A High-Resolution Imaging Method, Device, Equipment and Medium

Abstract

The present application relates to the field of sonar imaging technology, and discloses a high-resolution imaging method, device, equipment and medium, which are applied to a neural network model. The neural network model includes an encoding module, an imaging module and a super-resolution module. The method includes: the encoding module performs feature extraction and data dimensionality reduction on the sonar echo signals collected by each channel to obtain a complex envelope data matrix. The sonar echo signal represents the signal received by the sonar and reflected from the target position. The encoding module is trained with the sonar echo signal with added noise as the input and the complex envelope corresponding to the noise-free simulated sonar echo signal at different distances as the label. The imaging module generates an echo intensity imaging result corresponding to the complex envelope data matrix. The super-resolution module enhances the echo intensity imaging result to generate a high-resolution imaging result. The technical solution provided by the present application can improve the accuracy and efficiency of three-dimensional sonar imaging.

Description

Technical Field

[0001] This application relates to the field of sonar imaging technology, and particularly to a high-resolution imaging method, device, equipment and medium. Background Art

[0002] Traditional three-dimensional sonar imaging technology faces problems of low accuracy and slow efficiency. Especially in applications such as deep-sea exploration and ocean resource exploitation, traditional methods are difficult to provide sufficient resolution and real-time processing capabilities. Due to limitations in device performance and computational burden, traditional sonar imaging is difficult to meet the requirements for high-precision and high-efficiency imaging, resulting in poor performance in large-area detection and small-target recognition.

[0003] Therefore, improving the accuracy and efficiency of three-dimensional sonar imaging to meet the requirements of high-demanding applications such as deep-sea exploration is a technical problem that urgently needs to be solved currently. Summary of the Invention

[0004] This application provides a high-resolution imaging method, device, equipment and medium, which achieves the technical effect of improving the accuracy and efficiency of three-dimensional sonar imaging.

[0005] To achieve the above object, the main technical solutions adopted in this application include:

[0006] In a first aspect, an embodiment of this application provides a high-resolution imaging method, which is applied to a neural network model. The neural network model includes an encoding module, an imaging module and a super-resolution module. The method includes:

[0007] The encoding module performs feature extraction and data dimensionality reduction on the sonar echo signals collected by each channel to obtain a complex envelope data matrix. Among them, the sonar echo signal represents the signal received by the sonar and reflected from the target position. The encoding module is trained with the sonar echo signal with added noise as the input and the complex envelope corresponding to the noise-free simulated sonar echo signal at different distances as the label.

[0008] The imaging module generates an echo intensity imaging result corresponding to the complex envelope data matrix.

[0009] The super-resolution module enhances the echo intensity imaging result to generate a high-resolution imaging result.

[0010] A high-resolution imaging method provided by this embodiment extracts features and reduces the dimensionality of sonar echo signals collected by each channel through an encoding module to obtain a complex envelope data matrix, thereby effectively compressing and processing the original signal, removing redundant information, and extracting valuable features. This process provides a concise and efficient data input for subsequent imaging and enhancement processing, improving the efficiency of data processing. Subsequently, the imaging module converts the complex envelope data matrix into an echo intensity imaging result, accurately converting the sonar echo signal into a visual image, laying a foundation for subsequent image enhancement. Finally, the super-resolution module generates a high-resolution imaging result by enhancing the echo intensity imaging result, significantly improving the spatial resolution of the imaging, showing clearer and more accurate target details, and is particularly suitable for complex deep-sea detection environments.

[0011] In one embodiment, the encoding module extracts features and reduces the dimensionality of the sonar echo signals collected by each channel to obtain a complex envelope data matrix, including:

[0012] Use the one-dimensional convolutional layer and pooling layer in the encoding module to extract features from the sonar echo signal of each channel to obtain the temporal features corresponding to the sonar echo signal;

[0013] Flatten the temporal features and perform data dimensionality reduction through the fully connected layer in the encoding module to obtain the complex envelope data matrix.

[0014] In this embodiment, by using the one-dimensional convolutional layer and pooling layer in the encoding module to extract features from the sonar echo signal, the temporal features in the signal can be extracted. The convolutional layer can capture important features in the signal through local perception, while the pooling layer can reduce data redundancy, improve computational efficiency, and enhance the robustness of features, which helps to extract more accurate temporal information from complex sonar echoes. Through the flattening process and the dimensionality reduction operation of the fully connected layer, the dimensionality of the data can be further reduced, simplifying the complexity of subsequent processing. The dimensionality reduction process of the fully connected layer can retain the most important information while removing unnecessary noise, significantly improving the efficiency of data processing. This processing method not only helps to improve the calculation speed and meet the real-time requirements, but also ensures that the processed data can more accurately reflect the characteristics of the sonar echo signal, thereby improving the imaging accuracy, so as to meet the high-precision and high-efficiency imaging requirements such as deep-sea detection.

[0015] In one embodiment, the encoding module is trained in the following manner:

[0016] Obtain the training samples of the encoding module;

[0017] Use the training samples of the encoding module to train the one-dimensional convolutional layer, pooling layer, and fully connected layer in the encoding module to obtain the output of the encoding module;

[0018] Use the coded mean square error as the loss function of the coding module; wherein, the coded mean square error characterizes the difference between the training samples of the coding module and the output of the coding module.

[0019] Adopt the way of supervised learning, and optimize the parameters of the coding module through backpropagation and gradient descent until the loss function of the coding module meets the convergence condition, and obtain the trained coding module.

[0020] In this embodiment, by obtaining the training samples of the coding module, it is ensured that the coding module can learn effective signal features and effectively cope with noise interference. Using the structure of the convolutional neural network (CNN), the coding module can extract local and global features in the sonar signal, thereby improving the signal quality and reducing the influence of noise on the imaging result. Using the mean square error as the loss function further optimizes the output accuracy of the coding module to ensure more accurate signal reconstruction. Through the backpropagation and gradient descent optimization algorithms, the parameters are gradually adjusted, improving the generalization ability of the coding module to ensure stable signal recovery in complex environments. It not only improves the clarity and accuracy of sonar imaging, but also greatly improves the processing efficiency, enabling faster response to dynamic and complex detection tasks such as in the deep sea.

[0021] In one embodiment, the training samples of the coding module are obtained by the following method:

[0022] Simulate the sonar echo signal using the Rayleigh scattering formula to generate a noise-free simulated sonar echo signal.

[0023] Add noise to the noise-free simulated sonar echo signal to generate a noisy simulated sonar echo signal, and use the noisy simulated sonar echo signal as the input of the coding module.

[0024] Extract the complex envelopes of the noise-free simulated sonar echo signals at different distances to generate pure complex envelope data, and use the pure complex envelope data as the labels of the coding module.

[0025] Based on the input of the coding module and the label of the coding module, construct the input and label pairs of the coding module to obtain the training samples of the coding module.

[0026] In this embodiment, a noise-free sonar echo signal is simulated by using the Rayleigh scattering formula, and noise is added thereto to generate a noisy sonar echo signal, thereby simulating the noise interference in the actual environment. Then, the complex envelopes of the noise-free sonar echo signals at different distances are extracted to generate pure complex envelope data, which is used as the label of the encoding module to ensure high-quality extraction of the core features of the signal. Based on the noisy simulated sonar echo signal and the pure complex envelope data, training samples of the encoding module are constructed, enabling the encoding module to effectively extract useful information from the noisy signal, thereby improving the accuracy and efficiency of 3D sonar imaging and meeting the requirements of high-demanding applications such as deep-sea exploration.

[0027] In one embodiment, the extracting the complex envelopes of the noise-free simulated sonar echo signals at different distances to generate pure complex envelope data includes:

[0028] Dividing the far-field imaging space into multiple equidistant planes according to a preset distance resolution; wherein, the far-field imaging space represents the area far from the sonar emission source during the propagation of the sonar emission signal;

[0029] Determining the time series label corresponding to the equidistant plane; wherein, the time series label represents the position of the noise-free simulated sonar echo signal in the time series;

[0030] Based on the time series label, the noise-free simulated sonar echo signal, the center frequency of the sonar emission signal, the sampling time interval, and the time window length, determining the complex envelope corresponding to the noise-free simulated sonar echo signal at the equidistant plane;

[0031] Organizing the complex envelopes into a matrix form to generate the pure complex envelope data.

[0032] In this embodiment, by dividing the far-field imaging space into multiple equidistant planes and determining the time series label, the accuracy of target positioning is effectively improved, and signal aliasing is avoided. Then, based on the noise-free simulated sonar echo signal, the complex envelope is calculated and time-frequency analysis is performed to extract the core information of the echo signal, ensuring the high quality of the signal. Finally, by organizing the complex envelopes into a matrix form, the data processing flow is optimized, the efficiency is improved, and it is particularly suitable for large-scale data processing. It not only improves the imaging resolution but also ensures the fast and accurate positioning of targets in deep-sea exploration, meeting the requirements of high-precision detection.

