A method for detecting small underwater static targets based on feature fusion

Abstract

This invention relates to the field of underwater static target detection, and particularly to a method for detecting small underwater static targets based on feature fusion. The method includes: preprocessing, segmenting, and post-processing the echo data of a sonar image sequentially to obtain potential target regions; extracting Hu moment features from the potential target regions, and extracting CNN depth features from the potential target regions using a convolutional neural network; concatenating and fusing the extracted Hu moment features and CNN depth features to obtain the target classification result. This method can extract the vast majority of targets from the sonar image while preserving the integrity of the targets to the maximum extent; it fuses the low-level shape features (Hu moments) and CNN depth features of the targets to form a feature fusion network with stronger shape feature representation capabilities for classification, effectively improving the accuracy of target detection; this invention is applicable to underwater static small target detection under different imaging environments and has a wide range of application scenarios.

Description

Technical Field

[0001] This invention relates to the field of underwater static target detection, and in particular to a method for detecting small underwater static targets based on feature fusion. Background Technology

[0002] Underwater static small target detection is one of the hot research directions in the field of underwater acoustics. It effectively identifies and locates sonar images through autonomous detection algorithms. This research has also been widely applied in many aspects such as underwater salvage and search and rescue, oil exploration, and detection of suspicious objects. It has important research significance and application value for both improving my country's national defense science and technology capabilities and tapping marine resources.

[0003] When large amounts of sonar data are used for underwater static small target detection, traditional manual interpretation methods are inefficient and insufficient to meet mission requirements. How to achieve autonomous and efficient underwater static small target detection has become one of the urgent problems to be solved.

[0004] The complex and variable underwater imaging environment, the limited number of prior sonar data samples, and the small size of targets compared to the seabed background make it difficult to extract expressive features, thus exacerbating the difficulty of detecting small, static underwater targets. While deep features extracted by Convolutional Neural Networks (CNNs) are implicit features with rich expressive power, their multi-layered convolutional and pooling structures can easily lead to the loss of some low-level feature information when used for the detection and classification of small, static underwater targets, often resulting in missed detections and false positives. Summary of the Invention

[0005] The purpose of this invention is to achieve autonomous and accurate underwater static small target detection by proposing a feature fusion-based underwater static small target detection method. This invention addresses the characteristics of underwater static small targets and the requirements of underwater target detection tasks by fusing the shallow shape features (Hu moments) of the target with CNN deep features to achieve autonomous target detection, thereby effectively improving the detection performance of static small targets in different underwater environments.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution.

[0007] This invention proposes a method for detecting small underwater static targets based on feature fusion, the method comprising:

[0008] The echo data of the sonar image is preprocessed, segmented, and postprocessed sequentially to obtain the potential target region;

[0009] Extract the Hu moment features of the potential target region, and then extract the CNN depth features of the potential target region through a convolutional neural network;

[0010] The extracted Hu moment features and CNN deep features are concatenated and fused to obtain the target classification result.

[0011] As an improvement to the above technical solution, the preprocessing includes: performing regional scale correction on the sonar echo data and enhancing the sonar image; the postprocessing includes size constraint on the segmented sonar image.

[0012] As an improvement to the above technical solution, the regional scale correction of the sonar echo data specifically includes:

[0013] According to x′ ij =x ij / a j Regional scale correction is performed on the sonar echo data, where the correction factor a in the j-th column is... j =b j -b min / b max -b min b j Let b be the mean of column j. max and b min Let x' represent the maximum and minimum values ​​of the mean of each column, respectively. ij It is the corrected intensity value in the i-th row and j-th column, x ij Let x be the intensity value in the i-th row and j-th column. ij ∈X, where X is the raw echo data of a sonar image of size M×N, and M and N represent the width and height of the echo data.

[0014] As one of the improvements to the above technical solution, the enhancement of the sonar image specifically includes: transforming the echo data of the sonar image using segmented gamma correction;

[0015] The transformed echo value I′ is calculated as follows:

[0016]

[0017] Where γ1 represents the degree of stretching for echo values ​​with CDF > 95.44%, γ2 represents the degree of compression for echo values ​​with CDF ≤ 95.44%, CDF represents the cumulative distribution function, I1 represents echo values ​​with CDF > 95.44%, and I2 represents echo values ​​with CDF ≤ 95.44%.

