Synthetic aperture sonar image recognition method and system

Abstract

The application provides a synthetic aperture sonar image recognition method and system, which is used in the field of computer vision and comprises the following steps: acquiring a synthetic aperture sonar image; pre-processing the sonar image based on a Real-ESRGAN model of deep learning to obtain a high-quality sonar image after reconstruction; inputting the high-quality sonar image into a SimCLR model; extracting features of the sonar image through a fusion attention module in an encoder of the SimCLR model, wherein the fusion attention module comprises a channel attention mechanism and a spatial attention mechanism; and identifying and classifying the sonar image based on the extracted features through transfer learning. The scheme solves the problems of sample blur and sample scarcity in sonar image recognition and effectively improves the accuracy of sonar image recognition.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, and in particular relates to a synthetic aperture sonar image recognition method and system. Background Technology

[0002] Synthetic Aperture Sonar (SAS) is a primary means of long-range, high-resolution underwater imaging, and small-sample identification based on SAS images has significant application value. SAS, with its high-precision underwater imaging capabilities and high search efficiency, is widely used in various marine applications such as underwater resource exploration and development, military geographic information acquisition, underwater battlefield surveillance and intelligence gathering, and underwater early warning. Because SAS is not limited by narrow beamwidths in the azimuth direction, it can operate at lower frequencies, has a long detection range, and can effectively distinguish mines from surrounding rocks and other unidentified objects in complex seabed environments, providing accurate target information for subsequent mine clearance operations. It is increasingly valued in the field of marine exploration. With the development of computer vision technology, deep learning models can accurately and efficiently identify sonar target images. However, due to limitations such as the sonar imaging mechanism and acquisition costs, SAS images have fewer target pixels and a scarcer sample size compared to optical images.

[0003] Currently, to address the issue of small sample sizes, several approaches have been proposed: increasing the number of target image samples by performing scattering and geometric transformations on sonar images; using a data-driven model to generate realistic side-scan sonar measurement data from environmental input using CGAN; increasing the sample size by fusing optical target images with existing sonar background images; proposing a sample enhancement method for all types of side-scan sonar images involving image transformation, noise addition, and synthetic data generation techniques; and proposing the SGAN model to achieve effective sample enhancement. While these methods—using data augmentation, automatic data expansion, and synthesizing new samples through simulation software or generative models—can increase the number of sonar target image samples, they all rely on the original sonar images. Problems such as blurred sample images, low quality, and indistinct regions of interest remain unresolved, leading to difficulties in effective feature extraction and low recognition accuracy during the identification process.

[0004] To address the issue of insufficient pixels, this paper proposes a method to transfer optical target images into sonar background images using style transfer, resulting in simulated images. It also proposes using optical images as guide images and successfully generating high-quality sonar images using an improved CycleGAN network. Furthermore, a rasterization-based forward-looking sonar image simulator is proposed, combining ray tracing and rasterization techniques while incorporating the effects of sound pressure attenuation and multipath propagation into the model to improve the realism of the simulated sonar images. Finally, a Transformer-based generative adversarial network is proposed for despeccing sonar images to improve their clarity and quality. While these methods can improve sonar image quality by synthesizing pseudo-sonar samples through simulators and using generative adversarial networks to denoise blurred images, certain domain differences still exist in terms of texture and subtle features, failing to alleviate the sample problem under limited-lens conditions. In addition, uneven data distribution and insignificant intra-class differences lead to classification bias in imbalanced sonar datasets, where classes with more samples are more likely to be correctly identified than those with fewer samples. This phenomenon not only weakens the model's generalization performance but also severely impacts its classification accuracy in practical applications. Meanwhile, under conditions of unbalanced sparse data samples, directly using deep convolutional neural networks for sonar image recognition will result in severe overfitting, further leading to poor recognition accuracy. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a synthetic aperture sonar image recognition method and system to solve the problem of low recognition accuracy of classification models caused by the scarcity and ambiguity of samples in current sonar image recognition.

[0006] In a first aspect of the present invention, a synthetic aperture sonar image recognition method is provided, comprising:

[0007] Acquire synthetic aperture sonar images;

[0008] The sonar images are preprocessed using a deep learning-based Real-ESRGAN model to obtain high-quality reconstructed sonar images.

