An unsupervised active sonar small target area extraction method based on information theory

Abstract

The present invention relates to an unsupervised active sonar small target area extraction method based on information theory. According to the idea of a binary tree data structure, the image to be processed is divided into non-overlapping areas according to the principle of maximizing mutual information gain, and the suspected target area, that is, the area with the richest information, is extracted according to the maximum entropy principle. The present invention does not require too much prior information, but follows the relevant knowledge of information theory to realize the extraction of suspected target areas in active sonar echo maps. The effect is: a new perspective of unsupervised target area extraction method based on relevant knowledge of information theory is proposed. Under the premise of a given input image and end condition, the method divides the input image into multiple sub-areas according to the mutual information maximization principle, and extracts the sub-area corresponding to the maximum grayscale entropy as the potential target area, which provides a new idea for target area extraction.

Description

Technical Field

[0001] The present invention belongs to the field of region of interest extraction in target detection, and relates to an unsupervised active sonar small target region extraction method based on information theory. Background Art

[0002] Detecting quiet man-made targets (such as frogmen and UUVs) in active sonar systems remains one of the most challenging problems in ocean surveillance. Typically, signals collected by active sonar systems are processed using techniques such as beamforming and visualized as grayscale images, displaying the echo intensity within the acquisition range. After processing through processes such as background suppression and target tracking, the approximate range of the suspected target can be determined. However, the precise range and category of the target still require further determination through subsequent target detection and recognition.

[0003] In traditional optical image two-stage object detection, the method typically uses a selective search algorithm or RPN to pre-extract the target region, followed by neural network classification. The former requires input of prior information such as color, texture, size, and shape similarity between image regions, while the latter requires a large amount of pre-training data. Therefore, these algorithms are no longer efficient for sonar echograms. Summary of the Invention

[0004] Technical problems to be solved

[0005] In order to avoid the shortcomings of the existing technology, according to the characteristics of active sonar echograms, the present invention proposes an unsupervised active sonar small target area extraction method based on information theory.

[0006] Technical Solution

[0007] An unsupervised active sonar small target area extraction method based on information theory is characterized by the following steps:

[0008] Step 1. Set the end condition and initialization parameters: read the input image for target area extraction; set the sub-region division end condition to the number of sub-regions; create a sub-region storage list and initialize it to the read input image; create a sub-region position information storage list and initialize it to an empty node null;

[0009] The input image is an active sonar echo image;

[0010] The partitioning process uses a binary tree data structure to store partition location information and is initialized to contain only one root node null, the purpose of which is to store the location information of each sub-area division;

[0011] The number of divided sub-regions is greater than or equal to n, and the number of divisions m is log2n;

[0012] Step 2: Sub-region division: Read each element in the sub-region storage list as input and perform sub-region division respectively; during each sub-region division, the binary tree obtains information from the original input, traverses each horizontal and vertical division, finds the division that maximizes the total mutual information, and stores the corresponding partition position information in the sub-region position storage binary tree;

[0013] Step 3, update variables: calculate the sub-region and sub-region position information according to the leaf nodes in the sub-region position storage binary tree, and update the sub-region storage list and the sub-region position information storage list;

[0014] Step 4: Determine whether the end condition is met: Determine whether the number of sub-regions stored in the sub-region storage list is greater than or equal to the number of sub-regions set by the end condition. If so, proceed to the next step of determining the suspected target region; otherwise, return to step 2 to divide the sub-regions.

[0015] Step 5: Determine the suspected target area: Calculate the grayscale entropy of each sub-area in the sub-area storage list and sort them. The sub-area with the maximum grayscale entropy is the target area determined by this method.

[0016] Each division is performed by horizontal division or vertical division.

[0017] The input image img to be processed in the active sonar echo image dataset is read as a grayscale image.

[0018] The length L of the sub-region storage list list_img is used as the number of sub-region divisions, and the variable corresponding to 0 to L sub-region divisions is len=0:L-1.

[0019] The sub-area location information storage list is list_xy, and the first element of list_xy is initialized to [xmin, xmax, ymin, ymax], the purpose of which is to store detailed location information of sub-areas that need to be further divided, and the sub-areas are positioned using the upper left corner coordinates (xmin, ymin) and lower right corner coordinates (xmax, ymax).

[0020] The total mutual information Where i is the sub-region number, i∈[1,2], I i represents the mutual information of the ith subregion, N represents the number of pixels in the subregion before division, and n i represents the number of pixels in the ith sub-region after division, H(X) represents the grayscale entropy of the random variable X, H(X|Y) represents the conditional grayscale entropy of the random variable X under the given random variable Y, and H(Y i ) represents the grayscale entropy of the random variable corresponding to the i-th sub-region.