[0018] As an improvement to the above technical solution, the size constraint of the sonar image specifically includes: deleting regions in the segmented sonar image whose shape and size do not meet the detection requirements to obtain the filtered potential target regions.

[0019] As an improvement to the above technical solution, the deletion of regions in the segmented sonar image whose shape and size do not meet the detection requirements is specifically as follows: removing connected regions in the segmented sonar image with more than A data points and less than B data points, where A≥2000 and B≤500.

[0020] As an improvement to the above technical solution, the segmentation specifically includes: segmenting the sonar image using the MRF algorithm based on the Potts model.

[0021] As an improvement to the above technical solution, the convolutional neural network includes: three convolutional layers CONV-1, CONV-2 and CONV-3, three pooling layers MaxPool-1, MaxPool-2 and MaxPool-3, and three fully connected layers FC-1, FC-2 and FC-3.

[0022] As an improvement to the above technical solution, the step of extracting CNN deep features of potential target regions through convolutional neural networks specifically includes:

[0023] The potential target region is resized using bilinear interpolation and then input into a convolutional neural network.

[0024] The first feature map is obtained by extracting features from the input potential target region through convolutional layer CONV-1 and pooling layer MaxPool-1.

[0025] The first feature map is extracted by using a convolutional layer CONV-2 and a pooling layer MaxPool-2 to obtain the second feature map.

[0026] The second feature map is used to extract features from the second feature map to obtain the third feature map.

[0027] Finally, the CNN deep features are output through a fully connected layer FC-1.

[0028] As one improvement to the above technical solution, the step of concatenating and fusing the extracted Hu moment features and CNN deep features to obtain the target classification result specifically includes:

[0029] Construct a fully connected layer FC-4 and input the extracted Hu moment features into the fully connected layer FC-4 for dimension expansion;

[0030] The extended Hu moment features are concatenated and fused with the CNN deep features;

[0031] The spliced ​​and fused features are input into the fully connected layer FC-2 for mapping;

[0032] Finally, the target classification result is obtained through a fully connected layer FC-3.

[0033] The advantages of this invention compared to the prior art are:

[0034] 1) The first stage of region extraction uses a combination of preprocessing, segmentation and postprocessing to extract most of the targets in the acoustic image while preserving the integrity of the targets to the greatest extent.

[0035] 2) The second-stage feature fusion classification network fuses the low-level shape features of the target (Hu moments) and the deep features of the CNN to form a feature fusion network with stronger shape feature representation capabilities for classification, which effectively improves the accuracy of target detection;

[0036] 3) This invention is applicable to underwater static small target detection under different imaging environments and has a wide range of application scenarios. Attached Figure Description

[0037] Figure 1 This is the overall flowchart of the present invention;

[0038] Figure 2 This is a flowchart of the region extraction process;

[0039] Figure 3 It is the structure of a feature fusion network. Detailed Implementation

[0040] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0041] This invention detects small, static underwater targets in two stages: region extraction and potential target region classification. The first stage extracts potential target regions through four steps, including region scale correction, acoustic enhancement, Markov Random Field (MRF) segmentation, and scale constraint. The second stage combines target Hu moment features and CNN deep features to form a feature fusion network with stronger shape feature representation capabilities to classify the potential target regions extracted in the previous stage.

[0042] The specific technical solution of this invention is as follows: Figure 1 The diagram shown is an overall flowchart of the present invention, including:

[0043] S1, Region Extraction:

[0044] 1) Regional scale correction:

[0045] Because sound waves attenuate as they propagate in water, the echo intensity gradually weakens with increasing reception time, making it difficult to extract potential target areas from the far-field acoustic map. Utilizing the slow change in local gain of sonar echo intensity, region-scale correction can ensure a uniform distribution of echo data, facilitating subsequent processing.

[0046] Let the original echo data of the acoustic spectrum of size M×N be X={x ij}, i=1,…,M,j=1,…,N,x ij Let x' be the intensity value in the i-th row and j-th column. ij =x ij / a j Correct the regional data. Where a j =b j -b min / b max -b min b j Let b be the mean of column j. max and b min These represent the maximum and minimum values ​​of the mean for each column, respectively.