[0009] The high-quality sonar image is input into the SimCLR model, and the sonar image is feature extracted by the fusion attention module in the SimCLR model encoder. The fusion attention module includes a channel attention mechanism and a spatial attention mechanism.

[0010] Sonar images are identified and classified based on extracted features and through transfer learning.

[0011] In a second aspect of the present invention, a synthetic aperture sonar image recognition system is provided, comprising:

[0012] The image acquisition module is used to acquire synthetic aperture sonar images;

[0013] The image reconstruction module is used to preprocess sonar images using a deep learning-based Real-ESRGAN model to obtain high-quality reconstructed sonar images.

[0014] The feature extraction module is used to input the high-quality sonar image into the SimCLR model, and extract features from the sonar image through the fusion attention module in the SimCLR model encoder. The fusion attention module includes a channel attention mechanism and a spatial attention mechanism.

[0015] The image recognition module is used to identify and classify sonar images based on extracted features through transfer learning.

[0016] In a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor executes the computer program to implement the steps of the method as described in the first aspect of the present invention.

[0017] In a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method provided in the first aspect of the present invention.

[0018] In this embodiment of the invention, a super-resolution reconstruction algorithm is introduced by employing the Real-ESRGAN model to achieve pixel expansion and structural completion of blurred targets. A contrastive learning mechanism is used, and the SimCLR model is employed to achieve high-accuracy classification of small sample images after super-resolution reconstruction. At the same time, a fusion attention module is introduced within the contrastive learning framework, thereby reducing the number of classification model parameters while improving the recognition accuracy of sonar images and enhancing the generalization ability of the classification model. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a synthetic aperture sonar image recognition method according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the SimCLR model structure provided in one embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of the structure of a synthetic aperture sonar image recognition system according to an embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0025] It should be understood that the terms "comprising" and other similar expressions in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, or apparatus that includes a series of steps or units and is not limited to the listed steps or units. Furthermore, "first" and "second" are used to distinguish different objects and are not intended to describe a specific order.

[0026] Please see Figure 1 A flowchart illustrating a synthetic aperture sonar image recognition method provided in this embodiment of the invention includes:

[0027] S101. Obtain synthetic aperture sonar images;

[0028] Synthetic aperture sonar is a two-dimensional imaging sonar that utilizes a uniformly linearly moving acoustic array to form a large virtual synthetic aperture, thereby improving the lateral resolution of the sonar. Two-dimensional images acquired by synthetic aperture sonar can be used for target detection and recognition.

[0029] This includes data cleaning of synthetic aperture sonar images, removing duplicate data, processing missing, erroneous and abnormal data, and converting data formats.

[0030] S102. The sonar image is preprocessed using the Real-ESRGAN model based on deep learning to obtain a high-quality reconstructed sonar image.

[0031] Real-ESRGAN (Realistic Enhanced Super-Resolution Generative Adversarial Networks) is a deep learning-based super-resolution reconstruction framework that introduces generative adversarial networks (GANs) into super-resolution reconstruction, making the generated images more visually realistic.

[0032] The Real-ESRGAN model training in this embodiment is divided into two stages: the first stage uses the mean absolute error loss (L1 loss) to train a peak signal-to-noise ratio (PSNR) oriented model, and the second stage uses a combination of L1 loss, perceptual loss, and generative adversarial network (GAN) loss to train Real-ESRGAN.

[0033] The overall loss function is: ;

[0034] In the formula, λ and γ both represent weighting coefficients, and λ=1, γ=0.1;

[0035] ;

[0036] In the formula, Represents the true value. This represents the predicted value, and n represents the sample size.

[0037] ;

[0038] In the formula, This represents the features extracted from the pre-trained network. Representing the features of the target image, Features representing the generated image;

[0039] ;

[0040] Where BCELoss represents the binary cross-entropy loss, and D represents the discriminator. For the output of the generative model, ones represents a matrix of all 1s of the real image labels.

