Underwater robot sonar data preprocessing method and device based on convolution auto-encoder

Abstract

The invention relates to the technical field of underwater robot sonar data processing, in particular to an underwater robot sonar data processing method and device based on a convolutional auto-encoder, and the method comprises the following steps: S1, building a database, S2, carrying out standardization processing, S3, constructing a convolutional auto-encoder module, S4, training a preprocessing network, and S5, extracting features and reducing noise. According to the method, the convolutional auto-encoder model is adopted, the internal structure of the data can be automatically learned from the original data, good denoising and feature extraction capabilities are achieved, underwater noise, echo and other irrelevant information can be effectively removed, and therefore the definition and precision of the data are improved; space and depth information can be effectively enhanced, and obstacles and dangerous areas in the underwater environment can be identified more accurately; the method provided by the invention is an unsupervised learning method, can adapt to automatic optimization and processing flows of different underwater environments, and further enhances the adaptability, robustness and generalization ability of the underwater robot.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater robot sonar data processing, and in particular to an underwater robot sonar data preprocessing method and device based on a convolutional autoencoder. Background Art

[0002] Underwater robots, also known as unmanned remote-controlled submersibles, are robots that can perform various tasks underwater. They are widely used in many fields such as deep-sea exploration, marine scientific research, and underwater operations. Because the underwater environment is usually harsh and dangerous, and the diving depth of humans is limited, underwater robots have become an important tool for the development of marine resources. In underwater environments, robots often face complex obstacles and environmental interference when performing tasks, especially in turbid water bodies, where the effects of visual sensors and other sensors are seriously affected. In this environment, sonar, as an important underwater detection technology, has strong penetrating power and can effectively detect obstacles and the surrounding environment, and has gradually become one of the first choices for underwater robot detection equipment.

[0003] However, the quality of sonar data is often affected by underwater noise, echoes, stray signals and other factors, resulting in noise, distortion and low resolution in the data. These problems make it difficult for traditional sonar-based data processing methods to adapt to the needs of high-precision obstacle recognition and path planning in complex environments; in order to improve the quality and reliability of sonar data, underwater robots often need to rely on efficient preprocessing methods; traditional sonar data preprocessing methods, such as filtering, denoising and image enhancement, can improve data quality to a certain extent, but these methods usually rely on manually set rules or simple mathematical models, and are difficult to adapt to complex and changing underwater environments. Summary of the invention

[0004] The object of the present invention is to provide a method and device for preprocessing underwater robot sonar data based on a convolutional autoencoder, which can improve data quality and accuracy.

[0005] Based on the above purpose, the present invention adopts the following technical solution:

[0006] A method for preprocessing underwater robot sonar data based on convolutional autoencoder, comprising the following steps:

[0007] S1. Building a database: connecting with the underwater robot sonar, obtaining sonar-related data, and building a database of related sonar data;

[0008] S2, standardization processing: standardizing the underwater robot sonar data obtained in step S1 to obtain standardized sonar data;

[0009] S3. Construct a convolutional autoencoder CAE module: The CAE module includes an encoder, a decoder, and an output layer;

[0010] S4, training the preprocessing network: inputting the standardized sonar data obtained in step S2 as the original image into the CAE module obtained in step S3 for layer-by-layer unsupervised pre-training, and obtaining the sonar data preprocessing network after the training is completed;

[0011] S5, extracting features and reducing noise; using the sonar data preprocessing network in step S4 to extract deep features of the original image and perform noise reduction;

[0012] S6. Output data: output the data after noise reduction in step S5 through the output module.

[0013] Preferably, the specific process of building a database in step S1 includes:

[0014] Based on the data obtained in step S1, the original data is labeled and sorted, and then a sonar database is built.

[0015] Preferably, the specific process of the standardization processing in step S2 includes:

[0016] The acquired underwater robot sonar data is read and converted into a grayscale image; the stored image size is 128×128 pixels; the grayscale image is normalized: it is normalized to the range of [0, 1], and the normalization formula is as follows:

[0017]

[0018] In the formula, x represents the original data, min(x) and max(x) represent the minimum and maximum values ​​in the data set respectively, and x norm Represents normalized data.

[0019] Preferably, the specific process of constructing the CAE module in step S3 includes:

[0020] The CAE module includes an Encoder and a Decoder. The Encoder extracts low-level features of the image through two layers of convolution operations, and the Decoder restores the image to its original size through two layers of transposed convolution operations to reconstruct the denoised image.

[0021] Preferably, the encoder includes two 2D convolutional layers, and the decoder includes two 2D transposed convolutional layers; the CAE module includes an output layer, and the output layer includes a 2D transposed convolutional layer.

