CN113516626B

CN113516626B - Side-scan sonar submarine sand wave detection method based on multi-scale convolution and pooling strategy

Info

Publication number: CN113516626B
Application number: CN202110497412.0A
Authority: CN
Inventors: 年睿; 臧丽娜; 史可心; 何波; 于菲
Original assignee: Ocean University of China
Current assignee: Ocean University of China
Priority date: 2021-05-07
Filing date: 2021-05-07
Publication date: 2023-09-26
Anticipated expiration: 2041-05-07
Also published as: CN113516626A

Abstract

The invention provides a side-scan sonar submarine sand wave detection method based on a multi-scale convolution and pooling strategy. Deep semantic information of a submarine sand wave image is extracted based on a cumulative learning bilateral branch network, a pyramid convolution module and strip pooling are introduced according to the characteristics of the submarine sand wave, self-adaptive optimization and global information enhancement of characteristics are carried out, the problems that the number of submarine sand wave samples are unevenly distributed, morphological characteristics of the sand wave are various in a complex submarine environment, and marine reverberation noise interference is serious are effectively solved, the submarine sand wave is positioned by taking echo intensity as a research object through waveform matching, and morphological characteristics such as wavelength, wave height and asymmetry index of the submarine sand wave and migration trend of the submarine sand wave are analyzed, so that a foundation is laid for further investigation and analysis of the submarine sand wave.

Description

Side-scan sonar submarine sand wave detection method based on multi-scale convolution and pooling strategy

Technical Field

The invention particularly relates to a side-scan sonar submarine sand wave detection method based on a multi-scale convolution and pooling strategy, and belongs to the technical field of ocean intelligent information processing and underwater target detection and identification.

Background

Ocean exploration is mainly used for detecting various geophysical field characteristics, geological structures and mineral resources of the ocean floor and exploring the distribution characteristics of submarine placers and sand waves. Among these, seafloor sand waves are a topographical morphology that is commonly distributed throughout large sea areas, formed by a variety of environmental factors, such as sediment particle size, tidal characteristics, and associated residual currents. Submarine sand wave identification is an important mode for investigating submarine sediments, and is an important research content in the fields of marine sediment sounding, marine topography, marine geophysics, marine geological investigation and the like. The method for identifying the submarine sand waves generally comprises a numerical calculation method, a model calculation method, remote sensing detection and a field investigation method. Acoustic field investigation methods using side-scan sonar based on autonomous underwater vehicles (Autonomous Underwater Vehicle, AUV) or cabled underwater vehicles (Remotely Operated Vehicle, ROV) are currently widely used in the study of submarine sand waves.

The side-scan sonar is usually installed on two sides of the carrying device, the transducer array emits sound waves according to a certain frequency, the sound waves return when encountering sand waves and other submarine sediments, the receiving array of the side-scan sonar can receive sonar signals reflected, diffracted or scattered from the bottom frequency by frequency and splice echo intensities into a side-scan sonar image, different information of submarine characteristics is provided according to different frequencies, the submarine complex topography and topography are reflected, and the large-scale submarine sand wave investigation can be realized. The method for carrying out submarine sand wave investigation through the side-scan sonar is widely applied, provides a research foundation for further submarine sand wave simulation and predictive analysis, and currently faces a lot of challenges.

First, the quality of the acquired side-scan sonar image is affected due to the influence of the ocean reverberation noise and the noise of the detection equipment, the energy loss of the sound wave in the sea water, the limited single-pass information quantity of the sound wave, and the like, and the edge of the sand wave is unclear. The problems are that the traditional image recognition method is seriously hindered when the sand wave target features are manually extracted, all features are difficult to consider, the form of the submarine sand wave is fitted, and the processing speed and accuracy of sand wave recognition are affected.

Secondly, in the data acquired by the side-scan sonar, the number of images of a background area is far greater than that of images of sand waves, sample imbalance is presented, so that the deep learning network is enabled to pay more attention to the improvement of the accuracy of the samples of the background area, excessive negative sample gradients are applied to the images of the sand waves by the images of the background area, positive sample gradients of the sand waves are submerged, and therefore overfitting is caused, and the recognition accuracy of the sand wave data is seriously affected.

And thirdly, the tidal flow, the residual flow, the submarine substrate type and other hydrologic environments in different sea areas are different to cause the diversity of submarine topography characteristics, the geometrical characteristics of sand waves, the categories of the sand waves and the submarine environments are in multi-layer and complex nonlinear relations, and the shapes of the sand waves under different submarine topography are greatly different, so that the sand waves are required to be identified according to the surrounding environments and comprehensively judged, and the sand waves are required to have larger receptive field, stronger generalization performance and reasoning capacity.