[0041] Optional, Real-ESRGAN model training includes:

[0042] The training set images at a preset ratio are subjected to degradation processing, which includes one or more of blurring, resolution downsampling, random noise, and image compression.

[0043] The degraded image is reconstructed by a generator, and the reconstructed image is mixed with real images in the training set and input into a discriminator. The discriminator judges the mixed image.

[0044] The Real-ESRGAN model is optimized through adversarial training between the generator and the discriminator.

[0045] During training, Real-ESRGAN employs a series of complex image degradation techniques, such as blurring, resolution downsampling, random noise addition, and image compression, to maximize the creation of realistic scene information and reconstruct a low-resolution image set with many distortion factors for training. The n-order model in Real-ESRGAN contains n repeated degradation processes, each following the classical model:

[0046] ;

[0047] In the formula, D represents the entire degradation process, y represents the high-resolution image, and x represents the low-resolution image after degradation.

[0048] Blur operations are typically represented by a linear blur filter convolution. For a Gaussian blur kernel k with a kernel size of 2t+1, the element values ​​are sampled from a Gaussian distribution.

[0049] In the Real-ESRGAN model, the generator's main task is to upscale and restore degraded low-resolution images. After low-quality input data is fed into the generator, the reconstructed high-quality image is fused with its corresponding real image to obtain a more realistic, high-quality image that resembles the actual scene conditions. Finally, the reconstructed high-quality image is fed into the discriminator for discrimination.

[0050] In the Real-ESRGAN model, the spectrally normalized U-Net is used as the discriminator, which excels in reducing artifacts and enhancing detail. The core of the U-Net discriminator is to distinguish which image in an input image pair is real and which is reconstructed by the generator. Through adversarial training between the generator and the discriminator, Real-ESRGAN can continuously optimize its reconstruction algorithm, making the generated high-resolution images increasingly difficult for the discriminator to distinguish, thus gradually approaching or even surpassing the quality level of real images.

[0051] S103. Input high-quality sonar images into the SimCLR model, and extract features from the sonar images through the fusion attention module in the SimCLR model encoder. The fusion attention module includes a channel attention mechanism and a spatial attention mechanism.

[0052] SimCLR (Simple Contrastive Learning of Visual Representations) is a self-supervised contrastive learning framework that learns high-quality representations of images in an unsupervised manner. A SimCLR model typically includes a data augmentation module, an encoder, and a loss function.

[0053] In some embodiments, the encoder is a Resnet-18 network.

[0054] The SimCLR model has a built-in data augmentation module, eliminating the need for additional data augmentation on the dataset. This module randomly selects two different data augmentation methods for each data sample, such as cropping and brightness variation. The encoder incorporates a fusion attention module, which effectively extracts channel and spatial dimension features, suppresses redundant information, and reduces model complexity. Based on the similarity of image feature vectors, a contrast loss is calculated, and the InfoNCE loss function is used to enhance the similarity between similar samples, resulting in a more efficient representation. Figure 2 As shown, after the sonar image reconstructed by the Real-ESRGAN model is input into SimCLR, two different data augmentation methods are randomly selected to process it and obtain two related images. x i 、 x j The two images are fed into the encoder, which converts them into vectors in the same dimensional space. h i and h j Image feature values ​​z can be extracted by fusing the attention module. i 、 z j The feature vectors in the encoder can be input into downstream classification tasks to recognize and classify images based on transfer learning.

[0055] The fusion attention module includes channel attention and spatial attention mechanisms. The channel attention mechanism, also known as Efficient Channel Attention (ECA), inherits the core idea of ​​the SE module (SENet) to improve model performance by modeling the importance of feature channels, while also simplifying and innovating upon it. Through adaptive adjustment of one-dimensional convolution, it intelligently focuses on the channel feature combinations that have the most significant impact on model performance, dynamically adjusting the weights of different channels to strengthen important feature channels and suppress redundant information.