[0047] 2) Acoustic enhancement:

[0048] To initially distinguish the target from the seabed, piecewise gamma correction is used to enhance the acoustic image. According to the 2 sigma rule, echo values ​​with a cumulative distribution function (CDF) greater than 95.44% can be considered not to belong to the seabed. Therefore, the enhanced echo value I′ is calculated as follows:

[0049]

[0050] γ1 represents the degree of stretching for echo values ​​with CDF > 95.44%, and γ2 represents the degree of compression for echo values ​​with CDF ≤ 95.44%. Generally, γ1 > 1 and γ2 < 1. In this embodiment, after multiple experiments, it has been verified that γ1 = 1.2 and γ2 = 0.8 are more conducive to subsequent segmentation. I1 and I2 represent echo values ​​with CDF greater than and less than or equal to 95.44%, respectively.

[0051] 3) MRF segmentation:

[0052] The MRF algorithm based on the Potts model was used to segment the acoustic image, with the number of iterations set to 1 and the number of label categories set to 2.

[0053] 4) Dimensional constraints:

[0054] Based on prior knowledge of the target shape and size, regions in the segmented acoustic image that do not meet the detection requirements are deleted. After multiple experiments, connected regions with more than 2000 data points and less than 500 data points in the segmented acoustic image are removed, resulting in the filtered potential target regions. Figure 2 The diagram shown is a flowchart of the region extraction process.

[0055] S2. Potential Target Area Classification:

[0056] This stage classifies potential target regions using a constructed feature fusion classification network. The classification network, based on a basic CNN architecture, extracts seven-dimensional Hu moment features from the potential target regions and concatenates and fuses them at the back end of the feature fusion network to obtain the classification result. The basic CNN consists of three convolutional layers, three pooling layers, and three fully connected layers. The specific classification steps are as follows:

[0057] 1) Basic CNN feature extraction:

[0058] The potential target region is first resized to 64*64 using bilinear interpolation to fit the network input. Then, a 32*32 feature map is obtained through convolutional layer CONV-1 and pooling layer MaxPool-1. Next, a 16*16 feature map is obtained through CONV-2 and MaxPool-2 network layers. Then, an 8*8 feature map is obtained through CONV-3 and MaxPool-3 network layers. Finally, a 256-dimensional CNN deep feature map is output through fully connected layer FC-1. The convolutional kernel size is 3*3 with a stride of 1, the pooling kernel size is 2, and the activation function is ReLU.

[0059] 2) Hu moment feature extraction:

[0060] Extracting Hu moment features from potential target regions, specifically including:

[0061] Let f(x,y), (x=1,…,M, y=1,…,N) be the echo value of a sonar image of size M×N at (x,y). Then its p+q order geometric moment is defined as:

[0062]

[0063] Through geometric moment m pq Construct the p+q order central moments μ pq As shown in the following formula:

[0064]

[0065] Where (x0, y0) are the centroid coordinates of the sonar image, x0 = m 10 / m 00 y0=m 01 / m 00 .

[0066] The p+q-order normalized central moment η of the sonar image pq Defined as:

[0067]

[0068] in,

[0069] Through various linear combinations of the normalized central moment η pq seven Hu moments can be calculated as shown in the following formula:

[0070] H1 = η 20 + η 02

[0071]

[0072] H3 = (η 30 - 3η 12 ) 2 +(3η 21 - η 03 ) 2

[0073] H4 = (η 30 + η 12 ) 2 +(η 21 + η 03 ) 2

[0074] H5 = (η 30 - 3η 12 )(η 30 + η 12 )[(η 30 + η 12 ) 2 - 3(η 21 + η 03 ) 2 +(3η 21 - η 03 )(η 21 + η 03 )[3(η 30 + η 12 ) 2 -(η 21 + η 03 ) 2

[0075] H6 = (η 20 - η 02 )[(η 30 + η 12 ) 2 -(η 21 + η 03 ) 2 +4η[[ID=|103]] 11 (η 30 + η 12 )(η 21 + η 03 )

[0076] H7 = (3η​21 -η 03 )(η 30 +η 12 )[(η 30 +η 12 ) 2 -3(η 21 +η 03 ) 2 ]-(η 30 -3η 12 )(η 21 +η 03 )[3(η 30 +η 12 ) 2 -(η 21 +η 03 ) 2 ]

[0077] The above 7 Hu moments constitute the target's Hu feature vector H = (H1, H2, H3, H4, H5, H6, H7).

