Underwater small target feature extraction and classification identification method based on active sonar

Abstract

The invention relates to the field of underwater small target recognition, and discloses an underwater small target feature extraction and classification recognition method based on an active sonar, which comprises the following steps: S1, obtaining a sonar image, screening and sorting the sonar image, and after the sonar image is classified, carrying out median filtering noise reduction and data enhancement on the image to form a sonar image data set; step S2, constructing a sonar image target detection model YOLOv8s network model; s3, according to the established YOLOv8s network model, pre-training the model by using an optical image data set in a style conversion and transfer learning mode to obtain a basic parameter set of the model, and inputting a sonar image data set into the model for training to obtain a trained detection model; and S4, inputting a to-be-detected sonar image into the trained detection model to obtain a sonar image target detection result. The method provided by the invention facilitates the recognition process of an underwater small target.

Description

Technical Field

[0001] The present invention relates to the field of underwater small target recognition, and in particular to an underwater small target feature extraction and classification recognition method based on active sonar. Background Art

[0002] Active sonar is the main way to detect small underwater targets (such as frogmen, mines, etc.), and is of key significance for the protection of the underwater parts of offshore drilling platforms, important islands and ports. To achieve this goal, target classification and recognition technology plays a core role in active sonar signal processing. However, compared with other environments, the marine environment is more complex and particular for the propagation of acoustic signals, which makes active sonar face many challenges in the field of target detection. The presence of numerous scatterers in the ocean produces reverberation interference, which reduces the resolution of sonar signals and increases the difficulty of target identification.

[0003] The purpose of sonar image target detection is to extract the target area from the complex seabed reverberation background, which is a key step in image analysis. Only on the basis of accurate detection of sonar image targets can the features of underwater targets be extracted and parameters measured, making higher-level sonar image analysis and recognition possible. However, due to the complexity of the underwater acoustic field environment and the nonlinearity of sonar equipment imaging, the collected underwater sonar images have the characteristics of low contrast, poor imaging quality, and severe noise pollution. Since small underwater targets such as frogmen are small, the signal-to-noise ratio is very low, and they are often submerged in background noise, and the target brightness varies greatly at different distances, most filtering methods and detection methods are not satisfactory. Summary of the invention

[0004] The purpose of the present invention is to provide a method for extracting and classifying underwater small targets based on active sonar to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for extracting and classifying underwater small targets based on active sonar, comprising the following steps:

[0006] Step S1, obtaining sonar images and screening and sorting them, and after classification, performing median filtering to reduce noise and data enhancement on the images to form a sonar image data set;

[0007] Step S2, constructing a YOLOv8s network model for sonar image target detection model. The YOLOv8s model adds a high-level screening feature fusion pyramid to the Neck part of the network on the basis of the YOLOv8 network, introduces an MP module to replace the convolutional layer for downsampling; and introduces an ELA-SFF module to the feature fusion network for feature fusion;

[0008] Step S3, according to the established YOLOv8s network model, the model is pre-trained by using the optical image dataset through style conversion plus transfer learning to obtain the basic parameter set of the model, and then the sonar image dataset is input into the model for training to obtain a trained detection model;

[0009] Step S4, inputting the sonar image to be detected into the trained detection model to obtain the sonar image target detection result.

[0010] Preferably, the step S1 performs median filtering and denoising on the image, including: for pixels in the image, obtaining the grayscale values ​​of their neighborhood pixels and arranging them in order as a group, and storing the median of the group to replace the original pixel value; after the sonar image data set is annotated in yolo format, data enhancement is performed on the data set, including random cropping, rotation / affine transformation, random rotation, scaling, translation, perspective transformation and Mosaic enhancement.

[0011] Preferably, the high-level screening feature fusion pyramid includes a feature selection module and a feature fusion module. First, feature maps of different scales undergo a screening process in the feature selection module, and then, through a selective feature fusion mechanism, high-level and low-level information in these feature maps are synergistically integrated.

[0012] Preferably: the first branch of the MP module first passes through a maximum pooling layer for downsampling, and then passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number; the second branch first passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number, and then passes through a convolution layer with a convolution kernel of 3*3 and a step size of 2 for downsampling; finally, the results of the first branch and the second branch are combined to obtain the downsampled result.