Finally, regardless of the linear sand wave, the crescent sand wave has long-strip shape commonality, the waves are distributed in a discrete mode at intervals, the regular square pooling core limits the context flexibility of the sand wave feature capture and is unfavorable for effectively extracting the sand wave feature, and the traditional square pooling core is likely to introduce unnecessary connection such as noise and the like to interfere with the recognition of the submarine sand wave.

Disclosure of Invention

Aiming at the problems, the invention aims to provide a side-scan sonar submarine sand wave detection method based on a multi-scale convolution and pooling strategy, which comprises a side-scan sonar submarine sand wave identification method and a morphological analysis method.

In order to achieve the above purpose, the invention adopts the following specific technical scheme:

a side-scan sonar submarine sand wave detection method based on a multi-scale convolution and pooling strategy comprises the following steps:

s1: collecting submarine side-scan sonar data and attitude positioning data;

s2: preprocessing the submarine side-scan sonar data collected in the step S1;

s3: constructing a submarine sand wave identification network based on an improved convolutional neural network;

s4: training the submarine sand wave identification network obtained in the step S3;

s5: and (3) identifying the side-scan sonar testing data after the pretreatment of the S2 by using the trained submarine sand wave identification network of the S4, and outputting an identification result.

Further, in the step S1,

(1) Collecting data by using a side scan sonar submarine detection means: including underwater vehicle longitude lon, latitude lat, speedHeading h, pitch angle p, roll angle r, and raw echo intensity data X _i I=1..h; the original echo intensity data X is preprocessed and mapped into a side scan sonar image, and is defined as I, wherein the size is H multiplied by W, H is the height, and W is the width;

(2) High H by side-scan sonar detection _a Sonic pitch R _s Single-side detection range R _h The method comprises the steps of carrying out a first treatment on the surface of the Labeling the side-scan sonar image I, dividing the side-scan sonar image I into two types of sand waves and no sand waves, wherein the corresponding label is y, and each s is arranged according to the longitudinal interval _v Ping and lateral spacing per s _h Overlapping and blocking the echo intensity data points to form an image block P with the size of l _ij ,i＝1,...,(H-l)/s _v +1,j＝1,...,(W-l)/s _h +1 and respectively occupy the proportion lambda according to the training set, the test set and the verification set ₁ 、λ ₂ 、λ ₃ Distribution of sample data (lambda) ₁ +λ ₂ +λ ₃ ＝1)；

Further, in the step S2: the pretreatment comprises the following steps:

(1) And respectively carrying out blind area removal, noise elimination, intensity correction, speed correction and time-varying gain correction data preprocessing on the side-scan sonar image I.

(2) Adaptive filtering of echo intensities using ensemble empirical mode decomposition (Ensemble Empirical ModeDecomposition, EEMD) to obtain echo intensity data

(3) Calculating a correction proportion coefficient through speed correction to obtain echo intensity dataSetting the range distance delta h of the side-scan sonar image corresponding to the ith ping data _i Expressed as:

wherein i is E [1, H]，Δt _i The time difference between the reception of the ipping data and the reception of the (i + 1) th ping data for the sonar sensor,is the instantaneous speed of the AUV at the ith ping,/->Is the instantaneous voyage speed of AUV in the ith+1th ping, W is the width of echo intensity in the ith ping, +.>Is the AUV actual sailing distance, W/2R _s Representing the mapping coefficient between the width of the side-scan sonar image and the actual detection range of the seabed perpendicular to the navigation track direction, deltat _i Is the time difference between the i-th ping and the (i+1) -th ping of the echo received by the side scan sonar.

(4) And (3) compensating partial energy loss caused by the sound wave emitted by the side-scan sonar when the sound wave propagates in the sea water by a time-varying gain correction method:

wherein ,x_ij Is the original echo intensity X _i Is set for the j-th echo intensity value of (c),is the echo intensity value after filtering and velocity correction,/->Is a time-varying gainEcho intensity values after correction. A is an echo intensity correction coefficient, and B is an echo intensity correction bias.

Still further, the step (1) includes:

1) Adding white noise with a given amplitude to the echo intensity;

2) Performing an empirical mode decomposition (Empirical Mode Decomposition, EMD) on the signal with the echo intensities and the added white noise to obtain an intrinsic mode function (Instrinic Mode Function, IMF) component and a residual component;

3) This operation was repeated using a given number of trials; the number of IMFs was constant in each trial; all of the experiments can be written as such,

wherein ,X_i Is the i-th original signal in echo intensity; n is n _s Is white noise added in the s-th test; IMF (inertial measurement unit) _sp Is the p-th IMF component in the s-th trial; r is (r) _s Is the remaining component in the s-th trial;

4) Calculating the overall average value of all experiments, and decomposing the original signal X in the calculated echo intensity of the sonar by using the integrated empirical mode _i The final decomposition result of (2) is expressed as:

where S is the number of trials and r is the residual component of EEMD;

5) Select the first lot p ₁ IMF and denoising EEMD through a general threshold formula as follows:

wherein ,σ_p Is the standard deviation of noise, N is IMF _p Length of (2)A degree;

in determining the general threshold T _p Afterwards, we select a soft threshold function to filter noise component coefficients in the high frequency IMF;

6) Finally, the denoising method of EEMD comprises the following steps:

noise eliminating the echo intensity data to obtain The target features are further highlighted using histogram equalization as intensity correction.

