A sonar image classification method based on neural architecture search

Abstract

The application discloses a sonar image classification method based on neural architecture search, first, the sonar images obtained by the forward-looking sonar, three-dimensional imaging sonar and side-scan sonar are pretreated and screened; the sonar image data are divided into three subsets of a training set, a validation set and a test set; based on the characteristics of the sonar image, the search space and the search strategy are improved, the convolution block SIC Block with multi-scale feature extraction capability is proposed, and the DARTS-SIC algorithm is obtained; the subsets are input into the DARTS-SIC for neural architecture search, the best parameters are learned from the search space, and the best neural architecture is determined; the target sonar images of each class in the training set, the validation set and the test set are input into the best neural architecture, the model is retrained and tested, the test set recognition accuracy of each class of target is counted, and the final classification result is obtained. The application can effectively solve the problems of low efficiency of the model designed by artificial and low model recognition accuracy in the traditional method, and has a wide application prospect.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, specifically relating to a sonar image classification method based on neural architecture search. Background Technology

[0002] Optical imaging is of great value in Automatic Target Recognition (ATR), but its application in low-light and turbid underwater environments is severely limited. Imaging sonar continuously transmits and receives sonar signals during operation, presenting rich visual information within the observation area to achieve target detection. With the development of Imaging Sonar Systems (ISS), the application of sonar images in underwater target detection and recognition plays a crucial role in both military and civilian fields. In recent years, a series of classification algorithms based on sonar images have emerged in the field of underwater target recognition. These algorithms can generally be divided into two basic stages: feature extraction and target classification. However, traditional methods heavily rely on expert systems and lack adaptability to multi-scale features, resulting in problems such as weak robustness and low recognition rates. Furthermore, the human-involved feature extraction and classifier design process inevitably leads to the loss of some features and consumes a significant amount of human resources.

[0003] With the powerful feature extraction capabilities of Convolutional Neural Networks (CNNs), Deep Learning (DL) has become a compelling area of ​​innovation. However, manually designing CNN models is relatively inefficient due to the need to fine-tune numerous network parameters through ablation experiments. With the rapid development of Automated Machine Learning (AutoML), researchers are increasingly interested in Neural Architecture Search (NAS). However, research linking NAS to sonar image classification is relatively scarce compared to other fields, and NAS algorithms developed for other image classification tasks may not be suitable for sonar image datasets. Developing a NAS algorithm suitable for sonar image features, while simultaneously considering classification performance and adaptability to small sample sizes, is crucial. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a sonar image classification method based on neural architecture search. First, sonar images obtained from forward-looking sonar, 3D imaging sonar, and side-scan sonar are preprocessed and filtered. The sonar image data is divided into three subsets: a training set, a validation set, and a test set. Based on the characteristics of sonar images, the search space and search strategy are improved, and a convolutional block (SIC) with multi-scale feature extraction capabilities is proposed, resulting in the DARTS-SIC algorithm. The subsets are input into the DARTS-SIC algorithm for neural architecture search, learning the optimal parameters from the search space and determining the optimal neural architecture. The sonar images of various targets in the training, validation, and test sets are then fed into the optimal neural architecture for model retraining and testing. The accuracy of target recognition on the test set is statistically analyzed to obtain the final classification result. This invention effectively solves the problems of low efficiency and low recognition accuracy of manually designed models in traditional methods and has broad application prospects.

[0005] The technical solution adopted by this invention to solve its technical problem is as follows:

[0006] Step 1: Preprocess and filter the sonar images obtained from forward-looking sonar, 3D imaging sonar, and side-scan sonar;

[0007] Step 2: Divide the sonar image data obtained in Step 1 into three subsets: training set, validation set, and test set. Further divide the training set into training search subset and validation search subset.

[0008] Step 3: Based on the characteristics of sonar images, improve the search space and search strategy, propose a convolutional block SIC Block with multi-scale feature extraction capability, and obtain the DARTS-SIC algorithm;

[0009] Step 4: Input the labeled training search subset and validation search subset samples into DARTS-SIC to search for neural architectures, learn the optimal parameters from the search space, and determine the optimal neural architecture.

