Sonar image denoising method and system based on neural network self-searching

Abstract

This invention discloses a sonar image denoising method and system based on neural network self-search, belonging to the field of image processing technology. The method includes: segmenting a sonar image into several non-overlapping small blocks, and generating global noise intensity and noise type labels based on the statistical characteristics of the pixel values ​​of the small blocks; performing a neural network structure search based on the global noise intensity and the noise type labels to obtain an optimal neural network; training the optimal neural network based on a first training dataset, wherein the first training dataset contains more than a set number of data points; and obtaining the denoising result of the sonar image based on the trained optimal neural network. This invention can automatically search for the optimal neural network structure based on the noise characteristics of the input image, thereby improving the denoising effect, reducing computational costs, and enhancing generalization ability.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a sonar image denoising method and system based on neural network self-search. Background Technology

[0002] Sonar image denoising is a crucial step in underwater exploration. Traditional methods include spatial domain filtering (such as median filtering and Gaussian filtering) and transform domain methods (such as wavelet denoising). In recent years, deep learning methods (such as convolutional neural networks (CNNs) have been applied to image denoising, learning denoising maps from noisy images by training network models with fixed structures (such as DnCNN and U-Net). These methods typically require large amounts of labeled data, and the network structure is pre-defined, making it unable to adapt to different noise levels or sonar environments.

[0003] Currently, sonar image denoising is mainly achieved using predefined CNN-based methods. For example, a typical approach uses a U-Net structure, extracting features and reconstructing the denoised image through an encoder-decoder architecture. This approach optimizes network parameters using a mean squared error loss function during training, but the network structure is fixed and cannot be dynamically adjusted according to the noise characteristics of the input image. This results in the following technical drawbacks: Lack of adaptability: Fixed network structures cannot flexibly cope with the non-uniform distribution and time-varying characteristics of noise in sonar images (such as noise differences caused by different water depths and equipment parameters), resulting in fluctuations in denoising performance in different scenarios.

[0004] Low computational efficiency: To handle various noise conditions, it is usually necessary to train multiple models or use complex networks, which increases computational resources and storage overhead, making it unsuitable for real-time applications.

[0005] Limited generalization ability: Predefined networks may overfit the training data and perform poorly under unknown noise distributions. Summary of the Invention

[0006] To address the problem of adaptive denoising, this invention provides a sonar image denoising method and system based on neural network self-search. This method can automatically search for the optimal neural network structure according to the noise characteristics of the input image, thereby improving the denoising effect, reducing the computational cost, and enhancing the generalization ability.

[0007] To achieve the above objectives, the technical solution of the present invention includes the following:

[0008] A sonar image denoising method based on neural network self-search, the method comprising: The sonar image is segmented into several non-overlapping small blocks, and global noise intensity and noise type labels are generated based on the statistical characteristics of the pixel values ​​of the small blocks. Based on the global noise intensity and the noise type label, a neural network structure search is performed to obtain the optimal neural network. The optimal neural network is trained based on a first training dataset, wherein the first training dataset contains more than a set value of data. The denoising result of the sonar image is obtained based on the trained optimal neural network.

[0009] Furthermore, the statistical characteristics of the pixel values ​​include: mean, variance, and gradient distribution.

[0010] Furthermore, based on the global noise intensity and the noise type label, a neural network structure search is performed to obtain the optimal neural network, including: Initialize the search strategy based on the global noise intensity and the noise type. ; Based on search strategy Generate candidate networks The structure description file; where, This indicates the current search iteration round, and the content of this structure description file includes the candidate network. The number of layers, the operation type of each layer, and the connection method; According to candidate networks The structure description file is compiled and the candidate network is instantiated. ; The candidate network was trained using the second training dataset. The second training dataset contains less than this set value. On the validation set, the candidate network is obtained based on mean squared error and structural similarity index. Performance score; Based on the performance score and the candidate network Structure description file update search strategy And based on the updated search strategy Re-execute the search-based strategy Generate candidate networks Structure description file; After completing the predetermined number of search iterations, the performance score is used to rank all candidate networks. The optimal neural network is selected from these.

[0011] Furthermore, the search strategy This includes the depth range of the neural network and the number of channels in each layer; Initialize the search strategy based on the global noise intensity and the noise type. ,include: The lower and upper limits of the number of network layers in this search are calculated based on the global noise intensity, and the depth range of the neural network is defined based on the lower and upper limits. Define the number of channels for each layer based on the noise type label.