[0078] The extracted seven-dimensional Hu moment features of the original potential target region are input into the fully connected layer FC-4 for expansion, resulting in a 48-dimensional feature vector.

[0079] 3) Feature fusion and classification:

[0080] The extended 48-dimensional Hu moment features and 256-dimensional CNN deep features are concatenated and fused at the back end of the classification network to form a 304-dimensional fused feature. This fused feature is then input into the fully connected layer FC-2 and mapped to 128 dimensions. Finally, it passes through the fully connected layer FC-3 to obtain the target classification result. Figure 3 The diagram shows the structure of the feature fusion network.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A feature fusion based method for detecting underwater static small targets, the method comprising: processing echo data of a sonar image in sequence to obtain potential target regions through preprocessing, segmentation and post-processing; the preprocessing comprises: performing region scale correction on the sonar echo data and enhancing the sonar image using piecewise gamma correction; the segmentation adopts a MRF algorithm based on a Potts model; the post-processing comprises: performing size constraint on the segmented sonar image and deleting regions that do not meet the detection requirements in shape and size; extracting Hu moment features of the potential target regions and extracting CNN deep features of the potential target regions through a convolutional neural network; splicing and fusing the extracted Hu moment features and CNN deep features to obtain a target classification result; specifically comprising: constructing a full connection layer FC-4 and inputting the extracted Hu moment features into the full connection layer FC-4 for dimension expansion; splicing and fusing the expanded Hu moment features and the CNN deep features; inputting the spliced and fused features into a full connection layer FC-2 for mapping; and finally obtaining the target classification result through a full connection layer FC-3.

2. The method according to claim 1, wherein, the region scale correction on the sonar echo data specifically comprises: According to x′ ij = x ij / a j The regional scale correction is performed on the sonar echo data, wherein the jth column correction factor a j = b j -b min / b max -b min , b j is the jth column mean value, b max and b min respectively represent the maximum value and the minimum value of each column mean value, x′ ij is the corrected intensity value of the ith row and the jth column, x ij is the intensity value of the ith row and the jth column, x ij ∈ X, and X is the original echo data of the sonar image.

3. The method according to claim 1, wherein, the enhancement of the sonar image using piecewise gamma correction specifically comprises: transforming the echo data of the sonar image using piecewise gamma correction; the transformed echo value I' is calculated as follows: wherein γ1 represents the degree of stretching of the echo value with CDF>95.44%, γ2 represents the degree of compression of the echo value with CDF≤95.44%, CDF represents the cumulative distribution function, I1 represents the echo value with CDF>95.44%, and I2 represents the echo value with CDF≤95.44%.

4. The method according to claim 1, wherein, the size constraint on the segmented sonar image to delete regions that do not meet the detection requirements in shape and size specifically comprises: removing connected regions with more than A data points and less than B data points in the segmented sonar image, A≥2000 and B≤500.

5. The method according to claim 1, wherein, the convolutional neural network comprises: three convolutional layers CONV-1, CONV-2 and CONV-3, three pooling layers MaxPool-1, MaxPool-2 and MaxPool-3, and three full connection layers FC-1, FC-2 and FC-3.

6. The method according to claim 5, wherein, the extraction of the CNN deep features of the potential target regions through the convolutional neural network specifically comprises: inputting the potential target regions after size adjustment through bilinear interpolation into the convolutional neural network; extracting features of the input potential target regions through the convolutional layer CONV-1 and the pooling layer MaxPool-1 to obtain a first feature map; extracting features of the first feature map through the convolutional layer CONV-2 and the pooling layer MaxPool-2 to obtain a second feature map; extracting features of the second feature map through the convolutional layer CONV-3 and the pooling layer MaxPool-3 to obtain a third feature map; finally outputting the CNN deep features through the full connection layer FC-1.

Images