[0033] In one embodiment, the imaging module generating the echo intensity imaging result corresponding to the complex envelope data matrix includes:

[0034] Process the complex envelope data matrix using the two-dimensional convolutional layer in the imaging module to extract features and generate imaging results for each channel;

[0035] Perform data transformation on the imaging results of each channel to obtain preliminary imaging results;

[0036] Perform a deconvolution operation on the preliminary imaging results using the deconvolution layer in the imaging module to obtain the echo intensity imaging results.

[0037] In this embodiment, features in the complex envelope data are extracted through a two-dimensional convolutional layer to generate imaging results for each channel, which reflect different features of the signal. Then, the imaging results of each channel are subjected to data transformation to obtain preliminary imaging results. Finally, a deconvolution operation is performed on the preliminary imaging results using the deconvolution layer to improve the resolution and details of the image, and ultimately obtain the echo intensity imaging results. By extracting features through the convolutional layer, data transformation, and deconvolution operation, the imaging effect is gradually optimized to help improve the accuracy and clarity of underwater sonar imaging.

[0038] In one embodiment, the imaging module is trained in the following manner:

[0039] Obtain imaging module training samples; wherein, the imaging module training samples include target scatterer positions and delay beamforming results;

[0040] Use the imaging module training samples to train the two-dimensional convolutional layer and deconvolution layer in the imaging module to obtain an imaging module output;

[0041] Construct a total loss function for the imaging module based on imaging mean squared error and imaging cross entropy; wherein, the imaging mean squared error characterizes the difference between the imaging module output and the delay beamforming results, and the imaging cross entropy characterizes the difference between the imaging module output and the target scatterer positions;

[0042] Adopt a supervised learning method to optimize the parameters of the imaging module through backpropagation and gradient descent until the loss function of the imaging module meets the convergence condition to obtain a trained imaging module.

[0043] In this embodiment, by obtaining training samples of the target scatterer positions and the delay beamforming results, high-quality training data is provided for the imaging module. Then, these training samples are used to train the two-dimensional convolutional layer and the deconvolutional layer in the imaging module, thereby optimizing the target scatterer positioning and echo signal matching and enhancing the imaging ability. On this basis, by constructing a comprehensive loss function (including imaging mean square error and imaging cross-entropy), the difference between the output of the imaging module and the actual result is accurately measured, further optimizing the performance of the imaging module. Finally, supervised learning is adopted, and the parameters of the imaging module are optimized through backpropagation and gradient descent methods to ensure that the imaging module provides high-precision three-dimensional sonar imaging in complex environments.

[0044] In one embodiment, the training samples of the imaging module are obtained in the following manner:

[0045] Using simulation technology to simulate the complex envelope signal, generating a noise-free simulated complex envelope signal, and using the noise-free simulated complex envelope signal as the input of the imaging module;

[0046] Using simulation technology, according to the target scatterer position and the geometric structure of the sonar array, determining the delay of the sonar echo signal received by each element;

[0047] According to the delay, performing time alignment on the sonar echo signals received by each element;

[0048] Summing the aligned sonar echo signals to obtain the delay sum beamforming result, and using the target scatterer position and the delay sum beamforming result as the label of the imaging module;

[0049] Based on the input of the imaging module and the label of the imaging module, constructing an input and label pair of the imaging module to obtain the training samples of the imaging module.

[0050] In this embodiment, using the noise-free simulated complex envelope signal as the input effectively reduces noise interference, thereby improving the imaging accuracy. Secondly, accurately calculating the delay of the sonar echo signal and performing time alignment ensure that the signals received by different elements are synchronized, eliminating imaging distortion caused by time differences. On this basis, through the delay sum technology, the beam focusing effect is enhanced and the imaging effect is optimized. In addition, using the noise-free simulated complex envelope signal and accurate labels to generate the training samples of the imaging module supports the rapid learning of the imaging model, improving the training efficiency and the performance of the imaging model. It ensures that in complex environments such as the deep sea, the imaging system can achieve high-precision and high-efficiency three-dimensional sonar imaging, meeting the requirements of high-demand applications.

[0051] In one embodiment, the super-resolution module enhances the echo intensity imaging result to generate a high-resolution imaging result, including:

[0052] Perform a convolution operation on the echo intensity imaging result using the two-dimensional convolution layer in the super-resolution module to obtain a target feature map;

[0053] Use the sub-pixel convolution layer in the super-resolution module to rearrange the target feature map to generate the high-resolution imaging result.

[0054] In this embodiment, through the two-dimensional convolution layer and the sub-pixel convolution layer, the super-resolution module can significantly improve the accuracy and efficiency of three-dimensional sonar imaging. The two-dimensional convolution layer extracts more detailed information by performing convolution processing on the echo intensity imaging result, especially in terms of edges, textures, and depth features, thereby improving the accuracy of imaging. The sub-pixel convolution layer restores more details by enhancing the spatial resolution of the image, ensuring that the generated image is clearer. Further, it can not only provide more accurate high-resolution imaging results, but also process a large amount of echo data in a short time, meeting the dual requirements of accuracy and efficiency for high-demand applications such as deep-sea exploration.

[0055] In one embodiment, the super-resolution module is trained in the following manner:

[0056] Obtain super-resolution module training samples;

[0057] Use the super-resolution module training samples to train the two-dimensional convolution layer and the sub-pixel convolution layer in the super-resolution module to obtain the super-resolution module output;

[0058] Use super-resolution cross-entropy as the loss function of the super-resolution module; wherein, the super-resolution cross-entropy characterizes the difference between the super-resolution module training samples and the super-resolution module output;

[0059] Adopt the way of supervised learning, and optimize the parameters of the super-resolution module through backpropagation and gradient descent until the loss function of the super-resolution module meets the convergence condition to obtain the trained super-resolution module.

[0060] In this embodiment, by obtaining diverse training samples, optimizing the super-resolution module using super-resolution cross-entropy, and continuously optimizing the parameters of the super-resolution module in combination with the way of supervised learning, the finally obtained super-resolution module can significantly improve the accuracy and efficiency of three-dimensional sonar imaging. Thus, it ensures the accurate presentation of the target object and terrain, providing more reliable data support for deep-sea exploration tasks.

[0061] In one embodiment, the super-resolution module training samples are obtained in the following way:

[0062] Use simulation technology to generate a high-resolution sonar target scattering point image, and use the sonar target scattering point image as the label of the super-resolution module;

[0063] Perform delay beamforming processing on the sonar target scattering point image, and generate a corresponding low-resolution image through mean downsampling, and use the low-resolution image as the input of the super-resolution module;

[0064] Based on the input of the super-resolution module and the label of the super-resolution module, construct an input and label pair of the super-resolution module to obtain the training samples of the super-resolution module.

[0065] In this embodiment, simulation technology is used to generate a high-resolution sonar target scattering point image, and it is used as the label of the super-resolution module to provide real high-resolution target information for the super-resolution module; the sonar target scattering point image is processed through delay beamforming processing, and a low-resolution image is generated through mean downsampling, and the low-resolution image is used as the input of the super-resolution module; by constructing the input and label into an input and label pair, training samples of the super-resolution module are generated to help the super-resolution module recover a high-resolution image from the low-resolution input.

[0066] In a second aspect, an embodiment of the present application provides a high-resolution imaging device, and the device includes:

[0067] An encoding module, configured to perform feature extraction and data dimensionality reduction on the sonar echo signals collected by each channel to obtain a complex envelope data matrix; wherein, the sonar echo signal represents the signal received by the sonar reflected from the target position; the encoding module is trained based on the sonar echo signal with added noise as the input and the complex envelope corresponding to the noise-free simulated sonar echo signal at different distances as the label.

[0068] An imaging module, configured to generate an echo intensity imaging result corresponding to the complex envelope data matrix;

[0069] A super-resolution module, configured to enhance the echo intensity imaging result to generate a high-resolution imaging result.

[0070] In a third aspect, an embodiment of the present application provides a computer device, including:

[0071] A memory and a processor, which are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to execute the above-mentioned high-resolution imaging method.

[0072] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which computer instructions are stored, and the computer instructions are used to cause a computer to execute the above-mentioned high-resolution imaging method. Description of the Drawings

[0073] To more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0074] Figure 1 It is a flowchart of a high-resolution imaging method provided by an embodiment of the present application;

[0075] Figure 2 It is a flowchart of step S1 provided by an embodiment of the present application;

[0076] Figure 3 It is a flowchart obtained by training the encoding module provided by an embodiment of the present application in the following manner;

[0077] Figure 4 It is a flowchart of step S011 provided by an embodiment of the present application;

[0078] Figure 5 It is a flowchart of generating pure complex envelope data provided by an embodiment of the present application;

[0079] Figure 6 It is a flowchart of step S3 provided by an embodiment of the present application;

[0080] Figure 7 It is a flowchart obtained by training the imaging module provided by an embodiment of the present application in the following manner;

[0081] Figure 8 It is a flowchart of step S031 provided by an embodiment of the present application;

[0082] Figure 9 It is a flowchart of step S5 provided by an embodiment of the present application;

[0083] Figure 10 It is a flowchart obtained by training the super-resolution module provided by an embodiment of the present application in the following manner;

[0084] Figure 11 It is a flowchart of step S071 provided by an embodiment of the present application;

[0085] Figure 12 It is a schematic diagram of an application to a neural network model provided by an embodiment of the present application;

[0086] Figure 13 It is a block diagram of a high-resolution imaging device provided by an embodiment of the present application;

[0087] Figure 14Schematic structural diagram of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0088] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Apparently, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0089] The resolution of traditional sonar imaging systems is often limited by hardware devices, especially factors such as the aperture of the device and the performance of the sensors, resulting in poor performance when imaging distant and small-sized targets. The aperture of the sonar device (i.e., the size of the sonar array) is an important factor affecting the resolution of sonar imaging. The smaller the aperture, the lower the resolution because the directivity of the sonar beam is restricted, making it impossible to effectively distinguish targets at a relatively long distance or with a small size. When the target is at a relatively long distance or the target itself is small (such as small organisms or tiny geological structures in the deep sea), traditional sonar is difficult to provide clear and accurate images. The echo signal of the target may be interfered by noise, or due to insufficient resolution, the details of the target are blurred, resulting in poor imaging effects. Especially in the deep-sea environment, traditional sonar imaging methods are difficult to effectively capture subtle structural changes, limiting their application in deep-sea scientific research.

