A multi-angle sonar image target classification method based on a lightweight convolutional network

Abstract

The present application relates to the field of underwater target recognition, and particularly relates to a multi-angle sonar image target classification method based on a light convolutional network, which comprises the following steps: (1) processing sonar images to obtain a multi-angle sonar image dataset; (2) obtaining an image training set and a verification set; (3) constructing a multi-angle sonar image target classification model based on a light network; (4) iteratively training the above model using the dataset; and (5) obtaining a target classification result. The present application uses multi-angle sonar images to classify targets, which can more comprehensively describe the scattering characteristics of targets compared to single-angle images, thereby enabling attribute discrimination of targets with higher accuracy, and the light network model can also reduce resource consumption, thereby providing the possibility for practical applications.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic signal processing, and in particular to a multi-angle sonar image target classification method based on lightweight convolutional networks. Background Technology

[0002] In recent years, sonar imaging technology has matured significantly and has become a crucial method for acquiring underwater images, finding wide application in both military and civilian fields. Sonar equipment can provide intuitive global information about underwater targets, is independent of underwater light sources, and is highly practical. However, sonar images are mappings of target scattering information into two-dimensional images, and may suffer from aliasing and distortion, making them more difficult to interpret and understand visually. Therefore, identifying the true attributes of targets from sonar images has been a long-standing research focus.

[0003] With the development of deep learning theories, deep learning-based target recognition technology has been widely applied. Many scholars have designed different convolutional networks to detect and recognize targets in sonar images, achieving good results. However, most methods are based on single-angle sonar images. During sonar imaging, for the same observed target, the sonar images obtained at different angles will differ significantly, while different targets will show considerable similarity at a certain angle. Therefore, target recognition based on single-angle sonar images has certain limitations.

[0004] A multi-angle sonar image sequence of the same target contains the scattered echoes of the target at multiple angles, which can more finely characterize the scattering characteristics of the target. By fusing the target acoustic images from multiple perspectives into a single image, the common information of the target under different perspectives can be fully utilized. However, the choice of fusion method and the construction of the target classification network will affect the target classification effect.

[0005] In summary, there is an urgent need for a target classification method that can make full use of target information and can be flexibly deployed to existing equipment in the task of underwater target multi-angle sonar image classification, so as to improve the accuracy and efficiency of subsequent target classification. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and propose a multi-angle sonar image target classification method based on a lightweight convolutional network.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution.

[0008] This invention proposes a multi-angle sonar image target classification method based on a lightweight convolutional network, the method comprising:

[0009] The target's multi-angle sonar images are preprocessed and then fused to obtain multi-angle sonar images of the target.

[0010] The multi-angle sonar images of the target are input into a pre-selected and trained multi-angle sonar image target classification model based on a lightweight convolutional network to obtain the target classification results.

[0011] The multi-angle sonar image target classification model is obtained by improving the lightweight MobileNetV2 network. The improvement of the lightweight MobileNetV2 network includes: removing the network layer after the feature extraction convolutional layer Conv2 in the lightweight MobileNetV2 network, and modifying the number of channels of the feature extraction convolutional layer Conv2; adding a Flatten layer to achieve dimensionality reduction, adding a Dropout layer for random dropout, and finally adding a fully connected layer to classify the results.

[0012] As an improvement to the above technical solution, the multi-angle sonar image target classification model specifically includes: a feature extraction module and a classification module; wherein...

[0013] The feature extraction module includes an input layer (Input), a feature extraction convolutional layer (Conv1), bottleneck layers (Bottleneck1-Bottleneck7), a feature extraction convolutional layer (Conv2), a flatten layer, and a dropout layer; wherein, the feature extraction convolutional layer (Conv2) has 128 channels, and the dropout layer has a dropout rate of 0.5.

[0014] The classification module includes a fully connected layer.

[0015] As an improvement to the above technical solution, the processing procedure of the multi-angle sonar image target classification model specifically includes:

[0016] The Input layer feeds multi-angle sonar images of the target into the feature extraction convolutional layer Conv1;

[0017] The feature extraction convolutional layer Conv1 with 3×3 convolutional kernels extracts feature maps from multi-angle sonar images and inputs them into the bottleneck layers Bottleneck1-Bottleneck7.