Further, in the step S3, a submarine sand wave identification network based on an improved convolutional neural network is constructed, specifically:

(1) A cumulative-learning Bilateral Network (BBN) is used as a backbone Network for submarine sand wave identification, including convolutional learning branches and rebalancing branches. For the problem of training sample imbalance, the convolution learning branches are used for sampling training samples (u _c ,y _c ) Training samples (u) are obtained using a reverse sampling mechanism _r ,y _r ) As an input to the rebalancing branch, focus learning attention on the sand wave side-scan image data with a small number of training samples;

(2) Based on a cumulative learning bilateral branch network BBN as a main body, a pyramid convolution module, a strip pooling module and a teeterboard loss module are introduced to jointly construct a submarine sand wave identification network.

Further, in the step (1) of S3

1) Obtaining samples (u) using a reverse sampling mechanism _r ,y _r ) As an input of the rebalancing branch, the sampling probability P of the class m is calculated by an inverse sampler _m ：

wherein ,N_m 、N _j The number of samples of class m and class j, N _max Is the total number of samples and C is the total number of categories.

2) Through an accumulated learning strategy, the adjustment parameter alpha controls the weight and the classification loss occupied by the two branch characteristics, transfers the learning key points of bilateral branches, and focuses the sand wave side scanning image data with a small number of training samples.

Firstly, learning a general mode, and then gradually focusing on tail data; after convolution learning branches and rebalancing branches, the feature vectors f are output through corresponding full-connection layers respectively _c and f_r Then, the accumulated learning is input into the accumulated learning; if the total number of training iterations is T _max The current training iteration number is T, and the adjustment parameter alpha is:

alpha controls the two branch output eigenvectors f _c and f_r Is weighted, and the feature vector αf after weighting is weighted _c And (1-. Alpha.) f _r Classifier sent to corresponding branchAnd characterizer->In the method, output fusion is carried out in an element addition mode, and the fused C-dimensional output prediction characteristic z is expressed as follows:

the BBN loss function in the whole learning process is a seesaw loss function:

L＝αL _s (z _c )+(1-α)L _s (z _r ) (12)

wherein ,z_c ,z _r ＝{z ₁ ,z ₂ ,...,z _C "is category prediction, L _s (. Cndot.) represents the see-saw loss function.

Furthermore, the pyramid convolution module is introduced, which is specifically as follows:

(1) The pyramid convolution module is introduced into the cumulative learning bilateral branch network BBN. The convolution learning branch and the rebalancing branch have the same structure, and all parameters except the last residual block are shared. Analyzing input by utilizing a Local multiscale context aggregation module (Local multi-scale context aggregation module) and a Global multiscale context aggregation module (Global multi-scale context aggregation module) to enrich the feature capturing function of the accumulated learning bilateral branch network;

(2) For each branch network, firstly extracting a feature map through a convolution layer, then adding a network receptive field through Residual Block (Residual Block), then through cavity convolution (di conv) and pyramid convolution (PyConv), fusing the captured features of multiple convolution kernel scales through pyramid convolution, enabling the depth of a convolution kernel to change along each level to form full depth and connectivity, and obtaining the feature map of the initial dimension through improved Residual Block and convolution.

(3) Repeating the steps, and respectively obtaining a characteristic diagram by the convolution learning branch and the rebalancing branch through the multistage improved residual block. In the hole convolution process, the convolution scale is adjusted by adding the super-parameter hole rate on the basis of standard convolution, and the equivalent convolution kernel size after the hole convolution is introduced is k' =k+ (k-1) x (d-1), wherein k is the size of the original convolution kernel, and d is the super-parameter hole rate.