[0056] Spatial Attention Mechanism (SAM) places greater emphasis on the spatial dimension of feature maps, particularly focusing on the interactions between different pixels. This mechanism enables the model to consciously focus on regions rich in key feature information while ignoring regions with lower contribution. By dynamically adjusting the importance of each spatial location in the input image, it not only improves the specificity and effectiveness of feature representation but also effectively reduces interference from irrelevant information. Using SAM with adaptive multi-scale convolutional kernels (kernel sizes of 3, 5, and 7) not only effectively extracts target features at different scales but also enhances the model's adaptability.

[0057] By combining a spatial attention mechanism with adaptive multi-scale convolutional kernels with an efficient channel attention module, it can capture information in key regions and effectively enhance the feature contribution of important channels, thus achieving multi-dimensional attention enhancement. This improves model performance and better addresses the problem of image recognition accuracy in complex backgrounds.

[0058] In the SimCLR model training, model performance is evaluated based on the Top-1 accuracy function and the InfoNCE loss function. Top-1 accuracy represents the proportion of the class with the highest predicted probability that is the correct class.

[0059] Optionally, the encoder of the SimCLR model can be connected to the classification head of an image classification task, the classification head containing one or more fully connected layers for mapping features to a category space.

[0060] The classification head is the part of the model that maps the learned feature representations to class predictions. It is usually located at the end of the model, immediately following the feature extraction layer.

[0061] S104. Based on the extracted features, sonar images are identified and classified through transfer learning.

[0062] Features extracted by the fusion attention module are used for downstream classification tasks on sonar image data through transfer learning. Transfer learning is a machine learning method that applies the model or knowledge of the source task to the target task. In this embodiment, features are extracted using the encoder, and the features extracted from the source domain are directly used in the image classification model of the target domain for image recognition and classification through transfer learning. For example, the encoder of the SimCLR model is connected to the classification head of a mature classification model to achieve target feature classification in the image.

[0063] In this implementation, the Real-ESRGAN model based on deep learning is used to preprocess the sonar images to obtain high-quality sonar images. The reconstructed high-quality sonar images are then used in SimCLR to extract features through the fusion attention module. The sonar images are then identified and classified based on transfer learning, which can effectively improve the model's recognition accuracy and generalization performance.

[0064] To address the image blurring problem, a super-resolution reconstruction algorithm is introduced to achieve pixel augmentation and structural completion of blurred images. To address the scarce sample problem, a few-sample classification mechanism based on contrastive learning is used to achieve high-accuracy classification of small-sample sonar images after super-resolution reconstruction. To address the long-tail distribution problem, a fusion attention module is introduced within the contrastive learning framework to focus on regions of interest, further improving the classification performance of sonar images while reducing model parameters. Therefore, the accuracy of existing few-sample sonar image recognition is improved, effectively increasing recognition accuracy compared to traditional convolutional neural network classification models.

[0065] In some embodiments, during the SimCLR model training process, supervised feature extraction training is first performed on the target data, and then the extracted features are used in the downstream classification task. After obtaining the encoder trained through contrastive learning, the final fully connected layer is replaced with the classification head for the classification task.

[0066] It is understandable that, to verify the feature extraction effect of supervised learning, the convolutional layers of the encoder need to be frozen, i.e., only the classification head is trained, and the convolutional neural network is not trained. Without adding an improved attention mechanism, the encoder-frozen model has an accuracy of 72.1% on the training set, with a relatively high loss on the training set, but its accuracy on the validation set is consistent with the training set and shows an upward trend. This proves that supervised contrastive learning can apply the superior feature extraction capabilities of the upstream task to the downstream transfer learning task, which greatly improves the performance of image classification tasks. The suboptimal performance of the model on the training set is attributed to the frozen convolutional layers, which cannot update parameters, resulting in a large loss in the input classification feature vector. Subsequently, the freeze on the encoder convolutional layers was lifted, and training experiments were conducted. The model achieved an accuracy of 79.2% on the validation set. To further improve the classification accuracy, a hybrid attention mechanism and the SimCLR algorithm were used for feature extraction. On the test set, the hybrid attention mechanism and the SimCLR algorithm were used for target image recognition, achieving an accuracy of 87.5%, further improving the model performance. Overall, all models observed a rapid reduction and stabilization of loss in the initial stage. The second stage may initially exhibit significant loss fluctuations, but these fluctuations will eventually stabilize. Notably, the super-resolution reconstruction + hybrid attention + SimCLR model showed the most stable loss rate and the smallest overall variance in both training stages, indicating robust and consistent training.