[0090] In addition, in the process of data processing of traditional three-dimensional sonar imaging technology, multiple calculation and processing steps are usually required, including filtering, transformation, beamforming, point cloud extraction, etc. Each step will increase the computational burden, and these processes often take a long time. Therefore, in the rapid detection tasks of large ocean areas or scenarios that require real-time feedback, traditional methods often cannot meet the requirements. Imaging requires a long time for data processing, resulting in the inability to provide high-quality imaging results in a timely manner, which is obviously unacceptable for real-time decision-making or emergency operations.

[0091] Therefore, new imaging technologies need to be developed to improve the imaging resolution and overcome the bottlenecks in processing speed and accuracy of traditional methods. It is necessary not only to improve the detection accuracy but also to enhance the data processing efficiency to meet the high-efficiency imaging requirements in practical applications.

[0092] To solve the above technical problems, according to an embodiment of the present application, an embodiment of a high-resolution imaging method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0093] In this embodiment, a high-resolution imaging method is provided. Figure 1 It is a flowchart of a high-resolution imaging method provided by an embodiment of the present application. As Figure 1 shown, it is applied to a neural network model, and the neural network model includes an encoding module, an imaging module, and a super-resolution module; this process includes the following steps:

[0094] Step S1, the encoding module performs feature extraction and data dimensionality reduction on the sonar echo signals collected by each channel to obtain a complex envelope data matrix; wherein, the sonar echo signal represents the signal received by the sonar and reflected from the target position; the encoding module is trained with the sonar echo signal with added noise as the input and the complex envelope corresponding to the noise-free simulated sonar echo signal at different distances as the label.

[0095] Specifically, the encoding module mainly processes the sonar echo signals collected by each channel of the sonar array. These sonar echo signals are reflected from the target position and contain the feature information of the target. Through feature extraction, the encoding module can extract useful information for subsequent processing from the complex sonar echo signals, such as features like the amplitude and phase of the target. At the same time, data dimensionality reduction can reduce the amount of data, improve processing efficiency, remove redundant information, and make subsequent processing more efficient. The finally obtained complex envelope data matrix is a simplified signal representation form that retains key information and is convenient for further processing and analysis by subsequent modules. In addition, during the forward propagation process, a sliding window is used to perform streaming processing on the sonar echo signals. Specifically, the sampled data within the sliding window is first adjusted in shape to adapt to the input format of the encoding module. Subsequently, this adjusted data is input into the encoding module, and after processing, a pure complex envelope data matrix is output.

[0096] Step S3, the imaging module generates an echo intensity imaging result corresponding to the complex envelope data matrix.

[0097] Specifically, the imaging module mainly converts the complex envelope data matrix obtained by processing through the encoding module into an intuitive echo intensity imaging result. The complex envelope data matrix contains the amplitude and phase information of the sonar echo signal, and this information reflects the reflection characteristics of the target at different positions. The imaging module processes this data through operations such as signal enhancement, noise suppression, and spatial mapping, etc., to convert the complex envelope data into an imaging result, where the value of each pixel represents the echo intensity at the corresponding position. This imaging result can visually display the position, shape, and distribution of the target, providing an important visual basis for subsequent target recognition, classification, and analysis, thus playing an important role in fields such as marine exploration, underwater archaeology, and military reconnaissance.

[0098] Step S5, the super-resolution module enhances the echo intensity imaging result to generate a high-resolution imaging result.

[0099] Specifically, the super-resolution module focuses on processing and enhancing the echo intensity imaging result output by the imaging module. Through a series of complex algorithms and neural network structures, such as convolutional layers and sub-pixel convolutional layers, the super-resolution module can effectively improve the resolution of the image, making the imaging result clearer and more delicate. This process not only enhances the detail performance of the image but also improves the accuracy of target recognition, providing a higher-quality high-resolution imaging result.

[0100] A high-resolution imaging method provided in this embodiment, through the encoding module, extracts features and reduces the dimension of the sonar echo signals collected by each channel to obtain a complex envelope data matrix, thereby effectively compressing and processing the original signal, removing redundant information, and extracting valuable features. This process provides a streamlined and efficient data input for subsequent imaging and enhancement processing, improving the efficiency of data processing. Subsequently, the imaging module converts the complex envelope data matrix into an echo intensity imaging result, accurately converting the sonar echo signal into a visual image, laying a foundation for subsequent image enhancement. Finally, the super-resolution module generates a high-resolution imaging result by enhancing the echo intensity imaging result, significantly improving the spatial resolution of the imaging, showing clearer and more accurate target details, and is especially suitable for complex deep-sea exploration environments.

[0101] Figure 2 The flowchart of step S1 provided in the embodiment of this application, this process may include the following steps:

[0102] Step S11, use the one-dimensional convolutional layer and pooling layer in the encoding module to extract features from the sonar echo signal of each channel to obtain the temporal features corresponding to the sonar echo signal.

[0103] Specifically, in sonar signal processing, the encoding module is used to extract useful features from sonar echo signals. In a one-dimensional convolutional layer, the convolutional kernel slides over the sonar echo signals in each channel to extract temporal features. These features can capture local patterns and variations in the signal. The pooling layer follows the convolutional layer immediately and is used to downsample the extracted temporal features. The role of the pooling layer is to reduce the dimension of the data while retaining the most important features. The sonar echo signal refers to the signal received by the sonar system that is reflected back from the target location. The sonar echo signal of each channel refers to the signal received by one array element. Suppose there is a sonar system using a 32×32 planar array with a total of 1024 array elements. The signal received by each array element is a time series with a length of 16, and the sonar echo signal at each channel has 16 time points. The one-dimensional convolutional layer uses the convolutional kernel to slide over the signals in each channel, calculates the convolution operation, and extracts local features. The pooling layer is used to downsample the temporal features to reduce the data dimension. The encoding module consists of multiple one-dimensional convolutional layers and pooling layers, and the stacking of these layers can gradually extract the deep features of the signal. Repeating the above steps to further extract temporal features, the final output temporal features are (4, 1, 64). After the convolution and pooling operations, the ReLU (Rectified Linear Unit) activation function is used to introduce non-linearity. Through the ReLU activation function, the encoding module can process negative values after the convolution operation, making the output signal more non-linear, thus enhancing the learning ability of the encoding module for complex patterns. After passing through the ReLU activation function, the signal of each channel is converted into a set of features. These features will contain temporal information such as local patterns, variations, and fluctuations of the signal.

[0104] Step S13: Flatten the temporal features and perform data dimensionality reduction through the fully connected layer in the encoding module to obtain the complex envelope data matrix.

[0105] Specifically, the flattening process is to convert multi-dimensional data into one-dimensional data. In a neural network, the flattening process is usually performed after the convolutional layer and the pooling layer to input the data into the fully connected layer. The fully connected layer is used to adjust the dimension of the features and map them to the target dimension. Specifically, the shape of the feature vector obtained after flattening the temporal features of (4, 1, 64) is (1, 1, 256). The dimension of the features is adjusted through the fully connected layer and mapped to the feature vector of the target dimension (1, 1, 2048). The feature vector of the target dimension (1, 1, 2048) is adjusted to the shape of (2, 32, 32) to obtain the complex envelope data matrix. Each element of the matrix represents the amplitude and phase information at different positions of the complex envelope data, where 2 corresponds to the real part and the imaginary part of the complex envelope data respectively.

[0106] In this embodiment, the one-dimensional convolutional layer and pooling layer in the encoding module are used to extract features from the sonar echo signal, and the temporal features in the signal can be extracted. The convolutional layer can capture important features in the signal through local perception, while the pooling layer can reduce data redundancy, improve computational efficiency, and enhance the robustness of features, which helps to extract more accurate temporal information from complex sonar echoes. Through the flattening process and the dimensionality reduction operation of the fully connected layer, the data dimension can be further reduced, and the complexity of subsequent processing can be simplified. The dimensionality reduction process of the fully connected layer can retain the most important information while removing unnecessary noise, significantly improving the efficiency of data processing. This processing method not only helps to improve the calculation speed and meet the real-time requirements, but also ensures that the processed data can more accurately reflect the characteristics of the sonar echo signal, thereby improving the imaging accuracy and meeting the high-precision and high-efficiency imaging requirements such as deep-sea detection.