[0018] Each bottleneck layer (Bottleneck1-Bottleneck7) includes: a first standard convolution, a depthwise separable convolution, and a second standard convolution. First, the 1×1 first standard convolution performs dimensionality expansion; second, the 3×3 depthwise separable convolution extracts features; and finally, the 1×1 second standard convolution compresses the data. Each bottleneck layer (Bottleneck1-Bottleneck7) processes the feature map sequentially and inputs the processed feature map into the feature extraction convolutional layer (Conv2).

[0019] The number of channels in the feature extraction convolutional layer Conv2 is set to 128. The feature map is then processed and input into the Flatten layer.

[0020] The Flatten layer transforms the feature map into one dimension and inputs it into the Dropout layer;

[0021] The Dropout layer randomly discards feature maps;

[0022] The fully connected layer classifies the target and outputs the target classification result, that is, the probability that the target belongs to a certain category.

[0023] As an improvement to the above technical solution, the first standard convolution is activated using the function ReLU6(x), with the expression:

[0024] ReLU6(x) = min(6, max(x, 0))

[0025] x represents the input value of the activation function, min(·) represents the function that takes the minimum value, and max(·) represents the function that takes the maximum value;

[0026] The second standard convolution is activated using the function Linea(x), expressed as:

[0027] Linear(x) = x.

[0028] As an improvement to the above technical solution, the depthwise separable convolution uses a normalized BN layer for regularization, as expressed in the following expression:

[0029]

[0030] in, The output of the convolutional layer, where m represents the total number of nodes in the current network layer, and x... i μ represents the value of the i-th node in the current network layer. B and Let y represent the mean and variance, respectively. i This is the normalized output result. Represents the normalized x i The values ​​are ε, which represents the regularization factor, and γ and β, which are learnable reconstructable parameters.

[0031] The depthwise separable convolution is activated using the function ReLU6(x), expressed as:

[0032] ReLU6(x) = min(6, max(x, 0)).

[0033] As an improvement to the above technical solution, the method further includes: training a multi-angle sonar image target classification model, the training process specifically including:

[0034] Step 1) Acquire sonar images. After preprocessing the acquired sonar images, fuse multiple consecutive single-channel sonar images to obtain a multi-angle sonar image dataset with the same size, which will be used as the training dataset and the test dataset.

[0035] Step 2) Construct a multi-angle sonar image target classification model based on a lightweight convolutional network;

[0036] Step 3) Use the training dataset and the test dataset to iteratively train and test the multi-angle sonar image target classification model to obtain the trained multi-angle sonar image target classification model.

[0037] As an improvement to the above technical solution, step 1) specifically includes:

[0038] Step 1-1) Obtain N raw sonar images containing K types of targets. All N images are single-channel sonar images. Segment and upsample the images to obtain a set of images of the same size, X1, X2, ..., X... N and the corresponding label vectors Y1, Y2, ..., Y N ;

[0039] Steps 1-2) Divide the N sonar images into K categories and perform image fusion processing according to the categories. Fuse every P consecutive images in each category into a single P-channel image, i.e., X1' = X1 + X2 + ... + X P X2' = X2 + X3 + ... + X P+1 ;

[0040] After processing in step 1-2), steps 1-3) yield a set of images X1', X2', ..., X'. N-2K and the corresponding label vectors Y1', Y2', ..., Y N ' -2K Let them be denoted as X1', X2', ..., X' M With Y1',Y2',...,Y'M ;

[0041] Steps 1-4) Randomly select N1 images and their corresponding label vectors from the image set in Steps 1-3) to form the training dataset D. train The remaining M-N1 images and their corresponding label vectors constitute the test dataset D. test .