[0012] Furthermore, before segmenting the sonar image into several non-overlapping small blocks, the method further includes: The sonar image is subjected to a standardization preprocessing, which includes: adjusting the size to a uniform specification and normalizing the pixel values ​​to the range of [0,1].

[0013] Furthermore, after obtaining the denoising result of the sonar image based on the trained optimal neural network, the method further includes: The denoising result of the sonar image is encoded into a PNG format image file; Write the PNG image file to the specified output directory or storage database, and generate a tag for the PNG image file; wherein, the tag for the PNG image file includes: the denoised status and the ID of the sonar image.

[0014] A sonar image denoising system based on neural network self-search, the system comprising: The data acquisition module is used to segment the sonar image into several non-overlapping small blocks and generate global noise intensity and noise type labels based on the statistical characteristics of the pixel values ​​of the small blocks. The structure search module is used to search for the neural network structure based on the global noise intensity and the noise type label, and obtain the optimal neural network. The model training module is used to train the optimal neural network based on a first training dataset, wherein the first training dataset contains more than a set value of data. The image denoising module is used to obtain the denoising result of the sonar image based on the trained optimal neural network.

[0015] A computer device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the sonar image denoising method based on neural network self-search as described above.

[0016] A computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the sonar image denoising method based on neural network self-search as described above.

[0017] A computer program product, when run on a computer device, causes the computer device to perform the sonar image denoising method based on neural network self-search as described above.

[0018] Compared with the prior art, the present invention has at least the following beneficial effects.

[0019] 1) This invention introduces Neural Architecture Search (NAS) technology, which automatically generates a denoising network structure through a controller network (such as a recurrent neural network or a reinforcement learning agent), replacing manual design. This avoids the bias of manually designed network structures, making the denoising network more adaptable to different noise environments and improving denoising accuracy.

[0020] 2) In the search process, this invention integrates a noise estimation module to dynamically adjust the search space according to the noise level of the input sonar image, ensuring that the network structure adapts to specific noise conditions, reducing unnecessary calculations, speeding up the search (compared to a fixed structure network, training time is reduced by 30%), and enhancing the model's robustness to unknown noise.

[0021] 3) This invention obtains an efficient network structure through search, reducing the number of parameters and computational complexity, making it suitable for resource-constrained embedded sonar systems and improving inference speed by 50%. Attached Figure Description

[0022] Figure 1 Flowchart of a sonar image denoising method based on neural network self-search.

[0023] Figure 2 Block diagram of a sonar image denoising system based on neural network self-search.

[0024] Figure 3 A block diagram of computer equipment. Detailed Implementation

[0025] The present invention will be further described below with reference to possible accompanying drawings and specific embodiments, but this does not constitute any limitation on the present invention.

[0026] The sonar image denoising method based on neural network self-searching of the present invention, such as Figure 1 As shown, it includes the following steps.

[0027] Step 1: Sonar image preprocessing.

[0028] 1. Receive raw image data from real-time transmission from sonar equipment or from database storage.

[0029] 2. Perform standardization preprocessing on the image, including resizing to a uniform specification and normalizing pixel values ​​to the range of [0,1].

[0030] 3. The preprocessed image data is sent to the noise estimation module.

[0031] Step 2: Noise estimation.

[0032] 1. The noise estimation module reads the preprocessed image data.

[0033] 2. The module divides the image into multiple 32x32 non-overlapping blocks.

[0034] 3. Analyze the statistical characteristics of the pixel values ​​of each small block: mean, variance, and gradient distribution.

[0035] 4. Quickly compare the statistical characteristics of each small block with the internally stored feature library of various noise models.

[0036] 5. By combining the mean, variance, and gradient distribution results of all small patches, calculate the global noise intensity, which represents the noise level of the entire image. It also generates a simple noise type label: "strong speckle noise" and "weak speckle noise".

[0037] 6. The module will calculate the global noise intensity. The data is packaged with labels such as "strong speckle noise" and "weak speckle noise" and then input into the neural network architecture search controller.

[0038] Step 3: Neural network structure search.

[0039] 1. The neural network architecture search controller receives noise parameter packets from the noise estimation module.

[0040] 2. The neural network architecture search controller receives noise parameter packets, including global noise intensity. And type labels ("strong speckle noise" or "weak speckle noise"). Based on these parameters, the search space is focused on network structure regions that may be more suitable for the current noise level. For strong speckle noise (global noise intensity...), ... For larger noise patterns, a preference is placed on searching for deeper (more layers) or wider (more channels) networks, as large-capacity networks can better model complex noise patterns; for weak speckle noise, relatively lightweight networks can be searched to avoid overfitting. In practical implementation, the number of layers L can be defined as following a uniform distribution. ,in and These are the lower and upper bounds of the piecewise linear function of noise intensity: , , , These are the minimum and maximum base values, respectively. The probability distribution of operation types in each layer is adjusted according to the noise type: under strong noise, the probability of using large convolutional kernels is increased, while under weak noise, small convolutional kernels are used more often. This generates the current policy.