[0107] Figure 3 The flowchart for training the encoding module provided by the embodiment of the present application is obtained in the following manner, and this process may include the following steps:

[0108] Step S011, obtaining encoding module training samples.

[0109] Specifically, the sonar echo signal is simulated using the Rayleigh scattering formula to generate a noise-free sonar echo signal. Noise is added to the noise-free sonar echo signal to simulate the noise interference in the actual sonar system. These noisy sonar echo signals will be used as the input data of the encoding module. The noise-free sonar echo signal is processed to extract the complex envelope, and pure complex envelope data is generated as the label. These label data will be used for supervised learning to help the encoding module learn how to extract pure complex envelope data from the noisy sonar echo signal. The noisy sonar echo signal and the pure complex envelope data are paired to form a large number of encoding module training samples.

[0110] Step S013, using the encoding module training samples to train the one-dimensional convolutional layer, pooling layer, and fully connected layer in the encoding module to obtain the encoding module output.

[0111] Specifically, the one-dimensional convolutional layer is used to perform a convolution operation on the noisy sonar echo signal of each channel to extract temporal features. The ReLU activation function is used in the convolutional layer to introduce non-linearity, enabling the encoding module to learn more complex features. The pooling layer is used to downsample the convolved features to reduce the data dimension and computational complexity. The pooled features are flattened, and the dimension is adjusted through the fully connected layer to obtain the final encoding module output.

[0112] Step S015, using the encoding mean square error as the loss function of the encoding module; among them, the encoding mean square error characterizes the difference between the encoding module training samples and the encoding module output.

[0113] Specifically, the encoded mean squared error (MSE) is used as the loss function to calculate the difference between the output of the encoding module (predicted complex envelope data) and the label (clean complex envelope data).

[0114] In step S017, in a supervised learning manner, the parameters of the encoding module are optimized through backpropagation and gradient descent until the loss function of the encoding module meets the convergence condition, and the trained encoding module is obtained.

[0115] Specifically, the gradient of the loss function of the encoding module with respect to the parameters of the encoding module is calculated, and the gradient is propagated back to each layer of the network through backpropagation. The parameters of the encoding module are updated using the gradient descent method, and the value of the loss function of the encoding module is gradually reduced. When the value of the loss function of the encoding module no longer decreases significantly or reaches a certain preset threshold range, it is considered that the encoding module has converged. The convergence condition can be determined by setting the maximum number of iterations, the minimum loss threshold, etc. Through the above optimization process, the trained encoding module is obtained, which can extract clean complex envelope data from the sonar echo signal with noise.

[0116] In this embodiment, by obtaining the training samples of the encoding module, it is ensured that the encoding module can learn effective signal features and effectively cope with noise interference. Using the structure of the convolutional neural network (CNN), the encoding module can extract local and global features in the sonar signal, thereby improving the signal quality and reducing the impact of noise on the imaging result. Using the mean squared error as the loss function further optimizes the output accuracy of the encoding module and ensures more accurate signal reconstruction. Through the backpropagation and gradient descent optimization algorithms, the parameters are gradually adjusted, improving the generalization ability of the encoding module and ensuring stable signal recovery in complex environments. It not only improves the clarity and accuracy of sonar imaging, but also greatly improves the processing efficiency and can more quickly handle dynamic and complex detection tasks such as deep sea.

[0117] Figure 4 The flowchart of step S011 provided by the embodiment of the present application may include the following steps:

[0118] In step S0111, the sonar echo signal is simulated using the Rayleigh scattering formula to generate a noise-free simulated sonar echo signal.

[0119] Specifically, the Rayleigh scattering formula describes the scattering characteristics of sound waves in a medium and is applicable to the scattering of small particles or small targets. In sonar signal processing, the Rayleigh scattering model can be used to simulate the reflection characteristics of targets. Through the Rayleigh scattering formula, the reflection of small targets on the sonar transmitted signal can be effectively simulated, and then a noise-free simulated sonar echo signal can be generated.

[0120] Step S0113: Add noise to the noise-free simulated sonar echo signal to generate a noisy simulated sonar echo signal, and use the noisy simulated sonar echo signal as the input to the encoding module.

[0121] Specifically, to simulate the actual sonar environment, noise is added to the noise-free simulated sonar echo signal. The noise usually comes from environmental noise in seawater, electrical noise, etc.

[0122] Step S0115: Extract the complex envelopes of the noise-free simulated sonar echo signals at different distances to generate pure complex envelope data, and use the pure complex envelope data as the label of the encoding module.

[0123] Specifically, by dividing the far-field imaging space into multiple equidistant planes and determining the time series label. Then, based on the noise-free simulated sonar echo signal, calculate the complex envelope and perform time-frequency analysis to extract the core information of the echo signal, ensuring the high quality of the signal. Finally, generate pure complex envelope data by organizing the complex envelopes into a matrix form. And use it as the label of the encoding module for subsequent signal processing and imaging.

[0124] Step S0117: Based on the input of the encoding module and the label of the encoding module, construct the input and label pairs of the encoding module to obtain the training samples of the encoding module.

[0125] Specifically, the input is the noisy simulated sonar echo signal, which will be used to train the encoding module. The label is the pure complex envelope data, which represents the complex envelopes of the noise-free simulated sonar echo signal at different distances. Pair the input and the label to form the training samples of the encoding module. Each training sample includes a pair of input and label for supervised learning.

[0126] In this embodiment, a noise-free sonar echo signal is simulated by using the Rayleigh scattering formula, and on this basis, noise is added to generate a noisy sonar echo signal, thus simulating the noise interference in the actual environment. Then, the complex envelopes of the noise-free sonar echo signals at different distances are extracted to generate pure complex envelope data as the label of the encoding module, ensuring the high-quality extraction of the core features of the signal. Based on the noisy simulated sonar echo signal and the pure complex envelope data, the training samples of the encoding module are constructed, enabling the encoding module to effectively extract useful information from the noisy signal, thereby improving the accuracy and efficiency of three-dimensional sonar imaging and meeting the requirements of high-demanding applications such as deep-sea exploration.

[0127] Figure 5 The flowchart for generating pure complex envelope data provided by the embodiment of the present application may include the following steps:

[0128] Step S01151, divide the far-field imaging space into multiple equidistant surfaces according to a preset distance resolution; wherein, the far-field imaging space represents the area far from the sonar emission source during the propagation of the sonar emission signal.

[0129] Specifically, the far-field imaging space refers to the area far from the sonar emission source during the propagation of the sonar emission signal. In this area, the propagation of sound waves can be approximated as plane waves, and the characteristics of the signal are relatively stable, which is suitable for imaging processing. According to the preset distance resolution d r , divide the far-field imaging space into multiple equidistant surfaces, and the distance of each equidistant surface is r n = r0 + n × d r ; wherein, r n is the distance of the nth equidistant surface; r0 is the starting distance, and the minimum distance r0 that satisfies the far-field assumption is r0 ≥ D 2 / λ, D is the sonar aperture, and λ is the wavelength corresponding to the center frequency of the sonar emission signal; n = 0, 1, 2, …. The far-field assumption means that in the far-field imaging space, the propagation of sound waves can be approximated as plane waves to avoid the influence of near-field effects (such as wave diffraction and interference) on the imaging result.

[0130] Step S01153, determine the time series label corresponding to the equidistant surface; wherein, the time series label represents the position of the noise-free simulated sonar echo signal in the time series.

[0131] Specifically, for the distance r n of each equidistant surface, calculate the corresponding time series label I n for positioning the noise-free simulated sonar echo signal collected in the time series: I n = round(2f s × r n / c); wherein, f s is the sampling rate, that is, the number of samples of the noise-free simulated sonar echo signal collected per unit time; c is the speed of sound in the medium, and round is the rounding operation.

[0132] Step S01155, based on the time series label, the noise-free simulated sonar echo signal, the center frequency of the sonar emission signal, the sampling time interval, and the time window length, determine the complex envelope corresponding to the noise-free simulated sonar echo signal at the equidistant surface.

[0133] Specifically, for the distance r n of each equidistant surface, calculate the complex envelope R n of the center frequency component of the sonar emission signal:

[0134]

[0135] wherein, R n is the complex envelope corresponding to the simulated sonar echo signal without noise at the nth equidistant plane; s m is the simulated sonar echo signal without noise collected by the sonar, representing the signal value at the mth sampling point in the time series; f c is the center frequency of the sonar transmission signal, T is the sampling time interval, and L is the window length, representing the signal segment length used to calculate the complex envelope.

[0136] Step S01157, organize the complex envelope into a matrix form to generate pure complex envelope data.

[0137] Specifically, the complex envelope of each equidistant plane contains a real part and an imaginary part. Arrange the complex envelopes of each equidistant plane in distance order to form a matrix. The finally obtained matrix is the pure complex envelope data matrix (32, 32, 2), which can be used for subsequent signal processing and imaging.