[0010] Step 5: Input the sonar images of various targets in the training set, validation set, and test set into the optimal neural architecture obtained in Step 4, retrain the model, test it, and calculate the test set recognition accuracy of various targets to obtain the final classification result.

[0011] Further, step 1 specifically includes:

[0012] The image size was reset to [224, 224] using Lanczos interpolation; the original RGB image was converted to a grayscale image through grayscale processing, and weights were assigned based on the overall tone of the sonar image and the true tone of the target; noise reduction of the sonar image was achieved through a noise reduction method combining spatial domain and transform domain, and the images were filtered based on image quality and the degree of feature prominence.

[0013] Furthermore, step 2 specifically includes:

[0014] The sonar image data obtained in step 1 is divided into three subsets: 70% of the samples are assigned to the training set D. train 15% of the samples were assigned to the validation set D val The remaining 15% of the samples were allocated to the test set D. test and the corresponding tag set Y train Y val and Y test ; train set D train Further divided into training and search subsets D train-s and verify the search subset D val-s .

[0015] Furthermore, step 3 specifically includes:

[0016] The DARTS-SIC algorithm consists of two stages: a search stage and a verification stage. The input scale for both stages is [112, 112, 3], and the output is a 5-dimensional one-hot vector output by the softmax function.

[0017] The optimal units, namely ordinary connection units and jump connection units, are obtained by searching the optimized candidate operation space. The structure search process is regarded as a continuous relaxation of the architecture representation. The gradient descent method is used to search the architecture effectively. In each round of search, only the k architectures sampled from all candidate architectures are evaluated to assess their recognition rate. The sampling operation effectively reduces the computational resource consumption during the search.

[0018] Furthermore, step 4 specifically includes:

[0019] The neural architecture search includes: inputting labeled training and validation subsets into the DARTS-SIC algorithm to construct a supernetwork containing multiple candidate operations to perform the architecture search, where each operation represents a possible structural unit that the network may adopt; the search process utilizes gradient-based methods to optimize the weights and architecture parameters in the supernetwork, aiming to discover a network architecture that maximizes the target performance metrics; after the search is completed, based on the final values ​​of the architecture parameters, the operation with the highest weight is selected to determine the final network architecture; the target performance metrics include model complexity and classification accuracy.

[0020] Furthermore, step 5 specifically includes:

[0021] The specific steps for calculating the test set recognition accuracy for various target types are as follows: TP (predicted positive, actual positive); TN (predicted negative, actual negative); FP (predicted positive, actual negative); FN (predicted negative, actual positive); after obtaining TP, TN, FP, and FN, the accuracy is calculated using the formula... Calculate the recognition accuracy on the test set; where TP stands for True Positive: the number of times a positive class is correctly predicted as a positive class; TN stands for True Negative: the number of times a negative class is correctly predicted as a negative class; FP stands for False Positive: also known as Type I error, the number of times a negative class is incorrectly predicted as a positive class; and FN stands for False Negative: also known as Type II error, the number of times a positive class is incorrectly predicted as a negative class.

[0022] The beneficial effects of this invention are as follows:

[0023] This invention automates the design of neural network architectures, reducing the need for manual intervention and expert knowledge, and simplifying the model development process. Compared to traditional algorithms, this invention significantly improves classification accuracy on small-sample sonar image datasets. By introducing SIC Blocks, it can extract sonar image features at different scales, effectively enhancing the model's feature extraction capabilities. The test set classification accuracy outperforms other similar methods. It effectively addresses the inefficiencies of manual model design and low model recognition accuracy in traditional methods, and has broad application prospects. Attached Figure Description

[0024] Figure 1 This is a diagram of the DARTS-SIC algorithm framework of the present invention.

[0025] Figure 2 This section presents the number of samples and examples of various types of sonar image datasets used in embodiments of the present invention.

[0026] Figure 3 These are the test set recognition accuracy results of the algorithms in the embodiments of the present invention on the sonar image dataset.

[0027] Figure 4 The model complexity (FLOPs) of each algorithm in the embodiments of the present invention.