[0013] Preferably, the ELA-SFF module filters the necessary semantic information contained in the low-scale features by using the high-level features as weights. A given high-level feature and a low-scale feature are input. The high-level feature is first expanded using a transposed convolution (T-Conv) of size 2 and a kernel size of 3×3 to obtain the feature size. Then bilinear interpolation is used to up- or down-sample the high-level features. The ELA module is then used to convert the high-level features into corresponding attention weights. Finally, the filtered low-scale features are fused with the high-level features.

[0014] As a preferred embodiment, the YOLOv8s network model processes the image as follows: after the image is input, it passes through two convolutional layers, an ELA-SFF module and a convolutional layer, and then passes through a high-level screening feature fusion pyramid to obtain the first feature map P1, and then passes through the MP module and the high-level screening feature fusion pyramid to obtain the second feature map P2, and finally passes through the MP and the high-level screening feature fusion pyramid to obtain the feature map P3; the feature map P3 passes through the convolutional layer to obtain the feature map P4, and after P4 is upsampled, it is fused with P2 to obtain the feature Figure T1, then T1 passes through the high-level screening feature fusion pyramid and convolution layer to obtain feature map P5, P5 is upsampled and fused with P1 to obtain feature map T2, then T2 passes through the C2f module to obtain the output feature map D1; D1 passes through the convolution layer and is fused with P5 to obtain feature map T3, T3 passes through the C2f module to obtain the output feature map D2; D2 passes through the convolution layer and is fused with P4 to obtain feature map T4, T4 passes through the C2f module to obtain the output feature map D3; feature maps D1, D2, and D3 are output inference to obtain the final inference result of the model.

[0015] As a preferred embodiment, in step S3, adaptive instance regularization AdaIN is used to perform style conversion, which is expressed as follows:

[0016]

[0017] Where x is the optical image, y is the sonar image, μ and σ are the mean and variance of the pixel values ​​of the image, μ(y) refers to the mean of the sonar image, and σ(y) refers to the variance of the sonar image.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] The present invention applies deep learning to the field of sonar images and reasonably improves the network model. By improving the feature extraction network, features at different levels are effectively fused to optimize the information flow, the accuracy and robustness of the model when processing targets of different sizes and complexities are improved, and more efficient feature transfer is achieved in the network. By improving the feature fusion network, a lower resolution is used in the early stage to reduce the computational burden, and then the resolution is gradually increased to enhance detail recognition, thereby improving accuracy while maintaining efficiency, and applying transfer learning of style conversion. By performing style conversion on the source domain data set in a targeted manner, the dependence of transfer learning on the scale of the data set can be effectively alleviated. The present invention effectively improves the accuracy and efficiency of sonar image target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the method of the present invention;

[0021] Figure 2 This is a diagram showing the enhanced effect of the present invention. DETAILED DESCRIPTION

[0022] 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.

[0023] Example

[0024] See also Figure 1 , the method for feature extraction and classification of underwater small targets based on active sonar shown in the figure includes the following steps:

[0025] Step S1, obtaining sonar images and screening and sorting them, and after classification, performing median filtering to reduce noise and data enhancement on the images to form a sonar image data set;

[0026] Step S2, constructing a YOLOv8s network model for sonar image target detection model. The YOLOv8s model adds a high-level screening feature fusion pyramid to the Neck part of the network on the basis of the YOLOv8 network, introduces an MP module to replace the convolutional layer for downsampling; and introduces an ELA-SFF module to the feature fusion network for feature fusion;

[0027] Step S3, according to the established YOLOv8s network model, the model is pre-trained by using the optical image dataset through style conversion plus transfer learning to obtain the basic parameter set of the model, and then the sonar image dataset is input into the model for training to obtain a trained detection model;

[0028] Step S4, inputting the sonar image to be detected into the trained detection model to obtain the sonar image target detection result.