[0138] In this embodiment, by dividing the far-field imaging space into multiple equidistant planes and determining the time series label, the accuracy of target positioning is effectively improved, and signal aliasing is avoided. Then, based on the simulated sonar echo signal without noise, the complex envelope is calculated and time-frequency analysis is performed to extract the core information of the echo signal, ensuring the high quality of the signal. Finally, by organizing the complex envelope into a matrix form, the data processing flow is optimized, the efficiency is improved, and it is especially suitable for large-scale data processing. It not only improves the imaging resolution, but also ensures the fast and accurate positioning of the target in deep-sea detection, meeting the high-precision detection requirements.

[0139] Figure 6 This is the flowchart of step S3 provided by the embodiment of the present application, and this process may include the following steps:

[0140] Step S31, use the two-dimensional convolution layer in the imaging module to process the complex envelope data matrix, extract features and generate the imaging results of each channel.

[0141] Step S33, perform data deformation on the imaging results of each channel to obtain the preliminary imaging results.

[0142] Step S35, use the deconvolution layer in the imaging module to perform deconvolution operation on the preliminary imaging results to obtain the echo intensity imaging results.

[0143] Specifically, the complex envelope data matrix is input into the two-dimensional convolutional layer in the imaging module. The role of the convolutional layer is to extract features from the complex envelope data in the complex envelope data matrix. By learning the convolutional kernel, it can automatically identify and extract important features in the complex envelope data. These features include the spatial structure of the signal, frequency characteristics, etc. After being processed by the convolutional layer, each channel generates an imaging result, representing the response to different features of the complex envelope data. Then, data deformation is performed on the imaging results generated by each channel. Data deformation refers to adjusting the size or rearranging the imaging results to meet the requirements of subsequent processing. Through data deformation, a preliminary imaging result is obtained. This result is based on the features extracted by the convolutional layer but may still be a rough imaging image and has not been finely optimized. In the deconvolution step, the preliminary imaging result is processed by the deconvolution layer. The role of the deconvolution layer is to reverse the convolution operation, thereby restoring a more refined image. The deconvolution operation can be understood as an "upsampling" process of the image, aiming to enhance the resolution and details of the image, making the spatial distribution of the echo signal clearer and more accurate. After the deconvolution operation, the obtained result is the echo intensity imaging result, which represents the intensity distribution of the sonar echo signal at different positions and is usually used to represent the spatial layout of the target object or environmental features.

[0144] Specifically, the shape of the complex envelope data matrix is (32, 32, 2), where the 2 channels represent the real part and the imaginary part of the complex envelope data respectively. A two-dimensional convolutional layer is used to process the complex envelope data matrix to extract features, and 4096 channels are output. Each channel represents the imaging result of a pixel point in the far-field imaging space. The Sigmoid activation function is used to introduce non-linearity and limit the output value within the range of [0, 1]. The 4096-channel data output by the convolutional layer is deformed into the imaging result of a 64×64 imaging plane. A deconvolution layer is used. Here, the deconvolution layer uses two convolutional layers to implement the deconvolution operation to further improve the imaging accuracy, and finally a 64×64 echo intensity imaging result is obtained.

[0145] In this embodiment, features in the complex envelope data are extracted by the two-dimensional convolutional layer to generate imaging results for each channel, and these results reflect different features of the signal. Then, the imaging results of each channel are subjected to data deformation to obtain a preliminary imaging result. Finally, the deconvolution layer is used to perform a deconvolution operation on the preliminary imaging result to improve the resolution and details of the image, and finally an echo intensity imaging result is obtained. By extracting features through the convolutional layer, data deformation, and deconvolution operation, the imaging effect is gradually optimized, which helps to improve the accuracy and clarity of underwater sonar imaging.

[0146] Figure 7 The following is a flowchart for training the imaging module provided by the embodiment of the present application. This process may include the following steps:

[0147] Step S031: Obtain the training samples of the imaging module. Among them, the training samples of the imaging module include the target scatterer positions and the delay beamforming results.

[0148] Step S033: Use the training samples of the imaging module to train the two-dimensional convolutional layer and the deconvolutional layer in the imaging module to obtain the output of the imaging module.

[0149] Step S035: Construct the total loss function of the imaging module based on the imaging mean square error and the imaging cross-entropy. Among them, the imaging mean square error characterizes the difference between the output of the imaging module and the delay beamforming result, and the imaging cross-entropy characterizes the difference between the output of the imaging module and the target scatterer positions.

[0150] Step S037: Adopt the supervised learning method to optimize the parameters of the imaging module through backpropagation and gradient descent until the loss function of the imaging module meets the convergence condition, and obtain the trained imaging module.

[0151] Specifically, the target scatterer positions represent the position information of the target in the imaging space, usually a two-dimensional matrix, where the target positions are marked as 1 and other positions are marked as 0. The delay beamforming result is the result processed by the delay-and-sum (DAS) beamforming algorithm and is used as the reference label for training. Use the training samples of the imaging module to train the imaging module. The two-dimensional convolutional layer is used to extract the features of the input signal, and the deconvolutional layer is used to reconstruct the extracted features into the imaging result. After convolution and deconvolution processing, the imaging module outputs the imaging result. The total loss function of the imaging module consists of two parts, the imaging mean square error (MSE) and the imaging cross-entropy. The total loss function of the imaging module:

[0152]

[0153] where L MSE is the imaging mean square error, which measures the difference between the output y of the imaging module and the delay-and-sum beamforming result x DAS ; is the imaging cross-entropy, which measures the difference between the binarized result y 01 of the output of the imaging module and the 0-1 matrix x 01 corresponding to the target scatterer positions. α and β are weight parameters used to balance the two loss functions; y 01 is calculated as:

[0154]

[0155] where τ is the binarization threshold.

[0156] Calculate the gradient of the loss function of the computational imaging module with respect to the imaging module parameters, and propagate the gradient back to each layer of the network through backpropagation. Use the gradient descent method to update the imaging module parameters and gradually reduce the value of the loss function of the imaging module. When the value of the loss function of the imaging module no longer decreases significantly or reaches a certain preset threshold range, it is considered that the imaging module has converged. The convergence condition can be determined by setting the maximum number of iterations, the minimum loss threshold, etc. After the above optimization process, a trained imaging module is obtained, which can generate accurate echo intensity imaging results from the input complex envelope data matrix.

[0157] In this embodiment, by obtaining the training samples of the target scatterer position and the delay beamforming result, high-quality training data is provided for the imaging module. Then, these training samples are used to train the two-dimensional convolutional layer and the deconvolutional layer in the imaging module, so as to optimize the target scatterer positioning and echo signal matching and improve the imaging ability. On this basis, by constructing a comprehensive loss function (including imaging mean square error and imaging cross entropy), the difference between the output of the imaging module and the actual result is accurately measured, and the performance of the imaging module is further optimized. Finally, supervised learning is adopted to optimize the imaging module parameters through backpropagation and gradient descent methods to ensure that the imaging module provides high-precision three-dimensional sonar imaging in a complex environment.

[0158] Figure 8 It is a flowchart of step S031 provided by an embodiment of the present application, and this process may include the following steps:

[0159] Step S0311, use simulation technology to simulate the complex envelope signal, generate a noise-free simulated complex envelope signal, and use the noise-free simulated complex envelope signal as the input of the imaging module.

[0160] Specifically, use the Rayleigh scattering formula or other sonar signal models to generate a noise-free simulated complex envelope signal. These signals usually contain the reflection characteristics of the target and reflect the echo intensity and phase information of the target at different distances.

[0161] Step S0313, use simulation technology to determine the delay of the sonar echo signal received by each array element according to the target scatterer position and the geometric structure of the sonar array.

[0162] Specifically, the target scattering point position refers to the position of the point on the target that reflects sound waves in space, usually represented by a 0-1 matrix. A sonar array is a system composed of multiple acoustic sensors (array elements) for transmitting and receiving acoustic wave signals. Delay refers to the time difference of the sonar echo signal propagating from the target scattering point position to different array elements of the sonar array. Using the Rayleigh scattering formula or other sonar signal models combined with simulation software, simulate the propagation process of acoustic waves from each array element of the sonar array to the target scattering point position and then to each array element of the sonar array, and use Delay And Sum (DAS) beamforming to calculate the delay of the sonar echo signal received by each array element.

[0163] Step S0315: Align the sonar echo signals received by each array element according to the delay.

[0164] Specifically, align the sonar echo signals received by each array element according to the calculated time delay. The alignment can be achieved by delaying the signal in the time domain or adjusting the phase in the frequency domain.

[0165] Step S0317: Sum the aligned sonar echo signals to obtain the delay sum beamforming result, and use the target scattering point position and the delay sum beamforming result as the labels of the imaging module.

[0166] Specifically, Delay And Sum (DAS) beamforming sums the aligned sonar echo signals to calculate the echo intensity of the target in each direction, and obtains the delay sum beamforming result.

[0167] Step S0319: Based on the input of the imaging module and the labels of the imaging module, construct the input and label pairs of the imaging module to obtain the training samples of the imaging module.

[0168] Specifically, generate the training samples of the imaging module by combining the target scattering point position (label) and the delay sum beamforming result (input), and these data will be used to train the imaging module.