[0028] Figure 5 (a)(b)(c)(d) are, in turn, the confusion matrices of the classification results of each algorithm in the embodiments of the present invention.

[0029] Figure 6This is a diagram of the Normal unit structure obtained by DARTS-SIC search in an embodiment of the present invention.

[0030] Figure 7 This is a structural diagram of the Reduction unit obtained by DARTS-SIC search in an embodiment of the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] This invention addresses the shortcomings of manually designing models based on expert systems by proposing a sonar image classification method utilizing an optimized neural architecture search algorithm. This method, through dynamic design of the network model architecture, effectively enhances the category feature extraction capability of sonar images while maintaining model parameter efficiency. This invention, to some extent, solves the problem of poor classification performance and low accuracy of deep learning models on small sample datasets.

[0033] The technical solution adopted by this invention to solve its technical problem is as follows: A deep learning model is obtained through a neural architecture search algorithm, and multi-dimensional category features are extracted from underwater target sonar images, thereby achieving accurate classification of different categories of sonar images. This invention mainly includes the following steps:

[0034] (1) Preprocess and screen the sonar images obtained from forward-looking sonar, three-dimensional imaging sonar and side-scan sonar collected and organized from online channels.

[0035] (2) Divide the sonar image data constructed in (1) into three subsets: training set, validation set and test set, and further divide the training set into training search subset and validation search subset.

[0036] (3) Based on the characteristics of sonar images, the search space and search strategy are improved, and a convolutional block (SIC Block) with multi-scale feature extraction capability is proposed, resulting in the DARTS-SIC algorithm.

[0037] (4) Input the labeled training search subset and validation search subset samples into DARTS-SIC to search for neural architectures, learn the best parameters from the search space, and determine the best neural architecture.

[0038] (5) Input the sonar images of various targets in the training set, validation set and test set into the optimal neural architecture obtained by (4), retrain the model and test it, and statistically analyze the test set recognition accuracy of various targets to obtain the final classification result.

[0039] The specific methods for preprocessing and filtering in step (1) are as follows: the image size is reset to [224, 224], and the interpolation method used is Lanczos interpolation; the original RGB image is converted to a grayscale image through grayscale processing, and in actual processing, weight allocation is performed according to the overall tone of the sonar image and the true tone of the target; the purpose of image denoising is to filter out noise interference in the image and retain the texture, edge and other information of the sonar image as much as possible. This invention achieves sonar image denoising through a denoising method combining spatial domain and transform domain, and filters according to image quality and feature obviousness.

[0040] The specific method for constructing the sonar image dataset of underwater targets in step (2) is to divide all sample images in the dataset established in this invention into three subsets: 70% of the samples are allocated to the training set (D). train ), 15% of the samples were assigned to the validation set (D). val The remaining 15% of the samples were allocated to the test set (D). test ), and the corresponding tag set Y train Y val and Y test To establish the dataset for the search phase, this invention will use the training set (D) train ) is further divided into training search subsets (D train-s ) and validation search subset (D val-s ).

[0041] The DARTS-SIC algorithm described in step (3) includes two stages: a search stage and a verification stage. The input scale for both stages is [112, 112, 3], and the output is a 5-dimensional one-hot vector output by the softmax function. The optimal unit, namely the normal cell and the reduction cell, is obtained by searching the optimized candidate operation space. This invention treats the structure search process as a continuous relaxation of the architecture representation and uses gradient descent to effectively search the architecture. In each round of search, only k architectures sampled from all candidate architectures are evaluated to assess their recognition rate. The sampling operation effectively reduces the computational resource consumption during the search. The algorithm framework is as follows: Figure 1 As shown.

[0042] The neural architecture search phase described in step (4) includes: inputting labeled training and validation search subsets into the DARTS-SIC algorithm; constructing a supernetwork containing multiple candidate operations based on the idea of ​​differential architecture search to perform the architecture search, where each operation represents a structural unit that the network may adopt (such as different types of convolution or pooling operations). The search process utilizes gradient-based methods to optimize the weights and architecture parameters in the supernetwork, aiming to discover a network architecture that maximizes the target performance metrics (model complexity and classification accuracy in this invention). After the search is completed, the operation with the highest weight is selected to determine the final network architecture based on the final values ​​of the architecture parameters.