The strip-shaped pooling module is introduced and comprises the following concrete steps:

(1) Introducing a strip pooling module to the cumulative learning bilateral branch network BBN, and utilizing a narrow pooling core to better fit and identify the sand wave form; let the dimension be f _h ×f _w Is input feature map of (a)c represents the number of channels, and the channels are respectively fed into two parallel paths of horizontal pooling and vertical pooling in the process of strip pooling;

(2) An improved strip pooling strategy is established, and effective feature expressions are extracted from a row-by-row and column-by-column angle in parallel through equal weight combination of average pooling and maximum pooling:

the horizontal pooling output dimension is (f) _h 1, c)

The vertical pooling output dimension is (1, f _w C)

The two output features of the strip pool are expanded through convolution operation and respectively transformed into the size f _H ×f _W Horizontal bar expansion of (2)And vertical bar expansion->By y=y ^H +y ^V Will y ^H and y^V Combining; by z=scale (P _ij Sigma (f (y)) computation with global a priori feature output to create feature captures appropriate for the long strip morphology of the sand wave, assisting in subsea sand wave feature optimization and parameter learning, where Scale (·, ·) is element-by-element multiplicative addition, sigma is the activation function, and f is a 1 x 1 convolution.

The introduced seesaw loss module is specifically as follows:

(1) Introducing seesaw loss to a cumulative learning bilateral branch network BBN; introducing a distribution function mu (k) independent of the input samples and a smoothing parameter eta, optimizing to y' = (1-eta) y+eta mu (k) by sample Label regularization (Label SmoothingRegularization, LSR); according to the seesaw loss, dynamically rebalancing positive and negative sample gradients of each category through the proportion of the number of samples in the training process:

wherein ,y_m ' is to the artificial tag y _m Class labels after LSR are passed, so that stronger fault tolerance is achieved on sample marking errors; z _m ＝{z ₁ ,z ₂ ,...,z _C -category prediction;

for samples of class m, the negative sample gradient applied to class n is:

wherein ,S_mn As an adjustable adjustment factor between samples of different types, adjusting penalty imposed by the mth type on the nth type; by a relieving factor M _mn And a compensation factor C _mn Determining S _mn ＝M _mn C _mn ；

When category m occurs more frequently than category n, the mitigation factor is:

according to the sample number proportion N between the class N and the class m _n /N _m Reducing penalties on tail class n; the index p is a super-parameter of the adaptive amplitude;

if the prediction probability of the class n is greater than the class m, the compensation factor is:

penalty increase for class n (σ _n /σ _m ) ^q A multiple, wherein q is a hyper-parameter controlling the ratio; c (C) _mn When=1, C _mn By only applying the relieving factor M _mn The method comprises the steps of carrying out a first treatment on the surface of the Thereby avoiding misclassification of category n due to the overwhelming penalty mitigation for category n.

The invention introduces improved process analysis of pyramid convolution module, bar pooling module and teeterboard loss module:

(1) Introducing a pyramid convolution module:

the feature extraction aiming at the convolutional neural network can cause discontinuous phenomenon during the target feature extraction because the global information of the image is not effectively utilized due to the limitation of the receptive field. The seabed sand wave is a product of the combined action of the seabed and tide, the seabed slope factor, the sediment particle size and the material type can all influence the seabed sand wave, in a complex scene such as seabed sand wave identification, the receptive field is insufficient to capture the correlation of different positions in the scene, useful details can be lost, and the sand wave identification can be misjudged. The invention introduces a pyramid convolution module to the accumulated learning bilateral branch network, and utilizes a local multi-scale context aggregation module and a global multi-scale context aggregation module to effectively analyze input so as to obtain a feature capturing function of the extensive global information enrichment accumulated learning bilateral branch network; the pyramid convolution module is further optimized, the receptive field of the network is increased through combination with the cavity convolution, and the characteristics are enhanced through the optimization module combining the cavity convolution and the space pyramid convolution, so that the sand wave characteristic distinguishing and extracting capability is improved, and the global priori information of the submarine sand waves is effectively obtained.

(2) Introducing a strip pooling module:

the use of a conventional large square pooling kernel window for a subsea sand wave in the form of a long strip will inevitably contain contamination information from unrelated areas. The invention introduces the strip pooling module to the accumulative learning bilateral branch convolution network, and utilizes the narrow pooling core to better fit and identify the form of the sand wave, thereby avoiding the influence of the irrelevant area on the submarine sand wave identification and avoiding the influence of noise on the submarine sand wave identification to a certain extent. And the longer pooling core is convenient for capturing the relation among the sand wave stripes of the isolated area, so that the probability that the sand wave is misjudged as the background is reduced. The long and narrow pooling core can be suitable for identifying submarine sand waves in side-scan sonar data, and can aggregate global and local context information.