[0169] In this embodiment, the noise-free simulated complex envelope signal is used as the input, effectively reducing noise interference, thereby improving the imaging accuracy. Secondly, the delay of the sonar echo signal is accurately calculated and time-aligned, ensuring that the signals received by different array elements are synchronized and eliminating the imaging distortion caused by time differences. On this basis, through the delay sum technology, the beam focusing effect is enhanced and the imaging effect is optimized. In addition, using the noise-free simulated complex envelope signal and accurate labels to generate the training samples of the imaging module supports the rapid learning of the imaging model, improving the training efficiency and the performance of the imaging model. It ensures that in complex environments such as the deep sea, the imaging system can achieve high-precision and high-efficiency three-dimensional sonar imaging, meeting the requirements of high-demand applications.

[0170] Figure 9 This is the flowchart of step S5 provided by the embodiment of the present application. This process may include the following steps:

[0171] Step S51, perform a convolution operation on the echo intensity imaging result using the two-dimensional convolutional layer in the super-resolution module to obtain a target feature map.

[0172] Step S53, perform a rearrangement operation on the target feature map using the sub-pixel convolutional layer in the super-resolution module to generate a high-resolution imaging result.

[0173] Specifically, perform a convolution operation on the echo intensity imaging result through the first two-dimensional convolutional layer to output a 64-channel feature map. The second two-dimensional convolutional layer receives the output of the first two-dimensional convolutional layer, and the number of output channels and the feature map size remain unchanged. After the first two-dimensional convolutional layer and the second two-dimensional convolutional layer, the ReLU activation function is used to introduce non-linearity to enhance the network's expression ability. The third two-dimensional convolutional layer receives the output of the second two-dimensional convolutional layer and outputs a 4-channel target feature map to prepare for the subsequent sub-pixel convolutional layer. Finally, use the sub-pixel convolutional layer to perform a rearrangement operation on the target feature map, magnify the spatial size of the target feature map by 2 times, and finally obtain a high-resolution imaging result of 128×128.

[0174] In this embodiment, through the two-dimensional convolutional layer and the sub-pixel convolutional layer, the super-resolution module can significantly improve the accuracy and efficiency of three-dimensional sonar imaging. The two-dimensional convolutional layer extracts more detailed information by performing convolution processing on the echo intensity imaging result, especially in terms of edge, texture, and depth features, thereby improving the accuracy of imaging. The sub-pixel convolutional layer restores more details by enhancing the spatial resolution of the image, ensuring that the generated image is clearer. Further, it can not only provide more accurate high-resolution imaging results, but also process a large amount of echo data in a short time, meeting the dual requirements of accuracy and efficiency for high-demand applications such as deep-sea exploration.

[0175] Figure 10 This is the flowchart for obtaining the super-resolution module provided by the embodiment of the present application through training in the following manner. This process may include the following steps:

[0176] Step S071, obtain super-resolution module training samples.

[0177] Step S073, use the super-resolution module training samples to train the two-dimensional convolutional layer and the sub-pixel convolutional layer in the super-resolution module to obtain the super-resolution module output.

[0178] Step S075, use the super-resolution cross-entropy as the loss function of the super-resolution module; wherein, the super-resolution cross-entropy characterizes the difference between the super-resolution module training samples and the super-resolution module output.

[0179] Step S077: In a supervised learning manner, optimize the parameters of the super-resolution module through backpropagation and gradient descent until the loss function of the super-resolution module meets the convergence condition, obtaining a trained super-resolution module.

[0180] Specifically, the training samples of the super-resolution module include low-resolution images that have been downsampled or blurred as inputs, and use simulation technology to generate high-resolution sonar target scattering point images as labels to train the super-resolution module. The two-dimensional convolutional layer in the super-resolution module is used to extract the features of the sonar target scattering point image, obtaining the extracted feature map. The sub-pixel convolutional layer in the super-resolution module is used to magnify and restore the extracted feature map to generate the output of the super-resolution module. Use super-resolution cross-entropy as the loss function of the super-resolution module. Super-resolution cross-entropy can effectively measure the difference between the output image and the label (sonar target scattering point image), especially in pixel-level classification tasks.

[0181] Calculate the gradient of the loss function of the super-resolution module with respect to the parameters of the super-resolution module, and propagate the gradient back to each layer of the network through backpropagation. Use the gradient descent method to update the parameters of the super-resolution module, gradually reducing the value of the loss function of the super-resolution module. When the value of the loss function of the super-resolution module no longer decreases significantly, or reaches a certain preset threshold range, it is considered that the super-resolution module has converged. The convergence condition can be determined by setting the maximum number of iterations, the minimum loss threshold, etc. Through the above optimization process, a trained super-resolution module is obtained, which can generate high-resolution imaging results from the input low-resolution image (echo intensity imaging).

[0182] In this embodiment, by obtaining diverse training samples, optimizing the super-resolution module using super-resolution cross-entropy, and continuously optimizing the parameters of the super-resolution module in a supervised learning manner, the finally obtained super-resolution module can significantly improve the accuracy and efficiency of three-dimensional sonar imaging. Thus, it ensures the accurate presentation of the target object and terrain, providing more reliable data support for deep-sea exploration tasks.

[0183] Figure 11 This is the flowchart of step S071 provided by the embodiment of the present application. This process may include the following steps:

[0184] Step S0711: Use simulation technology to generate a high-resolution sonar target scattering point image, and use the sonar target scattering point image as the label of the super-resolution module.

[0185] Step S0713: Perform delayed beamforming processing on the sonar target scattering point image, and generate a corresponding low-resolution image through mean downsampling, using the low-resolution image as the input of the super-resolution module.

[0186] Step S0715: Based on the input of the super-resolution module and the label of the super-resolution module, construct an input and label pair of the super-resolution module to obtain the training samples of the super-resolution module.

[0187] In this embodiment, simulation technology is used to generate high-resolution sonar target scattering point images, which are used as the labels of the super-resolution module to provide real high-resolution target information for the super-resolution module; the sonar target scattering point images are processed through delay beamforming processing, and low-resolution images are generated through mean downsampling. The low-resolution images are used as the inputs of the super-resolution module; by constructing the inputs and labels into input and label pairs, training samples of the super-resolution module are generated to help the super-resolution module recover high-resolution images from low-resolution inputs.

[0188] The following describes the specific implementation of the present invention in combination with a specific application scenario. Refer to Figure 12 , applied to a neural network model, the neural network model includes an encoding module, an imaging module, and a super-resolution module. Figure 12 In (C, H, W) in [], it represents Channel × Height × Width; Reshape is for reshaping; Conv1d is a one-dimensional convolutional layer; ReLU is the ReLU activation function; the max pooling layer is MaxPool; Flatten is for flattening; FC is a fully connected layer; Conv2d is a two-dimensional convolutional layer; Sigmoid is the Sigmoid activation function; PixelShuffle is a sub-pixel convolutional layer. In this specific application scenario, the input in the encoding module is a 32×32 planar array, including sonar echo signals collected by 1024 array elements. The sequence length within the time window is L = 16, forming data with 1024 channels and 16 dimensions. Through two layers of one-dimensional convolution and pooling operations, temporal features are extracted. The convolutional layer uses the ReLU activation function to introduce non-linearity, and the pooling layer performs downsampling to reduce the data dimension. After the temporal features are flattened (Flattened), the dimension is adjusted through a fully connected layer and mapped to a shape of 32×32×2, where the 2 channels respectively correspond to the real part and the imaginary part of the complex envelope data. The output of the fully connected layer is adjusted to a shape of (32, 32, 2), split into the real part and the imaginary part, and finally a complex envelope data matrix is obtained.

[0189] The imaging module receives the complex envelope data matrix (32, 32, 2) from the encoding module, where the two channels respectively represent the real part and the imaginary part of the complex envelope data. Through a two-dimensional convolutional layer, with a convolutional kernel of 32×32, the two input channels are mapped to 4096 output channels. Each output channel represents the imaging result of a pixel point in the far-field imaging space. The convolutional layer uses the sigmoid activation function to introduce non-linearity. After convolution, the data of the 4096 output channels is reshaped into an imaging result of 64×64. Subsequently, two convolutional layers are used for deconvolution operations to further improve the imaging accuracy and output the echo intensity imaging result.

[0190] The super-resolution module first extracts features from the echo intensity imaging result through two two-dimensional convolutional layers: the first two-dimensional convolutional layer processes the echo intensity imaging result and outputs features of 64 channels. The second two-dimensional convolutional layer receives the output of the previous layer and keeps the number of output channels and the feature map size unchanged. The ReLU activation function is used after each convolution to introduce non-linearity and enhance the expression ability of the network. Then, a third two-dimensional convolutional layer is built to receive the 64-channel features output by the previous layer and output 4-channel features to prepare for the subsequent sub-pixel convolutional layer. Finally, the sub-pixel convolutional layer rearranges the 4-channel features to double the spatial size and obtain a high-resolution imaging result of 128×128.

[0191] Correspondingly, please refer to Figure 13 the block diagram of a high-resolution imaging device provided by an embodiment of the present application. The device includes:

[0192] An encoding module 101, configured to perform feature extraction and data dimensionality reduction on the sonar echo signals collected by each channel to obtain a complex envelope data matrix; wherein, the sonar echo signals represent the signals received by the sonar and reflected from the target position; the encoding module is trained with the sonar echo signals with added noise as the input and the complex envelopes corresponding to the noise-free simulated sonar echo signals at different distances as the labels.