[0067] To address the issues of scarce, low-quality, and unevenly distributed sonar image samples, this paper constructs an image recognition method that integrates super-resolution reconstruction (SMR) with supervised contrastive learning (SCLM). This method leverages the superior image processing capabilities of SMR and the small-sample recognition ability of SCLM. It generates datasets of higher quality than the original sonar images, effectively improving the recognition accuracy of sonar images under small-sample conditions. Comparative experiments show that introducing SCLM effectively suppresses model overfitting. Experimental results demonstrate that the recognition accuracy of the fusion of SMR and SCLM is 4.17% higher than that of existing DCNNs, particularly in the task of classifying synthetic aperture sonar images.

[0068] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0069] Figure 3 This is a schematic diagram of a synthetic aperture sonar image recognition system provided in an embodiment of the present invention. The system includes:

[0070] Image acquisition module 310 is used to acquire synthetic aperture sonar images;

[0071] The image reconstruction module 320 is used to preprocess the sonar image using a deep learning-based Real-ESRGAN model to obtain a high-quality reconstructed sonar image.

[0072] The process of preprocessing sonar images using a deep learning-based Real-ESRGAN model to obtain high-quality sonar images includes the following steps:

[0073] The training set images of a preset ratio are subjected to degradation processing, which includes one or more of blurring, resolution downsampling, random noise, and image compression.

[0074] The degraded image is reconstructed by a generator, and the reconstructed image is mixed with real images in the training set and input into a discriminator. The discriminator judges the mixed image.

[0075] The Real-ESRGAN model is optimized through adversarial training between the generator and the discriminator.

[0076] The feature extraction module 330 is used to input the high-quality sonar image into the SimCLR model and extract features from the sonar image through the fusion attention module in the SimCLR model encoder. The fusion attention module includes a channel attention mechanism and a spatial attention mechanism.

[0077] Image recognition module 340 is used to identify and classify sonar images based on extracted features through transfer learning.

[0078] Optionally, the feature extraction module 330 further includes:

[0079] The model evaluation unit is used to evaluate the model performance during SimCLR model training based on the Top-1 accuracy loss function and the InfoNCE loss function.

[0080] The encoder of the SimCLR model is connected to the classification head of the classification task. The classification head contains one or more fully connected layers for mapping features to the category space.

[0081] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and modules described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0082] Figure 4 This is a schematic diagram of an electronic device according to an embodiment of the present invention. The electronic device is used for synthetic aperture sonar image recognition and classification. Figure 4 As shown, the electronic device 4 in this embodiment includes a memory 410, a processor 420, and a system bus 430. The memory 410 includes an executable program 4101 stored thereon. As those skilled in the art will understand, Figure 4 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0083] The following is combined with Figure 4 A detailed introduction to each component of the electronic device:

[0084] The memory 410 can be used to store software programs and modules. The processor 420 executes various functional applications and data processing of the electronic device by running the software programs and modules stored in the memory 410. The memory 410 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device (such as cached data), etc. In addition, the memory 410 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0085] The memory 410 contains an executable program 4101 for interface generation. This executable program 4101 can be divided into one or more modules / units, which are stored in the memory 410 and executed by the processor 420 to perform tasks such as sonar image recognition. Each module / unit can be a series of computer program instruction segments capable of performing a specific function, describing the execution process of the executable program 4101 in the electronic device 4. For example, the executable program 4101 can be divided into functional modules such as an image acquisition module, an image reconstruction module, a feature extraction module, and an image recognition module.

[0086] The processor 420 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 410, and by calling data stored in the memory 410, it performs various functions and processes data, thereby monitoring the overall status of the electronic device. Optionally, the processor 420 may include one or more processing units; preferably, the processor 420 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, application programs, etc., and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor 420.