[0021] The specific operation of the sub-region division process is as follows: set the sub-region division position m, traverse m=1:(H+W), each value of the sub-region division position m represents a sub-region division, when the sub-region division position m<H, it is horizontal division, otherwise it is vertical division; calculate the mutual information I of the sub-region X before division corresponding to each value of the sub-region division position m and the further division of X into two sub-regions Y1 and Y2 m (X,Y), get the minimum mutual information I index (X, Y) corresponds to the partition position index; the position information of the two sub-areas Y1 and Y2 under the partition position index is represented by the corresponding nodes value1 and value2, and the nodes value1 and value2 are stored in the sub-area position storage binary tree tree. The calculation formula of node value1 and value2 is:

[0022]

[0023] value2=-value1

[0024] Among them, index is the minimum mutual information I index (X, Y) corresponds to the division position, and H is the height of the sub-region partimg.

[0025] The problem of distinguishing the division method of the sub-area division process is that if nodes value1 and value2 are floating-point numbers, it means that the sub-area division method is horizontal division, node value1 stores the upper sub-area position information, and node value2 stores the lower sub-area position information; otherwise, it means that the sub-area division method is vertical division, node value1 stores the left sub-area position information, and node value2 stores the right sub-area position information.

[0026] Beneficial effects

[0027] The present invention proposes an unsupervised active sonar small target area extraction method based on information theory. According to the idea of binary tree data structure, the image to be processed is divided into non-overlapping areas according to the principle of maximizing mutual information gain, and the suspected target area, that is, the area with the richest information, is extracted according to the maximum entropy principle. The present invention does not require too much prior information, but follows the relevant knowledge of information theory to realize the extraction of suspected target areas in active sonar echo maps. The effect is: a new perspective of unsupervised target area extraction method based on relevant knowledge of information theory is proposed. Under the premise of a given input image and end condition, this method divides the input image into multiple sub-areas according to the principle of maximizing mutual information, and extracts the sub-area corresponding to the maximum grayscale entropy as the potential target area, which provides a new idea for target area extraction.

[0028] The beneficial effects are reflected in:

[0029] 1. The method disclosed in the present invention adopts a region extraction method that divides an image into subregions based on maximizing mutual information gain and extracts the subregion corresponding to the maximum grayscale entropy as the target region. Compared with the selective search algorithm, it does not require too much prior information about the input image; compared with the RPN as a neural network, it does not require a pre-training process.

[0030] 2. On the active sonar echo map dataset evaluated, the method disclosed in the present invention achieved a good target area extraction effect. The average IOU of the target area extracted by the method and the actual target area corresponding to the manual calibration label reached 0.72. The IOU distribution histogram of the target area extracted by the method and the actual target area corresponding to the manual calibration label on the active sonar echo map dataset is shown in the figure below. Figure 2 shown. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a schematic diagram of the technical solution process of the present invention. To achieve better visualization effect, this figure and Figure 4 The active sonar echo images displayed have all been processed by mapping the grayscale values into pseudo-color in JET mode.

[0032] Figure 2 This is the IOU distribution histogram of the target area extracted by this method and the actual target area corresponding to the manual calibration label on the active sonar echo map dataset. The IOU range is 0 to 1, with a step size of 0.05, and the vertical axis represents the sample frequency within the corresponding IOU range.

[0033] Figure 3 Flowchart of a specific embodiment of the present invention.

[0034] Figure 4 Five comparison images of the target area extracted by this method (red frame) and the actual target area corresponding to the manual calibration label (green frame) are given. DETAILED DESCRIPTION

[0035] The present invention will now be further described with reference to the embodiments and accompanying drawings:

[0036] Step 1: Set the end condition and initialization parameters. In response to the target area extraction problem of the present invention, read the input image img to be processed in the active sonar echo map data set, and the reading method is a grayscale image. In response to the sub-region division problem in the present invention, set the sub-region division end condition as the number of divided sub-regions is greater than or equal to n (the number of divisions m is log2n); create a sub-region storage list list_img and initialize it to the read img, the purpose of which is to store the sub-regions that need to be further divided; read the length L of the sub-region storage list list_img as the number of sub-region divisions, and the variable corresponding to 0 to L sub-region divisions is len=0:L-1; create a sub-region location information storage binary tree tree and initialize it to contain only one root node null, the purpose of which is to store the location information of each divided sub-region; create a sub-region location information storage list list_xy, initialize the first element of list_xy to [xmin, xmax, ymin, ymax], the purpose of which is to store the detailed location information of the sub-region that needs to be further divided, and use the upper left corner coordinates (xmin, ymin) and lower right corner coordinates (xmax, ymax) of the sub-region for positioning.