(3) A seesaw loss module is introduced:

the cumulative learning bilateral branch network adopts Softmax to perform cross entropy loss calculation. However, background samples are exponentially more dominant than sand wave samples in seafloor sand wave identification, and they are highly likely to act as negative samples of sand waves. The learning process of the classifier produces a bias. Objects of the sand wave class are more likely to be misclassified as background classes. According to the invention, the teeterboard loss is introduced to replace the original Softmax cross entropy loss function, the ratio of the accumulated training sample number among different categories and the sample error classification during training are explored on line, and the real sonar data sample distribution is gradually approximated. The overwhelming punishment of the background sample to the submarine sand wave category is reduced through the synergistic effect of the relieving factor and the compensating factor, and meanwhile misjudgment of the background sample as the submarine sand wave caused by punishment is avoided. In addition, considering that the geometric forms of the submarine sand waves are complex and changeable, the requirements on the specialization and the detail of manual marking are high. According to the invention, the Seesaw Loss is optimized through sample tag regularization, noise is added to the tags to realize constraint on the accumulated learning bilateral branch network model, the problem of overfitting caused by insufficient submarine sand wave tags is solved, meanwhile, the generalization performance of the network can be effectively improved, and the network is promoted to adapt to the sand wave identification application in complex submarine topography and different sea areas.

The method comprises the steps of S6 in addition to the side-scan sonar submarine sand wave identification methods S1-S5: the method comprises the steps of analyzing the shape of the submarine sand wave, drawing the waveform of the sand wave by extracting the echo intensity waveform of the side-scan sonar image data, analyzing from the angle of the section of the sand wave, and researching the geometric shape analysis of the sand wave; the method specifically comprises the following steps:

(1) Taking the acquired echo intensity waveform as a search area, introducing a template to match and position sand waves, wherein the process is to move the template on the search area and calculate the similarity between the template and the target waveform;

(2) Using an envelope demodulation method, depicting and analyzing sand waves by locating peaks and troughs in the sounding profile;

(3) Checking the rate of change of the spectral sign of the sand wave waveform by using the zero-crossing rate, i.e. the number of times of going from positive to negative or vice versa in a given period;

(4) The cross-correlation coefficients are used to explore the salient features of the sand wave by analyzing the upper and lower envelopes of the sand wave profile waveform:

wherein ,respectively representing the upper envelope and the lower envelope of the echo intensity; /> and />Are respectively->Andis a variance of (2);

(5) Setting the wave crest to be the highest point along the sand wave, and the wavelength S of the sand wave _L Is defined as the horizontal distance from trough to trough, wave height S _H Defined as the vertical height difference between the peak and the adjacent trough of the sand wave, and the projection length of the ascending slope is defined as S _u Descending a slopeProjection length S _d The method comprises the steps of carrying out a first treatment on the surface of the The method comprises the following steps:

1) Actual wavelength S of sand wave _L The estimation of (2) is as follows:

wherein ,is the number of echo intensities occupied by sand waves in the side-scan sonar image, +.>Representing the maximum value of single-side echo intensity points in side-scan sonar images, R _H Representing the actual unilateral operation range of the side-scan sonar on the seabed;

2) The equilibrium relation between the observed sand wave height and the wavelength is analyzed by using a formula put forward by Flemming in field investigation, and then the sand wave height value is estimated as follows:

wherein ,M₁ Is the average wavelength coefficient, f ₁ Is the average wavelength index;

upper limit of wave heightRelationship with wavelength:

wherein ,M₂ For maximum wavelength coefficient, f ₂ Is the maximum wavelength index.

3) The ascending slope projection length S recorded in the process of matching and positioning the wave crest by the template _u And a falling slope projection length S _d The asymmetric morphology index R is estimated by the formula:

R＝S _u /S _d (23)。

the detection method comprises the following steps:

further, the step (1) specifically includes:

1) Using a set of Gaussian templates T _m To simulate the wave characteristics of sand waves:

wherein ,μ_t Is a mathematical expectation of a Gaussian curve, u _d Controlling the moving step length of the template in the abscissa; sigma (sigma) _t Is the variance of the Gaussian curve, controls the width of the curve, w _d Controlling the degree of dispersion of the template; b _t Determining the start point of the Gaussian curve s _d Controlling the moving step length of the template on the ordinate;

2) By root mean square errorAnd (3) measuring the similarity between the echo intensity waveform and the template, and screening out the best matching:

wherein , and T_m ＝[t ₁ ,t ₂ ,t _m ,…t _M ] _1×M The j-th sample data, which is the iping, is used for the template matching part and the template vector, respectively, M is the input sonar data +.>Is a length of (2);

3) According to taylor formula ln (1+x) =x-x ² /2+x ³ /3+....+(-1) ⁿ x ⁿ/n and taking the reciprocal of the root mean square error, logarithmically balancing the weight between the cross-correlation coefficient and the root mean square error, and providing a morphological similarity comprehensive template matching criterion based on a Taylor formula for combining the root mean square error and the cross-correlation coefficient:

the most ideal matching result is the root mean square error between the target waveform and the templateVery small, cross-correlation coefficient->Large, therefore, first of all +.>Then by logarithmic functionError of root mean square->Constraint is performed to balance the capabilities of the root mean square error and the cross correlation coefficient in the template matching algorithm.

