribution obtained in Step 1, obtain the probabilities of FN and FP through integration, where FN is a false negative, that is, misjudging the target as the background, and the probability is P fn ; FP is a false positive, that is, misjudging the background as the target, and the probability is P fp ;

[0009] Step 3, construct a loss function: construct a loss function about the background threshold through the probabilities of FN and FP, and solve the background threshold corresponding to the minimum value of the loss function:

[0010] Loss function:

[0011]

[0012] Among them, used to control the expected degree of FN and FP indicators, The value of represents that it is acceptable for FN to occur, but FP is not expected to occur; or it represents that it is acceptable for FP to occur, but FN is not expected to occur;

[0013] Step 4: Select the threshold thr corresponding to the minimum loss value in the loss function as the background suppression threshold.

[0014] The statistical fitting in the said step 1 adopts gamma fitting.

[0015] The said step 1 is: draw a histogram of pixel intensities, and select the gamma distribution with the best fitting degree to fit the pixel intensities, and obtain the corresponding shape parameter a and scale parameter b.

[0016] A method for separating a target and a reverberation background by using the obtained background suppression threshold, which is characterized in that: using the background threshold and background model obtained in step 4, calculating the separation boundary between the target and the reverberation background through a threshold method, and realizing the separation of the target and the reverberation background: First, calculate the weighted mean μ and weighted standard deviation σ of each pixel sequence, and set the threshold thr; then, the data falling within the interval [μ - thr * a, μ + thr * σ] is classified as background data, and the data falling outside this interval is classified as target data, and the separation result of the target and the background is obtained; finally, the background data in the separation result is suppressed or eliminated, and the final target image or image sequence is output.

[0017] Beneficial effects

[0018] A method for background threshold and separation of target and reverberation background based on performance evaluation proposed by the present invention mainly aims at the problem of threshold selection in background suppression. Applying statistical thinking to the problem of image background suppression, using distribution models and integral means to calculate the probabilities of performance indicators FN and FP, and designing a loss function using the association between FN and FP and the threshold. The parameters of the loss function can control the preference for expected performance indicators, and the threshold meeting the expected performance is obtained. Each pixel sequence has a separate threshold, and a more refined background suppression result can be obtained. The present invention provides a basis and reference for threshold selection, and improves the efficiency of image background suppression. Linking performance indicators and threshold selection to construct a loss function.

[0019] The technical effect of the present invention is as follows: A background threshold selection method based on performance evaluation is proposed, providing a more objective standard and reference for threshold selection. Since this method constructs a loss function through design, establishing the relationship between the performance metrics FN and FP and the background threshold, it can adjust the preference for the performance metrics of the expected background suppression result, and give the optimal threshold under the current performance preference, reducing the subjectivity of threshold selection and improving the performance of background suppression. The effect is as Figure 5 shown. Brief Description of the Drawings

[0020] Figure 1 is a flowchart of an image background threshold selection method based on performance evaluation.

[0021] Figure 2 is the experimental result of the statistical model of the target data in step 1.

[0022] Figure 3 is the experimental result of the statistical model of the reverberation data in step 1.

[0023] Figure 4 is the curve of the loss function under different parameter values in step 3.

[0024] Figure 5 is a comparison diagram of the background threshold selection method based on performance evaluation. Among them, (a) is the background suppression result without using this method, and (b) is the background suppression result using this method. The interfering pixels are further reduced, increasing the relative intensity of the target pixels. Detailed Embodiment

[0025] The present invention will be further described in combination with embodiments and drawings:

[0026] (1) Distribution fitting. For an image sequence, first perform statistical fitting on the target pixel intensity distribution and the background pixel intensity respectively. In the example, we draw the histogram of pixel intensity and select the gamma distribution with the best fitting degree to fit the pixel intensity, obtaining the corresponding shape parameter a and scale parameter b.

[0027] (2) Calculate the occurrence probability of performance metrics. After obtaining the fitting distribution of pixel intensity, calculate the occurrence probabilities of the performance metrics FN and FP through integration. FN represents false negative, that is, misjudging the target as the background; FP represents false positive, that is, misjudging the background as the target. Given a background model, it is easy to obtain its mean μ and standard deviation σ. We introduce a threshold thr. If a pixel value x falls within the interval [μ - thr * σ, μ + thr * σ], then this pixel value is judged as a background pixel, otherwise it is a target. Therefore, the probability P of FN occurring fnThe probability that the pixel value that should have been the target is misclassified as the background, that is, the probability that the target pixel value falls within the above interval; the probability of FP occurrence, P fp The probability that the pixel value that should have been the background is misclassified as the target, that is, the probability that the background pixel value falls outside the above interval. P fn and P fp Both are obtained by integration and are functions of thr.

[0028] (3) Construct the loss function. After calculating P fn and P fp , we designed the following loss function:

[0029]

[0030] where is used to control the expected degree of the FN and FP metrics. When is small, it means that the occurrence of FN can be accepted, and the occurrence of FP is not desired; when is large, it means that the occurrence of FP can be accepted, and the occurrence of FN is not desired. For each pixel sequence, we select the threshold value thr corresponding to the minimum loss value from its loss function and use this value for background suppression or provide a reference for researchers.

[0031] (4) Separate the target and reverberation background. For the obtained background model, separate the target and reverberation background according to its statistical characteristics and the threshold value thr obtained in the previous step. After the background model is obtained, separate the target data and background data of each pixel sequence. First, calculate the weighted mean μ and weighted standard deviation σ of each pixel sequence, and set the threshold thr; then, the data falling within the interval [μ - thr*σ, μ + thr*σ] is classified as background data, and the data falling outside this interval is classified as target data to obtain the separation result of the target and background; finally, suppress or eliminate the background data in the separation result and output the final target image or image sequence.

Claims

1. A method for selecting the background suppression threshold of a port active sonar based on performance evaluation, characterized in that The steps are as follows: Step 1, distribution fitting: perform statistical fitting on each pixel sequence of an image sequence respectively to obtain corresponding fitting parameters, including shape parameter a and scale parameter b; Step 2, Calculate the occurrence probability of performance metrics: For the fitted distribution obtained in Step 1, the probabilities of FN and FP are obtained through integration, where FN is false negative, that is, the target is misjudged as background, and the probability is P fn ; FP is false positive, that is, the background is misjudged as the target, and the probability is P fp ; Step 3, construct a loss function: construct a loss function regarding the background threshold through the probabilities of FN and FP, and solve for the background threshold corresponding to the minimum value of the loss function: Loss function: Among them, used to control the expected levels of FN and FP metrics, whose value indicates that FN occurrence is acceptable while FP occurrence is not desired; or that FP occurrence is acceptable while FN occurrence is not desired. Step 4: Select the threshold value thr corresponding to the minimum loss value in the loss function as the background suppression threshold; The statistical fitting in Step 1 adopts gamma fitting; Step 1 is: draw a histogram of pixel intensities, and select the gamma distribution with the best fitting degree to fit the pixel intensities to obtain the corresponding shape parameter a and scale parameter b.

2. A method for separating a target and a reverberation background by using the background suppression threshold obtained in claim 1, characterized in that: Using the background threshold and background model obtained in Step 4, calculate the separation boundary between the target and the reverberation background through a threshold method to achieve the separation of the target and the reverberation background: First, calculate the weighted mean μ and weighted standard deviation σ of each pixel sequence, and set the threshold thr; Then, the data falling within the interval [μ - thr*σ, μ + thr*σ] is classified as background data, and the data falling outside this interval is classified as target data to obtain the separation result of the target and the background; Finally, suppress or eliminate the background data in the separation result and output the final target image or image sequence.

Images