[0037] Step 2: Divide the subregions. Read the subregion partimg = list_img[len] to be processed in this loop, that is, the element corresponding to the index len in the subregion storage list list_img, and read the height H and width W of the subregion partimg. Set the subregion division method to horizontal division or vertical division. Regarding the physical meaning of the subregion division process, the subregion division process is shown in the channel expression (1), where the random variable X is the subregion before each subregion division (the input image at initialization), and the random variable Y is the output of further dividing the subregion X into two subregions.

[0038] X→Y (1)

[0039] The sub-region location storage binary tree tree obtains information from the sub-region. The total mutual information I(X,Y) it captures can be obtained by adding the mutual information available at the leaf nodes of the tree. The calculation formula is shown in formula (2), where i is the sub-region number, i∈[1,2], I i represents the mutual information of the ith subregion, N represents the number of pixels in the subregion before division, and n i represents the number of pixels in the ith sub-region after division, H(X) represents the grayscale entropy of the random variable X, H(X|Y) represents the conditional grayscale entropy of the random variable X under the given random variable Y, and H(Y i ) represents the grayscale entropy of the random variable corresponding to the i-th sub-region. The grayscale entropy H(α) of the random variable α is calculated as shown in formula (3), where p jis the probability that the grayscale value j (the value range of j is 0 to 255) appears in the random variable α.

[0040]

[0041]

[0042] For the specific operation of the sub-region division process, let the sub-region division position m, traverse m = 1: (H + W), each value of the sub-region division position m represents a sub-region division, and when the sub-region division position m < H, it is horizontal division, otherwise it is vertical division. Calculate the mutual information I of the sub-region X before division and the further division of X into two sub-regions Y1 and Y2 under each value of the sub-region division position m. m (X,Y), get the minimum mutual information I index (X, Y) corresponds to the partition position index. The position information of the two sub-regions Y1 and Y2 under the partition position index is represented by the corresponding nodes value1 and value2, and the nodes value1 and value2 are stored in the sub-region position storage binary tree tree. The calculation formula of the nodes value1 and value2 is shown in formula (4), where index is the minimum mutual information I index (X, Y) corresponds to the partition position, where H is the height of the subregion partimg. Regarding the partitioning method, if nodes value1 and value2 are floating-point numbers, the subregion partitioning method is horizontal. Node value1 stores the upper subregion position information, and node value2 stores the lower subregion position information. Otherwise, the subregion partitioning method is vertical. Node value1 stores the left subregion position information, and node value2 stores the right subregion position information.

[0043]

[0044] value2=-value1

[0045] After the sub-region division process is completed, it is determined whether all sub-regions in the list list_img have been read and processed. If so, step (3) of the sub-region division process is skipped and the next step (4) of variable update is entered; otherwise, the loop variable len is updated to len+1, and step (3) of the sub-region division process is repeated.

[0046] Step 3: Update variables. To update the values of sub-region-related variables, retrieve all leaf nodes in the binary tree tree that stores sub-region locations. Update the sub-region location information storage list list_xy to a list of the calculated sub-region location information for all leaf nodes. Update the sub-region storage list list_img to a list of the calculated sub-region location information for all leaf nodes. Update the length of the sub-region storage list list_img to L.

[0047] Step 4: Determine whether the end condition is met. To determine whether the end condition is met, determine the relationship between the length L of the sub-region storage list list_img and n. If the length L of the sub-region storage list list_img is less than n, return to step (3) of the sub-region division process and loop through step (3) of the sub-region division process L times. The loop variable corresponding to loops 0 to L is len = 0:L-1. Otherwise, proceed to the next step (6) of target region determination.

[0048] Step 5: Determine the target area. In order to determine the target area, the n elements contained in the sub-area storage list list_img represent the n sub-areas obtained by division, and the n elements contained in the sub-area position information storage list list_xy represent the coordinate position information of the n sub-areas. The grayscale entropy h is calculated and sorted for each sub-area partimg in the sub-area storage list list_img, and the sub-area partimg corresponding to the maximum grayscale entropy is obtained. The index i in the sub-area storage list list_img and the sub-area position information storage list list_xy can determine the suspected target area list_img[i] obtained by this method and the position information list_xy[i] of the suspected target area corresponding to the input image img. The comparison figure of some target area results (red box) extracted after processing by this method and the actual target area (green box) corresponding to the manual calibration label is shown in the figure below. Figure 4 shown.

[0049] During implementation, the dataset used in the present invention is an active sonar echo map dataset, which contains a total of 969 pairs of datasets. The dataset consists of an image set and a label set. The image set of this dataset is based on active sonar test data in a real ocean environment. The two-dimensional energy output pseudo-color image is obtained by subjecting the hydrophone array received signal to beamforming, matched filtering and other spatiotemporal processing. It contains a total of 969 sonar echo maps, and each sonar echo map has only one target. In order to evaluate the results of the region extraction method, this method uses the labelImg tool to manually annotate the actual location information of the target area (ground truth) to construct a label set. The label is stored in XML format, and the label content is (x1, y1, x2, y2), where (x1, y1) represents the coordinates of the upper left corner of the target area, and (x2, y2) represents the coordinates of the lower right corner of the target area.