Further, in the step (2), the upper and lower envelopes of the sand wave profile are formed into the reconstruction boundaries including the maximum and minimum signals of the echo intensities. Because the topography of the sand wave can generate a vibration signal with lower echo intensity, the envelope demodulation process reflects the amplitude change and direction of the sand wave signal by the extremum of the waveform, and the vibration with low signal-to-noise ratio is detected from a large amount of noise through filtering, so that the sand wave type is determined. The main fluctuation direction of the sand wave waveform can be obtained more clearly through extracting the waveform envelope, so that the noise influence can be eliminated, the sampling rate can be reduced, and the pressure of the storage space of the system can be relieved.

Further, in the step (3), an indirect clue of the sand wave frequency is provided by using the zero crossing rate, and the topological structure in the submarine sand wave is identified by calculating the number of zero crossings of the echo intensity, and the zero crossing rate Z' is expressed as:

wherein , and />The j-th sample of the ipping sonar signal>Echo intensities at K and (K-1), K being echo intensity +.>Length of->Is->Mean value of sgn [ · ]]Is a sign function.

The invention has the advantages and beneficial effects that:

the method is mainly used for identifying the submarine sand waves based on deep learning and is applied to submarine sand wave detection by combining waveform matching analysis. Deep semantic information of a submarine sand wave image is extracted based on a cumulative learning bilateral branch network, a pyramid convolution module and strip pooling are introduced according to the characteristics of the submarine sand wave, self-adaptive optimization and global information enhancement of characteristics are carried out, the problems that the number of submarine sand wave samples are unevenly distributed, morphological characteristics of the sand wave are various in a complex submarine environment, and marine reverberation noise interference is serious are effectively solved, the submarine sand wave is positioned by taking echo intensity as a research object through waveform matching, and morphological characteristics such as wavelength, wave height and asymmetry index of the submarine sand wave and migration trend of the submarine sand wave are analyzed, so that a foundation is laid for further investigation and analysis of the submarine sand wave.

The morphological parameters such as wavelength, wave height, asymmetry index and the like of the submarine sand wave are morphological and topological indexes of sand wave dynamics and are closely related to the tidal current size and direction of the seabed, sediment types, sand wave migration, marine engineering construction such as submarine pipelines, oil platforms and the like, so that the method is an essential parameter for sand wave research. The invention promotes the development and application of the deep learning method in marine geological investigation, and further carries out statistical analysis on morphological parameters of the sand wave, so as to predict migration of the sand wave.

Drawings

Fig. 1 is an overall flow chart of the present invention.

Fig. 2 is the AUV trace and manually noted sand wave region of example 1.

Fig. 3 is an image to be recognized of a submarine sand wave in example 1.

Fig. 4 is a block diagram of a subsea sand wave identification network in example 1.

Fig. 5 is a pyramid convolution block diagram in example 1.

Fig. 6 is a schematic diagram of the strip pooling scheme in example 1.

Fig. 7 is a view of a portion of the submarine sand wave image identification visualization of example 1.

Fig. 8 is a waveform detection diagram of a sand wave profile in example 1.

Fig. 9 is a sand wave envelope of example 1.

Fig. 10 is a background envelope diagram in example 1.

Fig. 11 is a sand wave waveform and zero crossing point profile in example 1.

Fig. 12 is a zero crossing rate statistic graph of the sand wave waveform in example 1.

Fig. 13 is a cross-correlation coefficient comparison chart in example 1.

Fig. 14 is a sand wave section parametric diagram in example 1.

Fig. 15 is a wavelength distribution map of sand waves in example 1.

Fig. 16 is a sand wave asymmetry index profile map in example 1.

Detailed Description

The invention will be further described in detail with reference to the accompanying drawings by way of specific examples in order to make the objects, embodiments and advantages of the invention more apparent.

Example 1: submarine sand waves of the gulf of Qingdao Jiang in 12 th 2019 are taken as objects for identification and morphological analysis.

A specific flow chart of this embodiment is shown in fig. 1.

In this embodiment, a section of AUV system (shown in FIG. 2) deployed from Qingdao Jiaozhou Bay in 2019 as shown in FIG. 3 is specifically adopted to perform submarine detection sand wave image as the image for identification and morphological analysis.

The following steps should be described in detail in connection with the drawings and the specific results and should be merely steps outlined in the summary.