[0193] An imaging module 103, configured to generate an echo intensity imaging result corresponding to the complex envelope data matrix;

[0194] A super-resolution module 105, configured to enhance the echo intensity imaging result and generate a high-resolution imaging result.

[0195] In some alternative embodiments, the encoding module 101 includes:

[0196] Use the one-dimensional convolutional layer and pooling layer in the encoding module to perform feature extraction on the sonar echo signals of each channel to obtain the temporal features corresponding to the sonar echo signals;

[0197] Flatten the time series features and perform data dimensionality reduction through the fully connected layer in the encoding module to obtain the complex envelope data matrix.

[0198] In some alternative embodiments, the encoding module is trained in the following manner:

[0199] Obtain the training samples of the encoding module;

[0200] Use the training samples of the encoding module to train the one-dimensional convolutional layer, pooling layer, and fully connected layer in the encoding module to obtain the output of the encoding module;

[0201] Use the encoding mean square error as the loss function of the encoding module; wherein, the encoding mean square error characterizes the difference between the training samples of the encoding module and the output of the encoding module;

[0202] Adopt the supervised learning method, and optimize the parameters of the encoding module through backpropagation and gradient descent until the loss function of the encoding module meets the convergence condition to obtain the trained encoding module.

[0203] In some alternative embodiments, the training samples of the encoding module are obtained in the following manner:

[0204] Simulate the sonar echo signal using the Rayleigh scattering formula to generate a noise-free simulated sonar echo signal;

[0205] Add noise to the noise-free simulated sonar echo signal to generate a noisy simulated sonar echo signal, and use the noisy simulated sonar echo signal as the input of the encoding module;

[0206] Extract the complex envelopes of the noise-free simulated sonar echo signals at different distances to generate pure complex envelope data, and use the pure complex envelope data as the labels of the encoding module;

[0207] Based on the input of the encoding module and the labels of the encoding module, construct the input and label pairs of the encoding module to obtain the training samples of the encoding module.

[0208] In some alternative embodiments, extracting the complex envelopes of the noise-free simulated sonar echo signals at different distances to generate pure complex envelope data includes:

[0209] According to the preset distance resolution, divide the far-field imaging space into multiple equidistant planes; wherein, the far-field imaging space represents the area far from the sonar transmitter during the propagation of the sonar transmission signal;

[0210] Determine the time series labels corresponding to the equidistant planes; wherein, the time series labels represent the positions of the noise-free simulated sonar echo signals in the time series;

[0211] Based on the time - series label, the noise - free simulated sonar echo signal, the center frequency of the sonar transmit signal, the sampling time interval, and the time - window length, determine the complex envelope corresponding to the noise - free simulated sonar echo signal at the equidistant plane;

[0212] Organize the complex envelope into a matrix form to generate pure complex envelope data.

[0213] In some alternative embodiments, the imaging module 103 includes:

[0214] Use the two - dimensional convolutional layer in the imaging module to process the complex envelope data matrix, extract features, and generate the imaging results for each channel;

[0215] Perform data transformation on the imaging results of each channel to obtain preliminary imaging results;

[0216] Use the de - convolutional layer in the imaging module to perform de - convolutional operations on the preliminary imaging results to obtain the echo intensity imaging results.

[0217] In some alternative embodiments, the imaging module is trained in the following manner:

[0218] Obtain the imaging module training samples; wherein, the imaging module training samples include the target scatterer positions and the delayed beamforming results;

[0219] Use the imaging module training samples to train the two - dimensional convolutional layer and the de - convolutional layer in the imaging module to obtain the imaging module output;

[0220] Construct the total loss function of the imaging module based on the imaging mean square error and the imaging cross - entropy; wherein, the imaging mean square error characterizes the difference between the imaging module output and the delayed beamforming results, and the imaging cross - entropy characterizes the difference between the imaging module output and the target scatterer positions;

[0221] Adopt the supervised learning method, and optimize the parameters of the imaging module through backpropagation and gradient descent until the loss function of the imaging module meets the convergence condition to obtain the trained imaging module.

[0222] In some alternative embodiments, the imaging module training samples are obtained in the following manner:

[0223] Use simulation technology to simulate the complex envelope signal, generate a noise - free simulated complex envelope signal, and use the noise - free simulated complex envelope signal as the input of the imaging module;

[0224] Use simulation technology to determine the delay of the sonar echo signal received by each array element according to the target scatterer positions and the geometric structure of the sonar array;

[0225] Align the sonar echo signals received by each array element according to the delay;

[0226] Sum the aligned sonar echo signals to obtain the delay sum beamforming result, and use the target scatterer position and the delay sum beamforming result as the labels of the imaging module;

[0227] Based on the input of the imaging module and the labels of the imaging module, construct the input and label pairs of the imaging module to obtain the imaging module training samples.

[0228] In some alternative embodiments, the super-resolution module 105 includes:

[0229] Perform a convolution operation on the echo intensity imaging result using the two-dimensional convolutional layer in the super-resolution module to obtain the target feature map;

[0230] Perform a rearrangement operation on the target feature map using the sub-pixel convolutional layer in the super-resolution module to generate a high-resolution imaging result.

[0231] In some alternative embodiments, the super-resolution module is trained in the following manner:

[0232] Obtain the super-resolution module training samples;

[0233] Use the super-resolution module training samples to train the two-dimensional convolutional layer and the sub-pixel convolutional layer in the super-resolution module to obtain the super-resolution module output;

[0234] Use the super-resolution cross-entropy as the loss function of the super-resolution module; wherein, the super-resolution cross-entropy characterizes the difference between the super-resolution module training samples and the super-resolution module output;

[0235] Adopt the supervised learning method, and optimize the parameters of the super-resolution module through backpropagation and gradient descent until the loss function of the super-resolution module meets the convergence condition to obtain the trained super-resolution module.

[0236] In some alternative embodiments, the super-resolution module training samples are obtained through the following method:

[0237] Use the simulation technology to generate a high-resolution sonar target scatterer image, and use the sonar target scatterer image as the label of the super-resolution module;

[0238] Perform delay beamforming processing on the sonar target scatterer image, and generate the corresponding low-resolution image through mean downsampling, and use the low-resolution image as the input of the super-resolution module;

[0239] Based on the input of the super-resolution module and the labels of the super-resolution module, construct the input and label pairs of the super-resolution module to obtain the super-resolution module training samples.

[0240] The further function descriptions of the above-mentioned various modules and units are the same as those in the corresponding embodiments above, and will not be elaborated here.

[0241] A high-resolution imaging device in this embodiment is presented in the form of functional modules. Here, the module refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and a memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0242] Please refer to Figure 14 , Figure 14 which is a schematic structural diagram of a computer device provided by an embodiment of the present application. As Figure 14 shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Each component communicates with each other using different buses and can be installed on a common motherboard or installed in other ways as needed. The processor can process instructions executed within the computer device, including instructions stored in the memory or on the memory to display graphical information of the GUI on an external input / output device (such as a display device coupled to the interface). In some alternative embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (such as a server array, a set of blade servers, or a multi-processor system). Figure 14 In

[0243] the example of one processor 10 is taken.

[0244] Among them, the memory 20 stores instructions executable by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiments.

[0245] The memory 20 may include a program storage area and a data storage area. The program storage area may store an operating system and application programs required for at least one function. The data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some alternative embodiments, the memory 20 may optionally include a memory remotely disposed relative to the processor 10, and these remote memories may be connected to the computer device through a network. Examples of the above networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0246] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk, or a solid-state drive; the memory 20 may further include a combination of the above types of memories.

[0247] The computer device further includes a communication interface 30 for the computer device to communicate with other devices or communication networks.

[0248] The embodiments of the present application also provide a computer-readable storage medium. The methods according to the embodiments of the present application may be implemented in hardware, firmware, or may be implemented as computer code that can be recorded on a storage medium, or may be implemented as computer code originally stored in a remote storage medium or a non-transitory machine-readable storage medium and downloaded through a network and to be stored in a local storage medium, so that the methods described herein may be stored in such software processed on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium may be a magnetic disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk, or a solid-state drive, etc.; further, the storage medium may also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code, and when the software or computer code is accessed and executed by the computer, the processor, or the hardware, the methods shown in the above embodiments are implemented.

[0249] The devices and modules illustrated in the above embodiments may be specifically implemented by computer chips or entities, or by products having a certain function. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0250] For convenience of description, when describing the above device, various modules are described separately according to their functions. Of course, when implementing the present application, the functions of each module can be implemented in one or more software and / or hardware.

[0251] Those skilled in the art should understand that the embodiments of the present application can be provided as methods and devices. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0252] The present application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses, and devices according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0253] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0254] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0255] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, commodity or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising one..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising said element.

[0256] Each embodiment in this specification is described in a progressive manner. For the identical or similar parts among the embodiments, reference can be made to each other, and the key point of each embodiment is to illustrate the differences from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and reference can be made to the corresponding parts in the method embodiments for the relevant content.