[0050] Figure 2 This is the IOU distribution histogram of the target area extracted by this method and the actual target area corresponding to the manual calibration label on the active sonar echo map dataset. The IOU range is 0 to 1, with a step size of 0.05, and the vertical axis represents the sample frequency within the corresponding IOU range.

Claims

1. An unsupervised active sonar small target area extraction method based on information theory, characterized by Here are the steps: Step 1. Set the end condition and initialization parameters: read the input image for target area extraction; set the end condition of sub-region division to be greater than or equal to n, and the number of divisions m to log2n; create a sub-region storage list and initialize it to the read input image; create a sub-region position information storage list and initialize it to an empty node null; The input image is an active sonar echograph input image; The partitioning process uses a binary tree data structure to store partition location information and is initialized to contain only one root node null, the purpose of which is to store the location information of each sub-area division; Step 2: Sub-region division: Read each element in the sub-region storage list as input and perform sub-region division respectively; during each sub-region division, the binary tree obtains information from the original input, traverses each horizontal and vertical division, finds the division that maximizes the total mutual information, and stores the corresponding partition position information in the sub-region position storage binary tree; Step 3, update variables: calculate the sub-region and sub-region position information according to the leaf nodes in the sub-region position storage binary tree, and update the sub-region storage list and the sub-region position information storage list; Step 4: Determine whether the end condition is met: Determine whether the number of sub-regions stored in the sub-region storage list is greater than or equal to the number of sub-regions set by the end condition. If so, proceed to the next step of determining the suspected target region; otherwise, return to step 2 to divide the sub-regions. Step 5: Determine the suspected target area: Calculate the grayscale entropy of each sub-area in the sub-area storage list and sort them. The sub-area with the maximum grayscale entropy is the target area determined by this method. The specific operation of the sub-region division process is as follows: let the sub-region division position m, w be the width of the sub-region partimg, traverse m=1:(H+W), each value of the sub-region division position m represents a sub-region division, when the sub-region division position m<H, it is horizontal division, otherwise it is vertical division; calculate the mutual information I of the sub-region X before division corresponding to each value of the sub-region division position m, and the mutual information I of further dividing X into two sub-regions Y1 and Y2 m (X,Y), get the minimum mutual information I index (X, Y) corresponds to the partition position index; the position information of the two sub-areas Y1 and Y2 under the partition position index is represented by the corresponding nodes value1 and value2, and the nodes value1 and value2 are stored in the sub-area position storage binary tree tree. The calculation formula of node value1 and value2 is: value2=-value1 Among them, index is the minimum mutual information I index (X, Y) corresponds to the partition position, H is the height of the sub-region partimg; The problem of distinguishing the division method of the sub-area division process is that if nodes value1 and value2 are floating-point numbers, it means that the sub-area division method is horizontal division, node value1 stores the upper sub-area position information, and node value2 stores the lower sub-area position information; otherwise, it means that the sub-area division method is vertical division, node value1 stores the left sub-area position information, and node value2 stores the right sub-area position information.

2. The unsupervised active sonar small target area extraction method based on information theory according to claim 1 is characterized in that: Each division is performed by horizontal division or vertical division.

3. The unsupervised active sonar small target area extraction method based on information theory according to claim 1 is characterized in that: The input image img to be processed in the active sonar echo map dataset is read as a grayscale image.

4. The unsupervised active sonar small target area extraction method based on information theory according to claim 1 is characterized in that: The length L of the sub-region storage list list_img is used as the number of sub-region divisions, and the variable corresponding to 0 to L sub-region divisions is len=0:L-1.

5. The unsupervised active sonar small target area extraction method based on information theory according to claim 1 is characterized in that: The sub-area location information storage list is list_xy, and the first element of list_xy is initialized to [xmin, xmax, ymin, ymax], the purpose of which is to store detailed location information of sub-areas that need to be further divided, and the sub-areas are positioned using the upper left corner coordinates (xmin, ymin) and lower right corner coordinates (xmax, ymax).

6. The unsupervised active sonar small target area extraction method based on information theory according to claim 1 is characterized by: The total mutual information Where i is the sub-region number, i∈[1,2], I i represents the mutual information of the ith subregion, N represents the number of pixels in the subregion before division, and n i represents the number of pixels in the ith sub-region after division, H(X) represents the grayscale entropy of the random variable X, H(X|Y) represents the conditional grayscale entropy of the random variable X under the given random variable Y, and H(Y i ) represents the grayscale entropy of the random variable corresponding to the i-th sub-region.

Images