[0029] In this embodiment, step S1 performs median filtering and denoising on the image, including: for pixels in the image, obtaining the grayscale values ​​of their neighborhood pixels and arranging them in order as a group, and storing the median of the group to replace the original pixel value; after the sonar image data set is annotated in yolo format, data enhancement is performed on the data set, including random cropping, rotation / affine transformation, random rotation, scaling, translation, perspective transformation and Mosaic enhancement.

[0030] In this embodiment, the high-level screening feature fusion pyramid includes a feature selection module and a feature fusion module. First, feature maps of different scales undergo a screening process in the feature selection module, and then, through the selective feature fusion mechanism, the high-level and low-level information in these feature maps are synergistically integrated.

[0031] In this embodiment, the first branch of the MP module first passes through a maximum pooling layer for downsampling, and then passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number; the second branch first passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number, and then passes through a convolution layer with a convolution kernel of 3*3 and a step size of 2 for downsampling; finally, the results of the first branch and the second branch are combined to obtain the downsampled result.

[0032] In this embodiment, the ELA-SFF module filters the necessary semantic information contained in the low-scale features by using high-level features as weights. A given high-level feature and a low-scale feature are input. The high-level feature is first expanded using a transposed convolution (T-Conv) of size 2 and a kernel size of 3×3 to obtain the feature size. Then bilinear interpolation is used to up- or down-sample the high-level features. The ELA module is then used to convert the high-level features into corresponding attention weights. Finally, the filtered low-scale features are fused with the high-level features.

[0033] In this embodiment, the image processing process of the YOLOv8s network model is as follows: after the image is input, it passes through two convolutional layers, an ELA-SFF module and a convolutional layer, and then passes through a high-level screening feature fusion pyramid to obtain the first feature map P1, and then passes through the MP module and the high-level screening feature fusion pyramid to obtain the second feature map P2, and finally passes through the MP and the high-level screening feature fusion pyramid to obtain the feature map P3; the feature map P3 passes through the convolutional layer to obtain the feature map P4, and after P4 is upsampled, it is fused with P2 to obtain the feature map T1, then T1 passes through the high-level screening feature fusion pyramid and the convolution layer to obtain the feature map P5, P5 is upsampled and fused with P1 to obtain the feature map T2, then T2 passes through the C2f module to obtain the output feature map D1; D1 passes through the convolution layer and is fused with P5 to obtain the feature map T3, T3 passes through the C2f module to obtain the output feature map D2; D2 passes through the convolution layer and is fused with P4 to obtain the feature map T4, T4 passes through the C2f module to obtain the output feature map D3; feature maps D1, D2, and D3 are output inference to obtain the final inference result of the model.

[0034] In this embodiment, in step S3, adaptive instance regularization AdaIN is used to perform style conversion, which is expressed as follows:

[0035]

[0036] Where x is the optical image, y is the sonar image, μ and σ are the mean and variance of the pixel values ​​of the image, μ(y) refers to the mean of the sonar image, and σ(y) refers to the variance of the sonar image.

[0037] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0038] The present invention selects the URPC2021 public dataset for experiments. URPC2021 is an underwater sonar dataset, which includes eight categories: human body, ball, round cage, square cage, tire, metal barrel, cube and cylinder. A total of 4000 sonar images are selected, of which 3200 images are used as training sets and 800 images are used as test sets for experiments. In order to verify the effectiveness of the algorithm in this paper, the sonar common target detection dataset (SCTD) is selected to verify the model of the present invention. This dataset has 357 sonar images. Due to the small amount of data, the dataset is randomly enhanced to 2142 images by randomly removing pixels, sharpening, affine transformation, changing brightness, randomizing hues, image flipping, and adding noise. The images are composed of three types of targets: aircraft, human body and ship. The image data enhancement results are shown in the figure. Figure 2 As shown, Figure 2 a is the original image, Figure 2 b is a randomly removed pixel map. Figure 2 c is the sharpening processing image, Figure 2 d is the brightness adjustment map, Figure 2 e is the adjustment tone map, Figure 2 f is the flipped graph.