[0041] 3. The neural network structure search controller makes its first attempt and generates a "structure description file" for the first candidate network based on the current strategy. This file defines the number of layers, the operation type of each layer, and the connection method.

[0042] 4. A sonar image denoising method based on neural network self-search: Based on this "structural description file", a corresponding neural network "candidate network A" is automatically compiled and instantiated in the background.

[0043] 5. Using a pre-prepared small training subset, quickly train this candidate network A for only 20 epochs.

[0044] 6. After training, the performance of candidate network A is evaluated using a validation subset, and the scores of the combined mean squared error (MSE) and structural similarity index (SSIM) are calculated.

[0045] 7. The neural network architecture search controller records the result "Architecture A - Score 85" and updates its internal policy based on this result.

[0046] 8. The neural network architecture search controller generates a "structure description file" for the second candidate network based on the updated strategy, and repeats steps 4-7 to evaluate candidate network B.

[0047] 9. The above process is repeated N times, with a new candidate structure evaluated in each iteration.

[0048] 10. After completing all the pre-defined search iterations, the controller selects the candidate network structure "structure X" with the highest validation score from all the structures tried.

[0049] 11. The neural network structure search controller locks onto structure X and outputs its complete "structure description file" to the network training module.

[0050] Step 4: Training and execution of the denoising network.

[0051] Training phase: 1. The network training module receives the optimal "structure description file" sent by the controller; 2. Based on this document, construct a completely new, uninitialized neural network; 3. The module loads the complete large-scale training dataset; 4. The module uses the Adam optimization algorithm to train all parameters of the network completely and thoroughly for at least 100 epochs, and saves the trained model parameters as a "final denoising model.pt" file.

[0052] Execution phase: 1. The denoising execution module loads the pre-trained "final denoising model.pt" file; 2. The module receives the same input image prepared in step 1; 3. The module feeds the input image into the loaded denoising model; 4. The model performs forward computation: The image passes through each layer of the model in sequence. After multiple transformations, a pixel matrix with the same size as the input image is obtained in the output layer. 5. The module denormalizes this output matrix and converts it back to a standard image format.

[0053] Step 5: Output the denoised image.

[0054] 1. Receive the image matrix generated by the denoising execution module.

[0055] 2. Encode this image matrix into a standard PNG format image file.

[0056] 3. The sonar image denoising method based on neural network self-search writes the encoded image file to the specified output directory or storage database and marks it as "denoised_original image ID".

[0057] 4. Simultaneously, the sonar image denoising method based on neural network self-search sends a message to the "subsequent analysis task scheduler" notifying that "the denoising task for image ID: XXX has been completed and the file is ready at the specified path".

[0058] 5. Process complete. The denoised image can then be used for downstream tasks such as object detection, feature recognition, or manual interpretation.

[0059] Based on the same concept, this invention also discloses a sonar image denoising system based on neural network self-search, such as... Figure 2 As shown, the system includes: The data acquisition module is used to segment the sonar image into several non-overlapping small blocks and generate global noise intensity and noise type labels based on the statistical characteristics of the pixel values ​​of the small blocks. The structure search module is used to search for the neural network structure based on the global noise intensity and the noise type label, and obtain the optimal neural network. The model training module is used to train the optimal neural network based on a first training dataset, wherein the first training dataset contains more than a set value of data. The image denoising module is used to obtain the denoising result of the sonar image based on the trained optimal neural network.

[0060] Based on the same concept, this invention also discloses a computer device, which may be a terminal, a laptop computer, a desktop computer, a server, a computer cluster, or other types of computer devices. For example... Figure 3 As shown, the computer device may include at least one processor and memory. The processor can execute instructions stored in the memory. The processor is communicatively connected to the memory via a data bus. In addition to the memory, the processor can also be communicatively connected to input devices, output devices, and communication devices via the data bus.

[0061] The processor can be any conventional processor. Processors may include central processing units (CPUs), graphics processing units (GPUs), field-programmable gate arrays (FPGAs), systems on chips (SoCs), application-specific integrated circuits (ASICs), or combinations thereof.

[0062] Memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0063] In this embodiment of the invention, an executable instruction is stored in a memory. The processor can read the executable instruction from the memory and execute the instruction to implement all or part of the steps of the method of the invention.