Step one, submarine detection is carried out by depending on an autonomous underwater vehicle (Autonomous Underwater Vehicle, AUV), and the submarine detection device is mainly configured as a side-scan sonar (Marine Sonic Sea Sca) and an underwater video camera (deep sea Power) &Light HD Multi SeaCam), doppler velocimeter (DVL Teledyne RDI), processor (Intel atom eBOX 530-820-FL), inertial navigation unit (KVH 1750 IMU), digital compass (OceanServer OS 5000), depth gauge (Measurement Specialties), underwater lighting (deep sea Power)&Light SLS-5100), acoustic communication (Teledyne BenthosATM-90), etc., wherein the side-scan sonar and the underwater video camera are mainly utilized to mainly take on the acquisition task of detection data. And obtain the longitude of the underwater vehicle ₁ Latitude l ₂ Speed ofHeading h, pitch angle p, roll angle r and raw echo intensity data X. The original echo intensity data is preprocessed and mapped into a side scan sonar image, and is defined as I, and the size is H multiplied by W, wherein H is the height, and W is the width。

Step two, setting side-scan sonar detection for measuring height H _a Sonic pitch R _s Single-side detection range R _h Specifically set H _a ＝10m，R _s ＝200m，R _h = 199.75m. Labeling the side-scan sonar image I by using labelme, taking the existence sand wave and the non-sand wave as corresponding labels y, overlapping and blocking according to longitudinal intervals e ping and transverse intervals f echo intensity points to form an image block u with the size of k x k _ij I=1, (H-k)/e+1, j=1, (W-k)/f+1, the specific 100 echo intensity point overlap separation into image blocks of size 224 x 224 u, training set: test set: validation set = 0.6:0.2:0.2.

Step three, carrying out blind area removal, noise elimination, intensity correction, speed correction and time-varying gain correction data preprocessing on original echo intensity data X to obtain a side-scan sonar image I, wherein the preprocessing work comprises the following steps:

(1) Adaptive filtering of echo intensities using ensemble empirical mode decomposition (Ensemble Empirical Mode Decomposition, EEMD) to obtain echo intensity data

(2) Calculating a correction proportion coefficient through speed correction to obtain echo intensity dataSetting the range distance delta h of the side-scan sonar image corresponding to the ith ping data _i Expressed as:

wherein i is E [1, H]，Δt _i The time difference between the reception of the ipping data and the reception of the (i + 1) th ping data for the sonar sensor,is the instantaneous voyage speed of the AUV at the iping,/>is the instantaneous voyage speed of AUV in the ith+1th ping, W is the width of echo intensity in the ith ping, +.>Is the AUV actual sailing distance, W/2R _s Representing the mapping coefficient between the width of the side-scan sonar image and the actual detection range of the seabed perpendicular to the navigation track direction, deltat _i Is the time difference between the i-th ping and the (i+1) -th ping of the echo received by the side scan sonar.

(3) And (3) compensating partial energy loss caused by the sound wave emitted by the side-scan sonar when the sound wave propagates in the sea water by a time-varying gain correction method:

wherein ,x_ij Is the original echo intensity X _i Is set for the j-th echo intensity value of (c),is the echo intensity value after filtering and velocity correction,/->Is the echo intensity value after time-varying gain correction. A is an echo intensity correction coefficient, and B is an echo intensity correction bias.

4. A side-scan sonar submarine sand wave identification method as defined in claim 3, wherein the step (2) comprises:

1) Adding white noise with a given amplitude to the echo intensity;

where S is the number of trials and r is the residual component of EEMD;

wherein ,σ_p Is the standard deviation of noise, N is IMF _p Is a length of (2);

6) Finally, the denoising method of EEMD comprises the following steps:

noise eliminating the echo intensity data to obtainThe target features are further highlighted using histogram equalization as intensity correction.

Step four, constructing a side-scan sonar submarine sand wave identification network based on a convolutional neural network, wherein a submarine sand wave identification network block diagram is shown in fig. 4, and specifically comprises the following steps:

(1) A cumulative learning Bilateral branch network (Bilate-BranchNetwork, BBN) was used as the backbone network for subsea sand wave identification, including convolutional learning branches and rebalancing branches. For the problem of training sample imbalance, the convolution learning branches are used for sampling training samples (u _c ,y _c ) Training samples (u) are obtained using a reverse sampling mechanism _r ,y _r ) As an input to the rebalancing branch, focus learning attention on the sand wave side-scan image data with a small number of training samples;

Firstly, learning a general mode, and then gradually focusing on tail data; after convolution learning branches and rebalancing branches, the feature vectors f are output through corresponding full-connection layers respectively _c and f_r Then takes it as input into accumulation theoryPerforming learning; if the total number of training iterations is T _max The current training iteration number is T, and the adjustment parameter alpha is:

the BBN loss function in the whole learning process is a seesaw loss function:

L＝αL _s (z _c )+(1-α)L _s (z _r ) (12)