[0039] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for extracting and classifying underwater small targets based on active sonar, characterized in that: The steps include: Step S1, obtaining sonar images and screening and sorting them, and after classification, performing median filtering to reduce noise and data enhancement on the images to form a sonar image data set; Step S2, constructing a YOLOv8s network model for sonar image target detection model, wherein the YOLOv8s model adds a high-level screening feature fusion pyramid to the Neck part of the network on the basis of the YOLOv8 network, and introduces an MP module to replace the convolutional layer for downsampling; The ELA-SFF module is introduced into the feature fusion network for feature fusion; Step S3, according to the established YOLOv8s network model, the model is pre-trained by using the optical image dataset through style conversion plus transfer learning to obtain the basic parameter set of the model, and then the sonar image dataset is input into the model for training to obtain a trained detection model; Step S4, inputting the sonar image to be detected into the trained detection model to obtain the sonar image target detection result.

2. The method for extracting and classifying underwater small targets based on active sonar according to claim 1 is characterized in that: The step S1 performs median filtering and denoising on the image, including: for pixels in the image, obtaining the grayscale values ​​of their neighborhood pixels and arranging them in order into a group, and storing the median value of the group to replace the original pixel value; after the sonar image data set is annotated in yolo format, data enhancement is performed on the data set, including random cropping, rotation / affine transformation, random rotation, scaling, translation, perspective transformation and Mosaic enhancement.

3. The method for extracting and classifying underwater small targets based on active sonar according to claim 2 is characterized in that: The high-level screening feature fusion pyramid includes a feature selection module and a feature fusion module. First, feature maps of different scales undergo a screening process in the feature selection module, and then, through a selective feature fusion mechanism, high-level and low-level information in these feature maps are synergistically integrated.

4. The method for extracting and classifying underwater small targets based on active sonar according to claim 3 is characterized in that: The first branch of the MP module first passes through a maximum pooling layer for downsampling, and then passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number; the second branch first passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number, and then passes through a convolution layer with a convolution kernel of 3*3 and a step size of 2 for downsampling; Finally, the results of the first branch and the second branch are combined to get the down-sampling result.

5. The method for extracting and classifying underwater small targets based on active sonar according to claim 4 is characterized in that: The ELA-SFF module filters the necessary semantic information contained in low-scale features by using high-level features as weights. A given high-level feature and a low-scale feature are input. The high-level feature is first expanded using a transposed convolution (T-Conv) of size 2 and a kernel size of 3×3 to obtain the feature size. Then bilinear interpolation is used to up- or down-sample the high-level features. The ELA module is then used to convert the high-level features into corresponding attention weights. Finally, the filtered low-scale features are fused with the high-level features.

6. The method for extracting and classifying underwater small targets based on active sonar according to claim 5, characterized in that: The image processing process of the YOLOv8s network model is as follows: after the image is input, it passes through two convolutional layers, an ELA-SFF module and a convolutional layer, and then passes through a high-level screening feature fusion pyramid to obtain the first feature map P1, and then passes through the MP module and the high-level screening feature fusion pyramid to obtain the second feature map P2, and finally passes through MP and the high-level screening feature fusion pyramid to obtain the feature map P3; the feature map P3 passes through the convolutional layer to obtain the feature map P4, and after P4 is upsampled, it is fused with P2 to obtain the feature map T1, Then T1 is filtered through a high-level feature fusion pyramid and convolutional layer to obtain feature map P5. After upsampling, P5 is fused with P1 to obtain feature map T2. Then T2 is passed through the C2f module to obtain the output feature map D1. D1 is fused with P5 through a convolutional layer to obtain feature map T3. T3 is passed through the C2f module to obtain the output feature map D2. D2 is fused with P4 after the convolutional layer to obtain feature map T4. T4 is passed through the C2f module to obtain the output feature map D3. Feature maps D1, D2, and D3 are output inferenced to obtain the final inference result of the model.

7. The method for extracting and classifying underwater small targets based on active sonar according to claim 6, characterized in that: In step S3, adaptive instance regularization AdaIN is used to perform style conversion, which is expressed as follows: Where x is the optical image, y is the sonar image, μ and σ are the mean and variance of the pixel values ​​of the image, μ(y) refers to the mean of the sonar image, and σ(y) refers to the variance of the sonar image.

Images