[0042] As an improvement to the above technical solution, step 3) specifically includes:

[0043] Step 3-1) For the constructed multi-angle sonar image target classification model based on a lightweight convolutional network, the parameters of each network layer are distributed by a Gaussian distribution G(μ,σ) with a mean of 0 and a variance of 1. 2 Random initialization;

[0044] Step 3-2) Use Optimizers to update the network parameters so that the error between the output image and the label converges to the optimal level;

[0045] Step 3-3) Calculate the model's loss value (Loss) using the cross-entropy loss function;

[0046] Steps 3-4) Use the Softmax classification function to transform the output of the last layer of the network into the probability of the input image belonging to each category. Softmax(z) i This yields the final classification result.

[0047] As an improvement to the above technical solution, the formula for calculating Loss in step 3-3) is:

[0048]

[0049] Where n represents the number of training set samples, K represents the total number of target categories, and q(x) j,k ) represents sample x j,k The predicted category, p(x) j,k ) represents sample x j,k This represents the probability of the current predicted category.

[0050] As an improvement to the above technical solution, in step 3-4), Softmax(z) i The formula for calculating ) is:

[0051]

[0052] Among them, z i This is the output value of the i-th node in the last network layer.

[0053] To enhance the applicability of the model in underwater scenarios, fully leverage the advantages of depthwise separable convolution in feature extraction, and improve the accuracy of target classification, this invention makes the following improvements based on the lightweight MobileNetV2 network: (1) In order to further reduce computational resources and save memory space during network training, while ensuring accuracy, the number of parameters and computational costs are fully considered. Through multiple experimental comparisons, this invention removes the network layers after the 9th layer in the model design and changes the number of channels of this convolutional layer from 1280 to 128. (2) In order to adapt to underwater target classification tasks and improve the convergence rate of the network while retaining target feature information, this invention uses the Flatten layer to reduce the data dimensionality, converts the three-dimensional output into one dimension, adds a Dropout layer to improve network fitting, sets the dropout rate to 0.5, and finally adds a fully connected layer to obtain the final classification result.

[0054] Reasons for improvement:

[0055] (1) The original lightweight MobileNetV2 network was trained using the ImageNet dataset, which contains 1.4 million images across 1000 image categories. However, underwater target images are difficult to acquire, resulting in a limited amount of data. Therefore, directly using the original network for training does not yield a good fit. (2) The output channels of the 9th layer of the MobileNetV2 network were increased from 320 to 1280. This increase in the number of channels consumes more computational resources. (3) The original lightweight network used average pooling (Avgpool) downsampling to reduce the number of features. This downsampling method retains more background information of the image and is not entirely suitable for underwater target classification tasks.

[0056] The advantages of this invention compared to the prior art are:

[0057] 1. By fusing information from single-channel sonar images from multiple angles, underwater targets can be classified using multi-channel sonar images. Compared to sonar images from a single angle, this method can more comprehensively describe the target's scattering characteristics, thereby enabling higher accuracy in target attribute identification.

[0058] 2. In order to reduce the computational cost of image fusion, a lightweight convolutional network is selected as the basic network model for feature extraction and target classification. This model can reduce the number of network parameters and computational cost, and improve the network's efficiency in classifying targets.

[0059] 3. To enhance the applicability of the network in underwater scenarios, the existing lightweight network is improved based on a full consideration of the number of parameters and computational cost. The improved network model can improve the convergence rate of the network while retaining the target feature information, making it more suitable for target classification tasks of sonar images and improving the accuracy of target classification. Attached Figure Description

[0060] Figure 1 This is a flowchart of the overall process of the multi-angle sonar image target classification method based on a lightweight convolutional network of the present invention.

[0061] Figure 2(a) is a schematic diagram of the bottleneck layer structure with a step size of 1; Figure 2(b) is a schematic diagram of the bottleneck layer structure with a step size of 2.

[0062] Figure 3 It is an improved lightweight convolutional network model diagram, including modules such as input layer, feature extraction convolutional layer, bottleneck layer, and dense connection layer. Detailed Implementation

[0063] This invention proposes a multi-angle sonar image target classification method based on a lightweight convolutional network, the method comprising:

[0064] Step 1) Preprocess the acquired sonar images by fusing three consecutive single-channel sonar images to obtain a multi-angle sonar image dataset with the same size.