[0043] The specific steps for calculating the accuracy of the test set identification in step (5) are as follows: TP: predicted positive, actual positive; TN: predicted negative, actual negative; FP: predicted positive, actual negative; FN: predicted negative, actual positive. After obtaining the above TP, TN, FP, and FN, the accuracy is calculated using the formula... Calculate the recognition accuracy of the test set.

[0044] Example:

[0045] All experiments in this invention were conducted on an Intel Xeon Platinum 8373C 2.6GHz CPU with 512GB of memory and two NVIDIA™ processors. The experiment was conducted on a GeForce RTX 3090 graphics processing unit (GPU). The operating system was CentOS Linux version 7.9.2009 (kernel), the code used was Python 3.9, and the deep learning framework was PyTorch 2.0. During feature extraction, the parameters of each layer of the neural network were configured in Anaconda Spyder. After configuration, the network model's search, training, and forward propagation files, written in Python, were compiled into an executable file to implement the network model's functionality. CUDA was used for GPU parallel acceleration during the experiment. The specific implementation method is as follows:

[0046] 1. During the construction of the sonar image dataset, multiple publicly available sonar image datasets were collected, totaling 476 sonar images across 5 categories: aircraft, tires, pipes, rocks, and bottles. Details and examples of each category are provided below. Figure 2 As shown. The dataset not only meets the requirements for architecture search in terms of sample quantity, but also contains sonar images with various imaging methods, which is beneficial for verifying the generalization of the network model. Lanczos interpolation is implemented by calling the OpenCV function library, and the truncated image is then set to [224, 224, 3]. The RGB image is converted to a grayscale image using the NumPy function library, with the conversion formula: I Gray= a×R+b×G+c×B; In order to eliminate the adverse effects caused by singular sample data, the pixel values ​​are normalized after grayscale processing, and the pixel values ​​are normalized from [0,255] to [0,1]; Through the typical block matching three-dimensional collaborative filtering algorithm - BM3D image denoising algorithm, the image is divided into blocks of a certain size. According to the similarity between image blocks, two-dimensional image blocks with similar structures are combined together to form a three-dimensional array. Then, the joint filtering method is used to process these three-dimensional arrays in the transform domain. Finally, through inverse transformation, the processed result is returned to the original image position, thereby obtaining the denoised image.

[0047] 2. Based on the stratified sampling rule, the obtained five types of target sonar image datasets were partitioned using the `random` and `shutil` libraries, distributing the image data into three different subsets: a training set comprising 70% of the total samples (333 sonar images); a validation set and a test set each comprising 15% of the total samples (71 and 72 sonar images respectively). Furthermore, the training set was further divided into a training search subset and a validation search subset at a 70%:30% ratio to facilitate application in the Neural Architecture Search (NAS) stage. The resulting training search subset and validation search subset contained 233 and 100 images respectively.

[0048] 3. The operation flow of the DARTS-SIC algorithm proposed in this invention based on DARTS (Differentiable Architecture Search) is as follows:

[0049] First, we define a discrete search space for neural network architectures. This includes a series of optional neural network operations (such as convolution, pooling, shortcut branching, etc.). Learnable weight parameters w are introduced to transform the selection weights of candidate operations into differentiable parameterized functions, where w represents the set of parameters for the selection weights. The neural network architecture is viewed as a directed acyclic graph (DAG), where nodes represent different layers or operations, and edges represent connections between operations. A hypergraph is then constructed using the parameterized functions and weights w. To represent the candidate structure. Weight optimization is performed using the training dataset, fine-tuning the parameters w by minimizing the loss function, and employing gradient descent for optimization. In each training iteration, the hypergraph is optimized... A successive operation selection process is performed to determine the operations that need to be retained in the final neural network structure. This step is differentiable, allowing the use of gradient information to update the weights of the parameterized function. The network performance is evaluated using a validation dataset to select the best-performing neural network structure. Finally, the above iterative process is repeated until a satisfactory neural network structure is found or the predetermined number of search iterations is reached.