[0257] The above are only the embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various modifications and changes can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

[0258] Although the embodiments of the present application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present application, and such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A high-resolution imaging method, characterized in that Applied to a neural network model, the neural network model includes an encoding module, an imaging module, and a super-resolution module; the method includes: The encoding module performs feature extraction and data dimensionality reduction on the sonar echo signals collected by each channel to obtain a complex envelope data matrix; wherein, the sonar echo signals represent the signals received by the sonar reflected from the target position; the encoding module is trained with the sonar echo signals with added noise as the input and the complex envelopes corresponding to the noise-free simulated sonar echo signals at different distances as the labels; wherein, the encoding module is trained in the following manner: obtaining encoding module training samples; using the encoding module training samples to train the one-dimensional convolutional layer, pooling layer, and fully connected layer in the encoding module to obtain an encoding module output; using encoding mean square error as the loss function of the encoding module; wherein, the encoding mean square error represents the difference between the encoding module training samples and the encoding module output; adopting the supervised learning method, optimizing the parameters of the encoding module through backpropagation and gradient descent until the loss function of the encoding module satisfies the convergence condition to obtain a trained encoding module; The imaging module generates an echo intensity imaging result corresponding to the complex envelope data matrix; wherein, the imaging module is trained in the following manner: obtaining imaging module training samples; wherein, the imaging module training samples include target scatterer positions and delayed beamforming results; using the imaging module training samples to train the two-dimensional convolutional layer and deconvolutional layer in the imaging module to obtain an imaging module output; constructing a total loss function of the imaging module based on imaging mean square error and imaging cross-entropy; wherein, the imaging mean square error represents the difference between the imaging module output and the delayed beamforming results, and the imaging cross-entropy represents the difference between the imaging module output and the target scatterer positions; adopting the supervised learning method, optimizing the parameters of the imaging module through backpropagation and gradient descent until the loss function of the imaging module satisfies the convergence condition to obtain a trained imaging module; The super-resolution module enhances the echo intensity imaging result to generate a high-resolution imaging result; wherein, the super-resolution module is trained in the following manner: obtaining super-resolution module training samples; using the super-resolution module training samples to train the two-dimensional convolutional layer and sub-pixel convolutional layer in the super-resolution module to obtain a super-resolution module output; using super-resolution cross-entropy as the loss function of the super-resolution module; wherein, the super-resolution cross-entropy represents the difference between the super-resolution module training samples and the super-resolution module output; adopting the supervised learning method, optimizing the parameters of the super-resolution module through backpropagation and gradient descent until the loss function of the super-resolution module satisfies the convergence condition to obtain a trained super-resolution module.

2. The method according to claim 1, wherein The encoding module performs feature extraction and data dimensionality reduction on the sonar echo signals collected by each channel to obtain a complex envelope data matrix, including: Use the one-dimensional convolutional layer and pooling layer in the encoding module to extract features from the sonar echo signals of each channel, and obtain the temporal features corresponding to the sonar echo signals; Flatten the temporal features and perform data dimensionality reduction through the fully connected layer in the encoding module to obtain the complex envelope data matrix.

3. The method according to claim 1, characterized in that, The training samples of the encoding module are obtained through the following method: Simulate the sonar echo signals using the Rayleigh scattering formula to generate noise-free simulated sonar echo signals; Add noise to the noise-free simulated sonar echo signals to generate noisy simulated sonar echo signals, and use the noisy simulated sonar echo signals as the input of the encoding module; Extract the complex envelopes of the noise-free simulated sonar echo signals at different distances to generate pure complex envelope data, and use the pure complex envelope data as the label of the encoding module; Based on the input of the encoding module and the label of the encoding module, construct the input and label pairs of the encoding module to obtain the training samples of the encoding module.

4. The method according to claim 3, characterized in that, The extraction of the complex envelopes of the noise-free simulated sonar echo signals at different distances to generate pure complex envelope data includes: According to the preset distance resolution, divide the far-field imaging space into multiple equidistant planes; where the far-field imaging space represents the area far from the sonar transmitter during the propagation of the sonar transmission signal; Determine the time series labels corresponding to the equidistant planes; where the time series labels represent the positions of the noise-free simulated sonar echo signals in the time series; Based on the time series labels, the noise-free simulated sonar echo signals, the center frequency of the sonar transmission signal, the sampling time interval, and the time window length, determine the complex envelopes corresponding to the noise-free simulated sonar echo signals at the equidistant planes; Organize the complex envelopes into a matrix form to generate the pure complex envelope data.

5. The method according to claim 1, wherein The imaging module generates the echo intensity imaging result corresponding to the complex envelope data matrix, including: Use the two-dimensional convolutional layer in the imaging module to process the complex envelope data matrix, extract features and generate the imaging results of each channel; Deform the imaging results of each channel to obtain the preliminary imaging result; Use the deconvolution layer in the imaging module to perform deconvolution operations on the preliminary imaging result to obtain the echo intensity imaging result.

6. The method according to claim 1, characterized in that, The training samples of the imaging module are obtained through the following method: Simulate the complex envelope signals using simulation technology to generate noise-free simulated complex envelope signals, and use the noise-free simulated complex envelope signals as the input of the imaging module; Use simulation technology to determine the delays of the sonar echo signals received by each array element according to the positions of the target scatterers and the geometric structure of the sonar array; Align the sonar echo signals received by each array element in time according to the delays; Sum the aligned sonar echo signals to obtain the delay sum beamforming result, and use the positions of the target scatterers and the delay sum beamforming result as the labels of the imaging module; Based on the input of the imaging module and the label of the imaging module, construct the input and label pair of the imaging module to obtain the training samples of the imaging module.

7. The method according to claim 1, characterized in that The super-resolution module enhances the echo intensity imaging result to generate a high-resolution imaging result, including: Performing a convolution operation on the echo intensity imaging result using the two-dimensional convolutional layer in the super-resolution module to obtain a target feature map; Performing a rearrangement operation on the target feature map using the sub-pixel convolutional layer in the super-resolution module to generate the high-resolution imaging result.

8. The method according to claim 1, characterized in that The training samples of the super-resolution module are obtained through the following method: Using simulation technology to generate a high-resolution sonar target scattering point image, and using the sonar target scattering point image as the label of the super-resolution module; Performing delayed beamforming processing on the sonar target scattering point image, and generating a corresponding low-resolution image through mean downsampling, and using the low-resolution image as the input of the super-resolution module; Based on the input of the super-resolution module and the label of the super-resolution module, construct the input and label pair of the super-resolution module to obtain the training samples of the super-resolution module.

9. A high-resolution imaging device, characterized in that, The device includes: An encoding module for extracting features and reducing the data dimension of the sonar echo signals collected by each channel to obtain a complex envelope data matrix; wherein, the sonar echo signals represent the signals received by the sonar reflected from the target position; the encoding module is trained based on the sonar echo signals with added noise as the input and the complex envelopes corresponding to the noise-free simulated sonar echo signals at different distances as the labels; wherein, the encoding module is trained in the following manner: obtaining the training samples of the encoding module; using the training samples of the encoding module to train the one-dimensional convolutional layer, pooling layer and fully-connected layer in the encoding module to obtain the output of the encoding module; using the encoding mean square error as the loss function of the encoding module; wherein, the encoding mean square error represents the difference between the training samples of the encoding module and the output of the encoding module; adopting the supervised learning method, and optimizing the parameters of the encoding module through backpropagation and gradient descent until the loss function of the encoding module meets the convergence condition to obtain the encoding module that has completed training; An imaging module for generating an echo intensity imaging result corresponding to the complex envelope data matrix; wherein, the imaging module is trained in the following manner: acquiring imaging module training samples; wherein, the imaging module training samples include target scatterer positions and delay beamforming results; using the imaging module training samples to train the two-dimensional convolutional layer and the deconvolutional layer in the imaging module to obtain an imaging module output; constructing a total loss function of the imaging module based on imaging mean square error and imaging cross entropy; wherein, the imaging mean square error characterizes the difference between the imaging module output and the delay beamforming result, and the imaging cross entropy characterizes the difference between the imaging module output and the target scatterer position; adopting a supervised learning method, and optimizing the parameters of the imaging module through backpropagation and gradient descent until the loss function of the imaging module meets the convergence condition to obtain a trained imaging module; A super-resolution module for enhancing the echo intensity imaging result to generate a high-resolution imaging result; wherein, the super-resolution module is trained in the following manner: acquiring super-resolution module training samples; using the super-resolution module training samples to train the two-dimensional convolutional layer and the sub-pixel convolutional layer in the super-resolution module to obtain a super-resolution module output; using super-resolution cross entropy as the loss function of the super-resolution module; wherein, the super-resolution cross entropy characterizes the difference between the super-resolution module training samples and the super-resolution module output; adopting a supervised learning method, and optimizing the parameters of the super-resolution module through backpropagation and gradient descent until the loss function of the super-resolution module meets the convergence condition to obtain a trained super-resolution module.

10. A computer device, characterized in that, Comprising: A memory and a processor, the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to execute the high-resolution imaging method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, Computer instructions are stored on the computer-readable storage medium, and the computer instructions are used to cause a computer to execute the high-resolution imaging method according to any one of claims 1 to 8.

Images