[0065] Step 2) Construct a multi-angle sonar image target classification model based on a lightweight convolutional network;

[0066] Step 3) Use the acquired dataset to iteratively train the above multi-angle sonar image target classification model;

[0067] Step 4) Use the image dataset to train and test the target classification model to obtain the target classification results.

[0068] To achieve the above objective, step 1) specifically includes:

[0069] Step 1-1) Obtain N original sonar images containing K types of targets. All are single-channel sonar images. Segment and upsample the images to obtain a set of images of the same size, X1, X2, ..., X... N and the corresponding label vectors Y1, Y2, ..., Y N ;

[0070] Steps 1-2) Divide the sonar images in the single-channel image set into K categories and perform image fusion processing according to the categories. Fuse every three consecutive images in each category into a 3-channel image, i.e., X'1 = X1 + X2 + X3, X'2 = X2 + X3 + X4, and so on.

[0071] After the above processing in steps 1-3), a total image set X'1, X'2, ..., X' is obtained. N-2K and the corresponding label vectors Y1', Y'2, ..., Y' N-2KLet them be denoted as X'1,X'2,...,X' M With Y1',Y'2,...,Y' M ;

[0072] Steps 1-4) Randomly select N1 images and their labels from the above image set to form the training dataset D. train The remaining N2 images and their labels constitute the training dataset D. test Where N1 + N2 = M,

[0073] To achieve the above objective, step 2) specifically includes:

[0074] Step 2-1) The multi-angle sonar image target classification model based on a lightweight convolutional network uses the lightweight MobileNetV2 network as the base network. The entire network model includes two parts: feature extraction and classification. The feature extraction network includes an input layer, a feature extraction convolutional layer Conv1, bottleneck layers Bottleneck1-Bottleneck7, a feature extraction convolutional layer Conv2, and a Flatten layer. After the feature map is converted into one dimension by the Flatten layer, the Dropout layer is used to randomly discard the feature map to improve the fitting effect of the network. The final classification network obtains the final target classification result through a fully connected layer.

[0075] Step 2-2) Bottleneck is the core part of the multi-angle sonar image target classification model based on lightweight convolutional networks. Each Bottleneck unit consists of two standard convolutions and one depthwise separable convolution. In the depthwise separable convolution module, the standard convolution is decomposed into a depthwise convolution and a pointwise convolution. The depthwise convolution convolves each channel of the input feature map with its corresponding convolution kernel, producing the same number of results as the number of input channels, thus filtering the input feature map. The pointwise convolution takes the result of the aforementioned depthwise convolution as input and uses a 1×1 convolution kernel to complete the linear combination of the output results of the depthwise convolution.

[0076] Steps 2-3) The bottleneck layer first performs dimensionality expansion using a 1×1 standard convolution, then extracts features using a 3×3 depthwise separable convolution, and finally compresses the data using a 1×1 standard convolution. The two standard convolutions are activated using ReLU6 and Linear functions, respectively. The depthwise separable convolution is regularized and activated using a normalized BN layer and the linear rectified function ReLU6.

[0077] ReLU6(x) = min(6, max(x, 0))

[0078] Linear(x) = x

[0079]

[0080] in, The output of the convolutional layer, μ B and Let y represent the mean and variance, respectively. i This is the normalized output; here, x i This represents the value of the i-th neuron (node) in the current network layer, where m is the total number of neurons (nodes) in the current network layer. normalized x i The values ​​γ and β are learnable reconfigurable parameters, and ε represents the regularization factor.

[0081] Steps 2-4) The original lightweight MobileNetV2 network uses the ImageNet dataset during training and uses Avgpool downsampling to reduce the number of features. Considering that the number and types of underwater target images are significantly different from the original dataset, in order to obtain a better fitting effect, the number of channels of the feature extraction convolutional layer Conv2 is reduced from 1280 to 128 to reduce computational overhead and memory usage. A Flatten layer is used to perform dimensionality reduction, converting the three-dimensional output of this layer into one dimension. A Dropout layer with a dropout rate of 0.5 is added afterward to randomly drop some units, thereby improving the network fitting. Finally, the target classification result is obtained through a fully connected layer.