[0087] The system bus 430 is used to connect various functional components inside the computer, transmitting data, address, and control information. Its type can be, for example, a PCI bus, an ISA bus, or a CAN bus. Instructions from the processor 420 are transmitted to the memory 410 via the bus, and the memory 410 sends data back to the processor 420. The system bus 430 is responsible for data and instruction exchange between the processor 420 and the memory 410. Of course, the system bus 430 can also connect to other devices, such as network interfaces and display devices.

[0088] In this embodiment of the invention, the executable program executed by the processing 420 of the electronic device includes:

[0089] Acquire synthetic aperture sonar images;

[0090] The sonar images are preprocessed using a deep learning-based Real-ESRGAN model to obtain high-quality reconstructed sonar images.

[0091] The high-quality sonar image is input into the SimCLR model, and the sonar image is feature extracted by the fusion attention module in the SimCLR model encoder. The fusion attention module includes a channel attention mechanism and a spatial attention mechanism.

[0092] Sonar images are identified and classified based on extracted features and through transfer learning.

[0093] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0094] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0095] The above-described 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 foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A synthetic aperture sonar image recognition method, characterized in that, include: Acquire synthetic aperture sonar images; The sonar images are preprocessed using a deep learning-based Real-ESRGAN model to obtain high-quality reconstructed sonar images. The high-quality sonar image is input into the SimCLR model, and the sonar image is feature extracted by the fusion attention module in the SimCLR model encoder. The fusion attention module includes a channel attention mechanism and a spatial attention mechanism. Sonar images are identified and classified based on extracted features and through transfer learning.

2. The method according to claim 1, characterized in that, The process of preprocessing sonar images using a deep learning-based Real-ESRGAN model to obtain high-quality reconstructed sonar images includes: The training set images at a preset ratio are subjected to degradation processing, which includes one or more of blurring, resolution downsampling, random noise, and image compression. The degraded image is reconstructed by a generator, and the reconstructed image is mixed with real images in the training set and input into a discriminator. The discriminator judges the mixed image. The Real-ESRGAN model is optimized through adversarial training between the generator and the discriminator.

3. The method according to claim 1, characterized in that, Before inputting the high-quality sonar image into the SimCLR model, the process also includes: In SimCLR model training, model performance is evaluated based on the Top-1 accuracy function and the InfoNCE loss function.

4. The method according to claim 1, characterized in that, Before inputting the high-quality sonar image into the SimCLR model, the process also includes: The encoder of the SimCLR model is connected to the classification head of the classification task, which contains one or more fully connected layers for mapping features to the category space.

5. A synthetic aperture sonar image recognition system, characterized in that, include: The image acquisition module is used to acquire synthetic aperture sonar images; The image reconstruction module is used to preprocess sonar images using a deep learning-based Real-ESRGAN model to obtain high-quality reconstructed sonar images. The feature extraction module is used to input the high-quality sonar image into the SimCLR model, and extract features from the sonar image through the fusion attention module in the SimCLR model encoder. The fusion attention module includes a channel attention mechanism and a spatial attention mechanism. The image recognition module is used to identify and classify sonar images based on extracted features through transfer learning.

6. The system according to claim 5, characterized in that, The process of preprocessing sonar images using a deep learning-based Real-ESRGAN model to obtain high-quality sonar images includes: The training set images at a preset ratio are subjected to degradation processing, which includes one or more of blurring, resolution downsampling, random noise, and image compression. The degraded image is reconstructed by a generator, and the reconstructed image is mixed with real images in the training set and input into a discriminator. The discriminator judges the mixed image. The Real-ESRGAN model is optimized through adversarial training between the generator and the discriminator.

7. The system according to claim 5, characterized in that, The feature extraction module further includes: The model evaluation unit is used to evaluate the model performance during SimCLR model training based on the Top-1 accuracy function and the InfoNCE loss function.

8. The system according to claim 5, characterized in that, Before inputting the high-quality sonar image into the SimCLR model, the process also includes: The encoder of the SimCLR model is connected to the classification head of the classification task, which contains one or more fully connected layers for mapping features to the category space.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of a synthetic aperture sonar image recognition method as described in any one of claims 1 to 4.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed, it implements the steps of a synthetic aperture sonar image recognition method as described in any one of claims 1 to 4.

Images