[0050] 4. To enhance the network model's ability to extract features at different scales, this invention proposes a convolutional block structure called SICBlock. In SICBlock, the input feature map is fed into a general 3×3 convolutional layer and a specified k×k convolutional layer, where k is set to 3, 5, or 7. While maintaining the same spatial dimension and number of channels, the outputs of each branch are summed by channel. The fused feature map is then processed by Layer Normalization (LN) and the Swish activation function, and then output through a Channel Attention Block (SE). Finally, they are output through a 1×1 convolutional layer, an LN layer, and a dropout layer. Notably, compared to the batch normalization (BN) layer commonly used in CNN architectures, LN is more suitable for addressing the problem of large statistical bias in normalization when dealing with small sample sizes. The aforementioned characteristics of the LN layer are highly compatible with the small sample task scenario encountered in this study; therefore, the invention uses an LN layer instead of a BN layer to achieve feature matrix normalization when constructing SICBlock. The SE layer gives the network the ability to fine-tune important features in the input data. This keen attention mechanism not only enhances the model's representational ability but also improves its performance in complex visual recognition tasks.

[0051] Furthermore, this invention incorporates the following operations into the search: 3×3 separable convolutional layers, 3×3 dilated separable convolutional layers, 3×3 max-pooling layers, 3×3 average-pooling layers, and shortcut branches. All operations have a stride of 1 (where applicable), and the convolutional feature maps are padded to maintain their spatial resolution. Units located at 1 / 3 and 2 / 3 of the total network depth are shrunk units, where the stride of each operation adjacent to the input node is 2.

[0052] 5. In the neural architecture search phase, labeled training and validation subsets are input into the DARTS-SIC algorithm. The algorithm learns the optimal parameters using these data and automatically optimizes the neural network architecture. Based on the ideas of DARTS, this invention transforms the neural network architecture search problem into a differentiable optimization problem, thereby enabling the possibility of using gradient descent for search. This method significantly reduces the time and computational resources required for the search and improves the efficiency of neural network architecture search. The specific operation is as follows: In formula (1), the mixed weights of a pair of nodes are defined as parameterized by the vector α(i,j), the neural architecture α is the upper-level variable, and the architecture model weights w are the lower-level variables.

[0053]

[0054]

[0055] In this invention, the symbol This represents the loss function. and Let represent the training loss and validation loss, respectively. This two-layer optimization problem is computationally intensive when calculating gradients; therefore, an approximation method is used in DARTS-SIC to solve this problem, as shown in equation (2):

[0056]

[0057] ε can be chosen in practice (it is 0 if a first-order approximation is used). With a suitable choice of ξ, the optimization algorithm can be guaranteed to converge to a fixed point. The two-layer optimization process is replaced by an alternating update optimization process through approximate calculation. First, the calculated model network weights w are iteratively updated to obtain w. * Then use w * As a constant, α is continuously updated to obtain the new operation weight α. * Then use the obtained α * The model network weights are further updated, and this process is repeated until the optimal model architecture is found.

[0058] 6. The configuration of experimental hyperparameters for the search and training phases is shown in Table 1. In the search phase, this invention utilizes the hyperparameters derived from D... train The two subsets are split, and the optimal parameter α is learned from the search space based on the search strategy in Section 5. After determining the optimal architecture through the optimal α obtained by the search, the algorithm will call D. train The samples in the table are used for retraining. During training, the modules obtained in the search phase are stacked according to the parameter settings in Table 1 to form the final model structure, and at the end of each duration, D is used. val The dataset is used to validate the trained model and save the corresponding predicted labels and Y values. val The comparison results obtained from the validation set are used to guide the direction of model training iterations. Finally, the saved optimal model weights are used to evaluate D. test Classify and classify according to the corresponding label set Y test To evaluate the recognition accuracy of the model, a total of 3 test experiments were conducted for each group. To verify the effectiveness of the algorithm of this invention on sonar image datasets, training experiments were carried out using a convolutional network model composed of ResNet-18, image salient region segmentation (GMR) and pyramid pooling (SPP), and a DARTS model under the same environment and standards as described above.