[0082] Steps 2-5) In the multi-angle sonar image target classification model based on a lightweight convolutional network, the image size of the input layer is (224,224,3). After passing through the feature extraction convolutional layer Conv1 with a 3×3 kernel, the feature map size is (112,112,32). Then, after passing through 7 bottleneck layers Bottleneck1-Bottleneck7, the feature map size becomes (112,112,16)(56,56,24)(28,28,32)(14,14,64)(14,14,96)(7,7,160)(7,7,320). After passing through the feature extraction convolutional layer Conv2 with a 3×3 kernel, the feature map size is (7,7,128). Then, the feature map is converted into one dimension through the Flatten layer, and the target is classified into K classes using a fully connected layer.

[0083] To achieve the above objective, step 3) specifically includes:

[0084] Step 3-1) When building a multi-angle sonar image target classification network based on a lightweight convolutional network, the parameters of each network layer are distributed using a Gaussian distribution G(μ,σ) with a mean of 0 and a variance of 1.2 Random initialization;

[0085] Step 3-2) Use Optimizers to update the network parameters so that the error between the output image and the label converges to the optimal level;

[0086] Step 3-3) Calculate the model's functional loss value using the categorical cross entropy loss function. The calculation formula is as follows:

[0087]

[0088] Where n represents the number of training set samples, K represents the total number of target categories, and q(x) j,k ) represents sample x j,k The predicted category, p(x) j,k ) represents sample x j,k This represents the probability of the current predicted category.

[0089] Steps 3-4) Use the Softmax classification function to convert the output of the last layer of the network into the probability of the input image belonging to each category, and obtain the final classification result:

[0090]

[0091] Among them, z i Let K be the output value of the i-th node, and K be the number of output nodes, i.e., the number of categories.

[0092] To achieve the above objective, step 4) specifically includes:

[0093] Step 4-1) Conduct comparative experiments using different network models and measure the final classification performance of each model using different evaluation metrics;

[0094] Step 4-2) Comparative experiments were conducted using single-channel sonar images and multi-angle sonar images to verify the classification performance of the proposed method.

[0095] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0096] Example

[0097] like Figure 1 As shown, embodiments of the present invention propose a multi-angle sonar image classification method based on lightweight convolutional networks, the implementation of which includes the following steps:

[0098] Step 1: Preprocess the acquired sonar images by fusing three consecutive single-channel sonar images to obtain a multi-angle sonar image dataset with the same size. The specific implementation is as follows:

[0099] Step 1-1) Obtain N original sonar images containing K types of targets. All are single-channel sonar images. Segment and upsample the images to obtain a set of images of the same size, X1, X2, ..., X... N and the corresponding label vectors Y1, Y2, ..., Y N ;

[0100] Steps 1-2) Divide the sonar images in the single-channel image set into K categories and perform image fusion processing according to the categories. Fuse every three consecutive images in each category into a 3-channel image, i.e., X'1 = X1 + X2 + X3, X'2 = X2 + X3 + X4, and so on.

[0101] After the above processing in steps 1-3), a total image set X'1, X'2, ..., X' is obtained. N-2K and the corresponding label vectors Y1', Y'2, ..., Y' N-2K Let them be denoted as X'1,X'2,...,X' M With Y1',Y'2,...,Y' M ;

[0102] Steps 1-4) Randomly select N1 images and their labels from the above image set to form the training dataset D. train The remaining N2 images and their labels constitute the training dataset D. test Where N1 + N2 = M,

[0103] Step 2: Construct a multi-angle sonar image target classification model based on a lightweight convolutional network. The specific implementation is as follows:

[0104] Step 2-1) The multi-angle sonar image target classification model based on a lightweight convolutional network uses the lightweight MobileNetV2 network as the base network. The entire network model includes two parts: feature extraction and classification. The feature extraction network includes an input layer, a feature extraction convolutional layer Conv1, bottleneck layers Bottleneck1-Bottleneck7, a feature extraction convolutional layer Conv2, and a Flatten layer. After the feature map is converted into one dimension by the Flatten layer, the Dropout layer is used to randomly discard the feature map to improve the fitting effect of the network. The final classification network obtains the final target classification result through a fully connected layer.