[0022] Preferably, the layers of the CAE module are connected by a RELU activation function, so that the i-th layer obtains the feature maps of all previous layers and uses them as input. The formula is as follows:

[0023] x i =H i ([x 0 , x 1 , …, x i-1 ])

[0024] In the formula, [x 0 , x 1 , …, x i-1 ] is the concatenation of the feature maps of layers 0, 1, 2, ..., i-1.

[0025] Preferably, the specific process of training the network in step S4 includes:

[0026] The original image is used as input, and the pre-training result is used as the target output. Through multiple rounds of training, the model learns how to effectively remove noise from the image.

[0027] The CAE module is used to perform layer-by-layer unsupervised pre-training. After the training is completed, the sonar data pre-processing network is obtained;

[0028] The mean square error (MSE) is used as the loss function to optimize the denoising performance of the model. The loss function expression is as follows:

[0029]

[0030] In the formula, n represents the number of samples, x i represents the true value of the i-th sample, y i Represents the model's predicted value for the i-th sample.

[0031] Preferably, the sonar data preprocessing network uses a loss function formula combining L1 loss and MSE as a loss function to optimize the denoising performance of the model; the loss function expression is as follows:

[0032]

[0033] in,

[0034]

[0035] In the formula, n represents the number of samples, x i represents the true value of the i-th sample, y i represents the predicted value of the model for the i-th sample, α and β are weight parameters that control L1 and MSE losses respectively.

[0036] Preferably, the specific process of extracting features and reducing noise in step S5 includes:

[0037] The sonar data preprocessing network obtained in step S4 is used to extract the deep features of the original image, and noise reduction is performed through signal reconstruction and residual link to obtain a denoised image;

[0038] The signal is disassembled and reconstructed through CAE, and the noise part removed by the model is represented by the difference between the original image and the denoised image; the denoised signal is obtained by calculating the residual signal. The signal is analyzed by combining MSE and peak signal-to-noise ratio (PSNR) to determine whether the model has accurately removed the noise and to judge the preprocessing effect of the reconstructed signal.

[0039] A sonar data preprocessing device for an underwater robot comprises a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, any step in the above-mentioned sonar data preprocessing method for an underwater robot based on a convolutional autoencoder is implemented.

[0040] The beneficial effects of the present invention include:

[0041] The present invention adopts a convolutional autoencoder (CAE) model, which, as an unsupervised learning model, can automatically learn the intrinsic structure of data from raw data and has good denoising and feature extraction capabilities. CAE compresses sonar data into a low-dimensional space through the structure of encoder and encoder, and can extract the most important features from it. In an underwater environment, this process can effectively remove underwater noise, echoes and other irrelevant information, thereby improving the clarity and accuracy of the data.

[0042] The CAE-based sonar data preprocessing method adopted in the present invention can not only improve the quality of sonar images, but also effectively enhance spatial and depth information, so that the underwater robot can more accurately identify obstacles and dangerous areas in the underwater environment, and provide more reliable input data for subsequent tasks such as obstacle detection, path planning and collision avoidance control, which significantly improves the autonomous navigation capability and task execution efficiency of the underwater robot.

[0043] Compared with traditional preprocessing methods, the CAE model adopted in the present invention can automatically capture the complex characteristics of the underwater environment and improve the representation ability of sonar data by learning a large amount of data through unsupervised learning methods; this enables the sonar data preprocessing method to adapt to different underwater environments and automatically optimize the preprocessing process to cope with the processing of diverse background noises without relying on manually set rules, further enhancing the adaptability, robustness and generalization ability of the underwater robot. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of the present invention;

[0045] Figure 2 Schematic diagram of the convolutional autoencoder neural network structure of the present invention;

[0046] Figure 3 A grayscale image converted from the sonar data of the present invention;

[0047] Figure 4 It is a sonar grayscale image processed by the convolutional autoencoder model of the present invention;

[0048] Figure 5 This is the sonar grayscale image after feature extraction and noise reduction in the present invention. DETAILED DESCRIPTION

[0049] The following is a further explanation of the present invention in conjunction with specific embodiments. Figure 1 As shown, the present invention is a method for preprocessing underwater robot sonar data based on convolutional autoencoder, comprising the following steps:

[0050] S1. Build a database: obtain data related to the underwater robot sonar, annotate and organize the original data, and build a database of related sonar data.

[0051] S2, standardization processing: standardize the underwater robot sonar data obtained in step S1 to obtain standardized sonar data:

[0052] The acquired underwater robot sonar data is read and converted into a grayscale image; the stored image size is 128×128 pixels, such as Figure 3 As shown; the grayscale image is normalized: it is normalized to the range of [0, 1], and the normalization formula is as follows:

[0053]

[0054] In the formula, x represents the original data, min(x) and max(x) represent the minimum and maximum values ​​in the data set respectively, and x norm Represents normalized data.