[0059] Table 1. Configuration of experimental hyperparameters

[0060]

[0061] from Figure 3The classification results of the sonar dataset established based on this invention show that the manually designed classic CNN model (ResNet-18) achieved the highest test set classification accuracy of 77.31%; the highest classification accuracy using the unoptimized DARTS method was 92.13%; and the convolutional network model composed of image salient region segmentation (GMR) and pyramid pooling (SPP) achieved the best average test set accuracy of 90.90%. The model structure obtained through DARTS-SIC search achieved the highest and average classification accuracies of 94.91% and 94.43% respectively on the test dataset. Furthermore, based on the results of the three sets of experiments, the NAS method outperformed the manually designed CNN model, meaning that in some cases, manual methods are not always the optimal solution. Figure 4 As shown, compared with ResNet-18 and GMR+SPP models, the algorithm of this invention has significant advantages. However, since SIC Block will increase the number of parameters to a certain extent, the number of FLOPs is higher than that of DARTS.

[0062] Figure 5 The confusion matrix is ​​displayed. Elements on the diagonal represent the accuracy for each class; higher values ​​indicate better performance. Elements outside the diagonal represent the misclassification rate; lower values ​​are better. Based on the confusion matrix distribution of DATS-SIC in three test experiments, elements from each class are largely clustered on the diagonal, showing significantly improved diagonal clustering compared to other models, especially ResNet-18, which is based on manually designed optical images. The results indicate that the DARTS-SIC method improves the classifier's accuracy on easily confused categories (bridge piers, stones) and reduces the occurrence of confusion between the two classes.

[0063] To demonstrate the contribution of the SIC Block layer to the model performance, this invention performs the following steps on the obtained cells: Figure 6 and Figure 7The visualization analysis is shown below. The visualization results indicate that the Normal units obtained by the DARTS-SIC method tend to combine SICBlock layers, depthwise separable convolutional layers, and dilated separable convolutional layers in an ordered manner. Furthermore, combining shortcut branches with other convolutional operations in the search space (such as depthwise separable convolution and dilated convolution) forms a residual structure that helps with training and accelerates model convergence. For the Reduction unit, it can be observed that the modules obtained by the DARTS-SIC method retain more jump connections in the deep connections to ensure that the Reduction unit has the ability to solve the vanishing and exploding gradient problems. The addition of the SIC Block layer enhances the multi-scale feature extraction capabilities of the Normal and Reduction units obtained by the DARTS-SIC search and provides more learnable parameters, thereby ensuring the final learning expressive power.

Claims

1. A sonar image classification method based on neural architecture search, characterized in that, Includes the following steps: Step 1: Preprocess and filter the sonar images obtained from forward-looking sonar, 3D imaging sonar, and side-scan sonar; Step 2: Divide the sonar image data obtained in Step 1 into three subsets: training set, validation set, and test set. Further divide the training set into training search subset and validation search subset. Step 3: Based on the characteristics of sonar images, improve the search space and search strategy, propose a convolutional block SIC Block with multi-scale feature extraction capability, and obtain the DARTS-SIC algorithm; Step 3-1: Define the discrete neural network architecture search space , including a series of optional neural network operations; introduce learnable weight parameters w , to weight the selection of candidate operations as a differentiable parameterized function; consider the neural network architecture as a directed acyclic graph (DAG), where nodes represent different layers or operations, and edges represent the connection relationship between operations; then construct a hypergraph w to represent the candidate structure by using the parameterized function and weight parameters ; perform weight optimization using the training data set, adjust the parameters by minimizing the loss function w , and use the gradient descent method for optimization; In each training iteration, by analyzing the hypergraph Perform sequential operation selection to determine the operations that need to be retained in the final neural network structure; The weights of the parameterized function are updated using gradient information, and the network performance is evaluated using a validation dataset to select the best-performing neural network structure. Finally, the above iterative process is repeated until a neural network structure that meets the requirements is found or the predetermined number of search iterations is reached. Step 3-2: In the SIC Block, the input feature map is fed into a general 3×3 convolutional layer and a specified... k ×k Convolutional layers, in which k Set to 3, 5, or 7; while maintaining the same spatial dimension and number of channels, sum the channels of the output of each branch; after the fused feature map is normalized by the layer and activated by the Swish activation function, it is output through the channel attention block; finally, it is output through a 1×1 convolutional layer, an LN layer, and a dropout layer. Step 3-3: The following operations were added to the search: 3×3 separable convolutional layer, 3×3 dilated separable convolutional layer, 3×3 max pooling layer, 3×3 average pooling layer and shortcut branch; the stride of all operations was 1, and the feature map after convolution was padded to maintain its spatial resolution. The cells located at 1 / 3 and 2 / 3 of the total network depth are shrinking cells, where the step size of each operation adjacent to the input node is 2; Step 4: Input the labeled training search subset and validation search subset samples into DARTS-SIC to search for neural architectures, learn the optimal parameters from the search space, and determine the optimal neural architecture. Step 5: Input the sonar images of various targets in the training set, validation set, and test set into the optimal neural architecture obtained in Step 4, retrain the model, test it, and calculate the test set recognition accuracy of various targets to obtain the final classification result.