[0105] Step 2-2) The bottleneck layer is the core part of the multi-angle sonar image target classification model based on a lightweight convolutional network. Each Bottleneck unit consists of two standard convolutions and one depthwise separable convolution. In the depthwise separable convolution module, the standard convolution is decomposed into a depthwise convolution and a pointwise convolution. The depthwise convolution convolves each channel of the input feature map with its corresponding convolution kernel, producing the same number of results as the number of input channels, thus filtering the input feature map. The pointwise convolution takes the result of the aforementioned depthwise convolution as input and uses a 1×1 convolution kernel to complete the linear combination of the output results of the depthwise convolution. As shown in Figure 2(a) and Figure 2(b), these are schematic diagrams of the bottleneck layer structure with stride size of 1 and stride size of 2, respectively.

[0106] Steps 2-3) The bottleneck layer first performs dimensionality expansion using a 1×1 standard convolution, then extracts features using a 3×3 depthwise separable convolution, and finally compresses the data using a 1×1 standard convolution. The two standard convolutions are activated using ReLU6 and Linear functions, respectively. The depthwise separable convolution is regularized and activated using a normalized BN layer and the linear rectified function ReLU6.

[0107] ReLU6(x) = min(6, max(x, 0))

[0108] Linear(x) = x

[0109]

[0110] in, The output of the convolutional layer, μ B and Let y represent the mean and variance, respectively. i This is the normalized output result;

[0111] Steps 2-4) The original lightweight MobileNetV2 network uses the ImageNet dataset during training, employing Avgpool downsampling to reduce the number of features. Considering the significant difference in the number and types of underwater target images compared to the original dataset, to achieve better fitting results, the number of channels in the feature extraction convolutional layer Conv2 is reduced from 1280 to 128, thus reducing computational overhead and memory usage. A Flatten layer is used for dimensionality reduction, transforming the three-dimensional output of this layer into one dimension. A Dropout layer with a dropout rate of 0.5 is then added to randomly discard some units, thereby improving network fitting. Finally, a fully connected layer is used to obtain the target classification result. Figure 3The diagram shown is of an improved lightweight convolutional network model, which includes modules such as an input layer, a feature extraction convolutional layer, a bottleneck layer, and a densely connected layer.

[0112] Steps 2-5) In the multi-angle sonar image target classification model based on a lightweight convolutional network, the image size of the input layer is (224,224,3). After passing through the feature extraction convolutional layer Conv1 with a 3×3 kernel, the feature map size is (112,112,32). Then, after passing through 7 bottleneck layers Bottleneck1-Bottleneck7, the feature map size becomes (112,112,16)(56,56,24)(28,28,32)(14,14,64)(14,14,96)(7,7,160)(7,7,320). After passing through the feature extraction convolutional layer Conv2 with a 3×3 kernel, the feature map size is (7,7,128). Then, the feature map is converted to one dimension through the Flatten layer, and the target is classified into K classes using a fully connected layer.

[0113] Step 3: Iteratively train the multi-angle sonar image target classification model using the acquired dataset. The specific implementation is as follows:

[0114] Step 3-1) When building a multi-angle sonar image target classification network based on a lightweight convolutional network, the parameters of each network layer are distributed using a Gaussian distribution G(μ,σ) with a mean of 0 and a variance of 1. 2 Random initialization;

[0115] Step 3-2) Use Optimizers to update the network parameters so that the error between the output image and the label converges to the optimal level;

[0116] Step 3-3) Calculate the model's functional loss value using the categorical cross entropy loss function. The calculation formula is as follows:

[0117]

[0118] Where n represents the number of training set samples, K represents the total number of target categories, and q(x) j,k ) represents sample x j,k The predicted category, p(x) j,k ) represents sample x j,k This represents the probability of the current predicted category.

[0119] Steps 3-4) Use the Softmax classification function to convert the output of the last layer of the network into the probability of the input image belonging to each category, and obtain the final classification result:

[0120]

[0121] Among them, z i Let K be the output value of the i-th node, and K be the number of output nodes, i.e., the number of categories.