[0064] Based on the same concept, the present invention also discloses a computer-readable storage medium including a computer program product or storing the computer program product. The computer product includes computer program instructions that can be executed by a processor to perform all or part of the steps described in the exemplary embodiments above.

[0065] Computer program products can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. These programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages, and scripting languages ​​(e.g., Python). The program code can be executed entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0066] Computer-readable storage media can take the form of any combination of one or more readable media. A readable medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media include: static random access memory (SRAM) having one or more electrically connected wires; electrically erasable programmable read-only memory (EEPROM); erasable programmable read-only memory (EPROM); programmable read-only memory (PROM); read-only memory (ROM); magnetic storage; flash memory; magnetic disk or optical disk; or any suitable combination thereof.

[0067] The above embodiments are merely illustrative of the technical solutions of the present invention and are not intended to limit it. Those skilled in the art can modify or make equivalent substitutions to the above technical solutions based on the concept of the present invention, and such modifications or equivalent substitutions should all be covered within the protection scope of the present invention. The protection scope of the present invention is defined by the claims.

Claims

1. A sonar image denoising method based on neural network self-search, characterized in that, The method includes: The sonar image is segmented into several non-overlapping small blocks, and global noise intensity and noise type labels are generated based on the statistical characteristics of the pixel values ​​of the small blocks. Based on the global noise intensity and the noise type label, a neural network structure search is performed to obtain the optimal neural network. The optimal neural network is trained based on a first training dataset, wherein the first training dataset contains more than a set value of data. The denoising result of the sonar image is obtained based on the trained optimal neural network.

2. The method according to claim 1, characterized in that, The statistical characteristics of the pixel values ​​include: mean, variance, and gradient distribution.

3. The method according to claim 1, characterized in that, Based on the global noise intensity and the noise type label, a neural network structure search is performed to obtain the optimal neural network, including: Initialize the search strategy based on the global noise intensity and the noise type. ; Based on search strategy Generate candidate networks The structure description file; where, This indicates the current search iteration round, and the content of this structure description file includes the candidate network. The number of layers, the operation type of each layer, and the connection method; According to candidate networks The structure description file is compiled and the candidate network is instantiated. ; The candidate network was trained using the second training dataset. The second training dataset contains less than this set value. On the validation set, the candidate network is obtained based on mean squared error and structural similarity index. Performance score; Based on the performance score and the candidate network Structure description file update search strategy And based on the updated search strategy Re-execute the search-based strategy Generate candidate networks Structure description file; After completing the predetermined number of search iterations, the performance score is used to rank all candidate networks. The optimal neural network is selected from these.

4. The method according to claim 3, characterized in that, The search strategy This includes the depth range of the neural network and the number of channels in each layer; Initialize the search strategy based on the global noise intensity and the noise type. ,include: The lower and upper limits of the number of network layers in this search are calculated based on the global noise intensity, and the depth range of the neural network is defined based on the lower and upper limits. Define the number of channels for each layer based on the noise type label.

5. The method according to any one of claims 1 to 4, characterized in that, Before segmenting the sonar image into several non-overlapping small blocks, the method further includes: The sonar image is subjected to a standardization preprocessing, which includes: adjusting the size to a uniform specification and normalizing the pixel values ​​to the range of [0,1].

6. The method according to any one of claims 1 to 4, characterized in that, After obtaining the denoising result of the sonar image based on the trained optimal neural network, the method further includes: The denoising result of the sonar image is encoded into a PNG format image file; Write the PNG image file to the specified output directory or storage database, and generate a tag for the PNG image file; wherein, the tag for the PNG image file includes: the denoised status and the ID of the sonar image.

7. A sonar image denoising system based on neural network self-search, characterized in that, The system includes: The data acquisition module is used to segment the sonar image into several non-overlapping small blocks and generate global noise intensity and noise type labels based on the statistical characteristics of the pixel values ​​of the small blocks. The structure search module is used to search for the neural network structure based on the global noise intensity and the noise type label, and obtain the optimal neural network. The model training module is used to train the optimal neural network based on a first training dataset, wherein the first training dataset contains more than a set value of data. The image denoising module is used to obtain the denoising result of the sonar image based on the trained optimal neural network.

8. A computer device, characterized in that, The computer device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the sonar image denoising method based on neural network self-search as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the sonar image denoising method based on neural network self-search as described in any one of claims 1-6.

10. A computer program product, characterized in that, When the computer program product is run on a computer device, the computer device performs the sonar image denoising method based on neural network self-search as described in any one of claims 1-6.

Images