[0055] S3. Build the convolutional autoencoder (CAE) module:

[0056] The CAE module includes: encoder, decoder and output layer; the encoder consists of two 2D convolution layers, each convolution kernel size is 3×3, and the low-level features of the image are extracted through two layers of convolution operation; the encoder consists of two 2D transposed convolution layers, each convolution kernel size is 3×3, and the image is restored to its original size through two layers of transposed convolution operation to reconstruct the denoised image; finally, the image is output by an output layer, which consists of one 2D transposed convolution layer, and the convolution kernel size is 3×3, as shown in Figure 2 As shown; the sonar grayscale image after CAE model processing is as follows Figure 4 shown.

[0057] The layers of the network are connected through RELU to obtain the feature maps x of all layers before the i-th layer. 0 , x 1 , …, x i-1 , taking it as input, expressed as:

[0058] x i =H i ([x 0 , x 1 , …, x i-1 ])

[0059] In the formula, [x 0 , x 1 , …, x i-1 ] is the concatenation of the feature maps of layers 0, 1, 2, …, i-1.

[0060] Table 1. Neural network structure based on convolutional autoencoder

[0061]

[0062] As shown in Table 1, the output of each layer in the model is regarded as the input of the following layer:

[0063] (1) After the data is input, first 0 ) performs a 3×3 convolution to adjust the number of channels from 1 to 64, and then uses 3x3 convolution to obtain a 32-channel feature layer. At this time, a feature x with a shape of (128, 128, 32) is obtained. 1 .

[0064] (2) x 1 As the input of the 2D transposed convolutional layer, first x 1Perform a transposed convolution with a convolution kernel of 3×3, and keep the number of channels unchanged, that is, shape (128, 128, 32). Then use the second convolution kernel of 3×3 to obtain a 64-channel feature with a shape of (128, 128, 64). The output of the 64-channel is used as input to the output layer and perform the third transposed convolution operation to obtain an output x with a shape of (128, 128, 1). 2 .

[0065] (3) Transpose the convolution to obtain x 2 and the initial input x 0 Perform residual link and convert x 2 and x 0 Add, and then perform grayscale image permutation and pixel-by-pixel addition operations in sequence to obtain a new feature y; feature y is the predicted value of the sample, which also retains the initial input x 0 And after convolution processing x 2 characteristics.

[0066] S4, training preprocessing network: the standardized sonar data obtained in step S2 is used as the original image input into the CAE module obtained in step S3 for layer-by-layer unsupervised pre-training. After the training is completed, the sonar data preprocessing network is obtained:

[0067] The original image is used as input, and the pre-training result is used as the target output. Through multiple rounds of training, the model learns how to effectively remove noise from the image.

[0068] The CAE module is used to perform layer-by-layer unsupervised pre-training. After the training is completed, the sonar data pre-processing network is obtained;

[0069] At the same time, the mean square error is used as the loss function to optimize the denoising performance of the model; the loss function expression is as follows:

[0070]

[0071] In the formula, n represents the number of samples, x i represents the true value of the i-th sample, y i Represents the model's predicted value for the i-th sample.

[0072] The sonar data preprocessing network uses the loss function formula that combines L1 loss and MSE as the loss function to optimize the denoising performance of the model; the loss function expression is as follows:

[0073]

[0074] in,

[0075]

[0076] In the formula, n represents the number of samples, x i represents the true value of the i-th sample, y i represents the predicted value of the model for the i-th sample, α and β are weight parameters that control L1 and MSE losses respectively.

[0077] S5, extracting features and reducing noise; using the sonar data preprocessing network in step S4 to extract deep features and perform noise reduction;

[0078] The sonar data preprocessing network obtained in step S4 is used to extract the deep features of the original image, and the denoising process is performed through signal reconstruction and residual linking to obtain a denoised image; the sonar grayscale image obtained by constructing a new neural network through residual linking and reorganization method is as follows: Figure 5 As shown, the grayscale image is Figure 3 , Figure 4 In comparison, after denoising, the target area becomes clearer and the noise is significantly reduced. Then CAE can be used to disassemble and reconstruct the signal, and the difference between the original image and the denoised image represents the noise removed by the model; by calculating the residual signal, the denoised signal is obtained. By combining the MSE and PSNR analysis model, it can be determined whether the signal has accurately removed the noise, thereby judging the effect of the reconstructed signal preprocessing, and it can be used as a loss function to further optimize the model.

[0079] The difference between the original image and the denoised image represents the noise removed by the model; by calculating the residual signal, it is possible to analyze whether the model has accurately removed the noise and determine the effect of the reconstructed signal preprocessing.