[0122] Step 4: Train and test the target classification model using the image dataset to obtain the target classification results. The specific implementation is as follows:

[0123] Step 4-1) Conduct comparative experiments using different network models and measure the final classification performance of each model using different evaluation metrics;

[0124] Step 4-2) Comparative experiments were conducted using single-channel sonar images and multi-angle sonar images to verify the classification performance of the proposed method.

[0125] The technical effects of this invention will be further explained below with reference to simulation experiments:

[0126] Conditions and content of the simulation experiment:

[0127] The simulation platform consisted of a Windows system and a Keras environment built on a TensorFlow backend. The experimental computer had an i7-10750H CPU and an RTX 2070 GPU. The simulation dataset used in the experiment was obtained from 3D modeling software and included 1526 images of targets with three shapes: spherical, cylindrical, and frustum. During the simulation, the grazing angle between the sonar and the target ranged from 30° to 45°, and the angle between the target axis and the incident sound wave ranged from 0° to 180°. The simulation dataset was used to assist in training the model.

[0128] To verify the performance of the proposed method, 80% of the image data was randomly selected for training, and the remaining data was used as a test set to verify the network performance. The network cost function was chosen as the classification cross-entropy, calculating the distance between the predicted value and the true label. The Optimizers were used to optimize the parameters of the entire network, with a dropout rate of 0.5, a batch size of 16, a learning rate of 0.0001, and 100 iterations. Ten random experiments were conducted using VGG, MobileNetV2, and the improved network of this invention, and the average values ​​were calculated. Table 1 shows the classification accuracy and parameter count of several network models on the dataset presented in this paper.

[0129] Table 1 Classification performance of different models / %

[0130]

[0131] As shown in Table 1, the model proposed in this invention has the highest classification accuracy. The VGGNet model has the second highest classification accuracy, but its parameter count is too large. Compared with the original MobileNetV2, the model proposed in this invention has advantages in both classification accuracy and parameter count. Using the model proposed in this invention, a comparative experiment was conducted using single-channel image data. The classification results are shown in Table 2. The results show that multi-channel data can make full use of image information and is more conducive to target classification.

[0132] Table 2 Classification performance under different data / %

[0133]

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-angle sonar image target classification method based on a lightweight convolutional network, the method comprising: preprocessing a plurality of angle single-channel sonar images of a target to obtain a multi-angle sonar image of the target through information fusion; inputting the multi-angle sonar image of the target into a multi-angle sonar image target classification model based on a lightweight convolutional network that is preselected, established and trained to obtain a target classification result; wherein the multi-angle sonar image target classification model is obtained by improving a lightweight MobileNetV2 network; the improvement of the lightweight MobileNetV2 network comprises: removing network layers after a feature extraction convolutional layer Conv2 in the lightweight MobileNetV2 network, and modifying the number of channels of the feature extraction convolutional layer Conv2; adding a Flatten layer to realize dimension reduction operation, and adding a Dropout layer for random discarding, and finally adding a fully connected layer to classify the result; the multi-angle sonar image target classification model specifically comprises: a feature extraction module and a classification module; wherein, the feature extraction module comprises an input layer Input, a feature extraction convolutional layer Conv1, bottleneck layers Bottleneck1-Bottleneck7, a feature extraction convolutional layer Conv2, a Flatten layer and a Dropout layer; wherein the number of channels of the feature extraction convolutional layer Conv2 is 128, and the dropout rate of the Dropout layer is 0.5; the classification module comprises a fully connected layer; the processing process of the multi-angle sonar image target classification model specifically comprises: the input layer Input sends the multi-angle sonar image of the target to the feature extraction convolutional layer Conv1; the feature extraction convolutional layer Conv1 with a 3x3 convolutional kernel extracts feature maps from the multi-angle sonar image and inputs them to the bottleneck layers Bottleneck1-Bottleneck7; each Bottleneck in the bottleneck layers Bottleneck1-Bottleneck7 comprises: a first standard convolution, a deep separable convolution and a second standard convolution; first, the 1x1 first standard convolution performs dimension expansion operation, second, the 3x3 deep separable convolution performs feature extraction, and finally, the 1x1 second standard convolution compresses the data; each Bottleneck in the bottleneck layers Bottleneck1-Bottleneck7 processes the feature maps in turn and inputs the processed feature maps to the feature extraction convolutional layer Conv2; the feature extraction convolutional layer Conv2 is set to 128 channels to process the feature maps and input them to the Flatten layer; the Flatten layer converts the feature maps into one dimension and inputs them to the Dropout layer; the Dropout layer randomly discards the feature maps; the fully connected layer classifies the target and outputs the target classification result, i.e., the probability that the target belongs to a certain category; the first standard convolution uses the function ReLU6(x) for activation, and the expression is: ReLU6(x) = min(6, max(x, 0)) where x represents an input value of the activation function, min(·) represents a minimum function, and max(·) represents a maximum function; the second standard convolution is activated using a function Linea(x) and has an expression of: Linear(x) = x.