2. The sonar image classification method based on neural architecture search according to claim 1, characterized in that, Step 1 specifically involves: The image size was reset to [224, 224] using Lanczos interpolation; the original RGB image was converted to a grayscale image through grayscale processing, and weights were assigned based on the overall tone of the sonar image and the true tone of the target; noise reduction of the sonar image was achieved through a noise reduction method combining spatial domain and transform domain, and the images were filtered based on image quality and the degree of feature prominence.

3. The sonar image classification method based on neural architecture search according to claim 2, characterized in that, Step 2 specifically involves: The sonar image data obtained in step 1 is divided into three subsets: 70% of the samples are allocated to the training set. D train 15% of the samples were allocated to the validation set. D val The remaining 15% of the samples were allocated to the test set. D test and the corresponding tag set Y train , Y val and Y test ; training set D train Further divided into training and search subsets D train-s and verify search subset D val-s .

4. The sonar image classification method based on neural architecture search according to claim 3, characterized in that, Step 3 specifically involves: The DARTS-SIC algorithm consists of two stages: a search stage and a verification stage. The input scale for both stages is [112, 112, 3], and the output is a 5-dimensional one-hot vector output by the softmax function. The optimal units, namely ordinary connection units and jump connection units, are obtained by searching the optimized candidate operation space. The structure search process is regarded as a continuous relaxation of the architecture representation, and gradient descent is used to effectively search the architecture. In each round of search, only the architecture sampled from all candidate architectures is calculated. k The architecture is evaluated for its recognition rate, and the sampling operation effectively reduces the consumption of computing resources during the search.

5. The sonar image classification method based on neural architecture search according to claim 4, characterized in that, Step 4 specifically involves: The neural architecture search includes: inputting labeled training and validation subsets into the DARTS-SIC algorithm to construct a supernetwork containing multiple candidate operations to perform the architecture search, where each operation represents a possible structural unit that the network may adopt; the search process utilizes gradient-based methods to optimize the weights and architecture parameters in the supernetwork, aiming to discover a network architecture that maximizes the target performance metrics; after the search is completed, based on the final values ​​of the architecture parameters, the operation with the highest weight is selected to determine the final network architecture; the target performance metrics include model complexity and classification accuracy.

6. The sonar image classification method based on neural architecture search according to claim 5, characterized in that, Step 5 specifically involves: The specific steps for calculating the test set recognition accuracy of various targets are statistical analysis. TP The prediction was positive, and the actual result was positive. TN The prediction was negative, and the actual result was negative. FP The prediction was positive, but the actual result was negative. FN The prediction was negative, but the actual result was positive; thus, the above was obtained. TP, TN, FP and FN Then, through the formula Calculate the recognition accuracy on the test set; Where TP stands for True, which is the number of times a positive class is correctly predicted as a positive class; TN stands for True Negative, which is the number of times a negative class is correctly predicted as a negative class; FP stands for False Positive, also known as Type I error, which is the number of times a negative class is incorrectly predicted as a positive class; and FN stands for False Negative, also known as Type II error, which is the number of times a positive class is incorrectly predicted as a negative class.

Images