[0080] S6. Output data: output the data after noise reduction in step S5 through the output module.

[0081] The above description is only a further explanation of the present invention in combination with specific embodiments. All descriptions do not limit the scope of protection of the present invention. Any changes or replacement schemes that can be easily thought of by any technician in this field within the technical scope disclosed by the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for preprocessing underwater robot sonar data based on convolutional autoencoder, characterized in that: The following steps are involved: S1. Build a database: connect with the underwater robot sonar, obtain sonar-related data, annotate and organize the original data, and build a database of related sonar data; S2, standardization processing: standardizing the underwater robot sonar data obtained in step S1 to obtain standardized sonar data; S3. Construct a convolutional autoencoder CAE module: the CAE module includes an encoder, a decoder and an output layer; S4, training the preprocessing network: inputting the standardized sonar data obtained in step S2 as the original image into the CAE module obtained in step S3 for layer-by-layer unsupervised pre-training, and obtaining the sonar data preprocessing network after the training is completed; S5, extract features and reduce noise; Using the sonar data preprocessing network in step S4 to extract deep features of the original image and perform noise reduction processing; S6. Output data: output the data after noise reduction in step S5 through the output module.

2. The underwater robot sonar data preprocessing method based on convolutional autoencoder according to claim 1 is characterized in that: The specific process of the standardization process in step S2 includes: The acquired underwater robot sonar data is read and converted into a grayscale image; the grayscale image is normalized: it is normalized to the range of [0,1], and the normalization formula is as follows: In the formula, x represents the original data, min(x) and max(x) represent the minimum and maximum values ​​in the data set respectively, and x norm Represents normalized data.

3. The underwater robot sonar data preprocessing method based on convolutional autoencoder according to claim 1 is characterized in that: The specific process of constructing the CAE module in step S3 includes: The Encoder extracts low-level features of the image through two layers of convolution operations, and the Decoder restores the image to its original size through two layers of transposed convolution operations to reconstruct the denoised image.

4. The underwater robot sonar data preprocessing method based on convolutional autoencoder according to claim 3 is characterized in that: The Encoder includes two 2D convolutional layers, the Decoder includes two 2D transposed convolutional layers; and the output layer includes a 2D transposed convolutional layer.

5. The underwater robot sonar data preprocessing method based on convolutional autoencoder according to claim 4 is characterized in that: The layers of the CAE module are connected by the RELU activation function, so that the i-th layer obtains the feature maps of all previous layers and uses them as input. The formula is as follows: x i =H i ([x0,x1,…,x i-1 ]) In the formula, [x0,x1,…,x i-1 ] is the concatenation of the feature maps of layers 0, 1, 2, …, i-1.

6. The underwater robot sonar data preprocessing method based on convolutional autoencoder according to claim 1 is characterized in that: The specific process of the network training in step S4 includes: The original image is used as input, and the pre-training result is used as the target output. Through multiple rounds of training, the model learns how to effectively remove noise from the image. The CAE module is used to perform layer-by-layer unsupervised pre-training. After the training is completed, the sonar data pre-processing network is obtained; The mean square error (MSE) is used as the loss function to optimize the denoising performance of the model. The loss function expression is as follows: In the formula, n represents the number of samples, x i represents the true value of the i-th sample, y i Represents the model's predicted value for the i-th sample.

7. The underwater robot sonar data preprocessing method based on convolutional autoencoder according to claim 6 is characterized in that: The sonar data preprocessing network uses a loss function formula combining L1 loss and mean square error as a loss function to optimize the denoising performance of the model; the loss function expression is as follows: in, In the formula, n represents the number of samples, x i represents the true value of the i-th sample, y i represents the predicted value of the model for the i-th sample, α and β are weight parameters that control L1 and MSE losses respectively.

8. The underwater robot sonar data preprocessing method based on convolutional autoencoder according to claim 7 is characterized in that: The specific process of extracting features and reducing noise in step S5 includes: The sonar data preprocessing network obtained in step S4 is used to extract the deep features of the original image, and noise reduction is performed through signal reconstruction and residual link to obtain a denoised image; The signal is disassembled and reconstructed through CAE, and the noise part removed by the model is represented by the difference between the original image and the denoised image; the denoised signal is obtained by calculating the residual signal. The signal is analyzed by combining the mean square error and peak signal-to-noise ratio to determine whether the model has accurately removed the noise and to judge the preprocessing effect of the reconstructed signal.

9. A sonar data preprocessing device for an underwater robot, comprising a memory and a processor, wherein a computer program is stored in the memory, and wherein: When the processor executes the computer program, it implements any step in the underwater robot sonar data preprocessing method based on convolutional autoencoder as described in any one of claims 1-9.

Images