2. The method according to claim 1, wherein, the deep separable convolution is regularized using a standardization BN layer and has an expression of: where, is the output result of the convolution layer, m represents the total number of nodes of the current network layer, x i represents the value of the i-th node of the current network layer, μ B and respectively represent the mean and variance, y i is the normalized output result, represents the normalized x i value, ε represents a regularization factor, and γ and β are learnable reconstruction parameters. the deep separable convolution is activated using a function ReLU6(x) and has an expression of: ReLU6(x) = min(6, max(x, 0)).

3. The method of claim 1, wherein the method is based on a lightweight convolutional network. The method further comprises training the multi-angle sonar image target classification model, and the training process specifically comprises: Step 1) obtaining a sonar image, pre-processing the obtained sonar image, fusing multiple continuous single-channel sonar images to obtain a multi-angle sonar image dataset with the same size, and taking the multi-angle sonar image dataset as a training dataset and a test dataset; Step 2) constructing a multi-angle sonar image target classification model based on a lightweight convolutional network; Step 3) iteratively training and testing the multi-angle sonar image target classification model using the training dataset and the test dataset respectively to obtain a trained multi-angle sonar image target classification model.

4. The method according to claim 3, wherein, The step 1) specifically comprises: Step 1-1) Obtain N original sonar images containing K-class targets, and each of the N images is a single-channel sonar image; segment and up-sample the images, and obtain a set of images X1, X2,..., X N and corresponding label vectors Y1, Y2,..., Y N ; Step 1-2) divide N sonar images into K categories and perform image fusion processing by category, fuse the images in each category into a P-channel image every continuous P, that is, X'1=X1+X2+...+X P , X'2=X2+X3+...+X P+1 ; Step 1-3) After the processing of step 1-2) is completed, a set of images X1', X'2,..., X'Nis obtained N-2K and the corresponding label vectors Y1', Y'2,..., Y'N N-2K , denoted as X'1, X'2,..., X'N M and Y1', Y'2,..., Y'N M ; Step 1-4) Randomly select N1 images and corresponding label vectors from the image set of step 1-3) to form a training data set D train , and the remaining M-N1 images and corresponding label vectors form a test data set D test .

5. The method according to claim 3, wherein, The step 3) specifically comprises: Step 3-1) For the constructed multi-angle sonar image target classification model based on lightweight convolutional network, the parameters of each network layer are randomly initialized with Gaussian distribution G(μ,σ 2 ) with mean 0 and variance 1. Step 3-2) updating network parameters using an Optimizers optimizer so that an error between an output image and a label converges to an optimal level; Step 3-3) calculating a function loss value Loss of the model by using a cross-entropy loss function; Step 3-4) The output of the last layer of the network is converted into probabilities that the input image belongs to each class using the Softmax classification function Softmax(z i ), resulting in the final classification result.

6. The method according to claim 5, wherein, In the step 3-3), a calculation formula of the Loss is: wherein n represents the number of training set samples, K represents the total number of target categories, q(x j,k ) represents the predicted category of the sample x j,k , and p(x j,k ) represents the probability of the sample x j,k being the current predicted category.

7. The method according to claim 5, wherein, The calculation formula of Softmax(z i ) in the step 3-4) is: where z i is the output value of the last network layer for the i-th node.

Images