(2) Based on a cumulative learning bilateral branch network as a main body, a pyramid convolution module, a strip pooling module and a teeterboard loss are introduced to form a submarine sand wave identification network:

the horizontal pooling output dimension is (f) _h 1, c)

The vertical pooling output dimension is (1, f _w C)/>

The introduced seesaw loss module is specifically as follows:

(1) Introducing seesaw loss to a cumulative learning bilateral branch network BBN; introducing a distribution function mu (k) independent of the input samples and a smoothing parameter eta, optimizing to y' = (1-eta) y+eta mu (k) by sample label regularization (Label Smoothing Regularization, LSR); according to the seesaw loss, dynamically rebalancing positive and negative sample gradients of each category through the proportion of the number of samples in the training process:

for samples of class m, the negative sample gradient applied to class n is:

1) The pyramid convolution module is introduced into the cumulative learning bilateral branch network BBN. The convolution learning branch and the rebalancing branch have the same structure, and all parameters except the last residual block are shared. Analyzing input by utilizing a Local multiscale context aggregation module (Local multi-scale context aggregation module) and a Global multiscale context aggregation module (Global multi-scale context aggregation module) to enrich the feature capturing function of the accumulated learning bilateral branch network;

2) For each branch network, firstly extracting a feature map through a convolution layer, then adding a network receptive field through Residual Block (Residual Block), then through cavity convolution (di conv) and pyramid convolution (PyConv), fusing the captured features of multiple convolution kernel scales through pyramid convolution, enabling the depth of a convolution kernel to change along each level to form full depth and connectivity, and obtaining the feature map of the initial dimension through improved Residual Block and convolution.

3) Repeating the steps, and respectively obtaining a characteristic diagram by the convolution learning branch and the rebalancing branch through the multistage improved residual block. In the hole convolution process, the convolution scale is adjusted by adding the super-parameter hole rate on the basis of standard convolution, and the equivalent convolution kernel size after the hole convolution is introduced is k' =k+ (k-1) x (d-1), wherein k is the size of the original convolution kernel, and d is the super-parameter hole rate.

1) Introducing a strip pooling module to the cumulative learning bilateral branch network BBN, and utilizing a narrow pooling core to better fit and identify the sand wave form; let the dimension be f _h ×f _w Is input feature map of (a)c represents the number of channels, and the channels are respectively fed into two parallel paths of horizontal pooling and vertical pooling in the process of strip pooling;

2) An improved strip pooling strategy is established, and effective feature expressions are extracted from a row-by-row and column-by-column angle in parallel through equal weight combination of average pooling and maximum pooling:

the horizontal pooling output dimension is (f) _h 1, c)

The vertical pooling output dimension is (1, f _w C)

The introduced seesaw loss module is specifically as follows:

1) Introducing seesaw loss to a cumulative learning bilateral branch network BBN; introducing a distribution function mu (k) independent of the input samples and a smoothing parameter eta, optimizing to y' = (1-eta) y+eta mu (k) by sample label regularization (Label Smoothing Regularization, LSR); according to the seesaw loss, dynamically rebalancing positive and negative sample gradients of each category through the proportion of the number of samples in the training process:

for samples of class m, the negative sample gradient applied to class n is:

penalty increase for class n (σ _n /σ _m ) ^q Multiple, wherein q is the superstrate of the control ratioParameters; c (C) _mn When=1, C _mn By only applying the relieving factor M _mn The method comprises the steps of carrying out a first treatment on the surface of the Thereby avoiding misclassification of category n due to the overwhelming penalty mitigation for category n.

The result of the submarine sand wave identification experiment is visualized as shown in fig. 7, and the effect of identifying part of submarine sand wave images is shown. Fig. (a) is an original sonar image of the removed water body, and fig. (b) is a group Truth of fig. (a). Fig. (c) and (d) are raster image visualizations of the result of the recognition of the side-scan sonar image by the Resnet50 and the algorithm of the present invention, respectively. Graph (e) and graph (f) are the projections of the recognition results of the Resnet50 and the algorithm of the invention, respectively, in grid form into the original sonar image (c). By comparing the Resnet50 algorithm with the recognition result of the invention, the realization of a better recognition effect of the invention can be obviously seen. In the recognition effect of the Resnet50, there is a case where the background is erroneously recognized as a sand wave, and there is a case where the background is erroneously recognized as a background for a region where the sand wave is gentle.

Sand wave waveform analysis method:

and eleventh, drawing a sand wave waveform by extracting the echo intensity of the side-scan sonar, analyzing from a sand wave section angle, and researching the geometric form of the sand wave. The geometric analysis of the sand wave comprises the following steps:

(1) And taking the acquired echo intensity waveform as a search area, introducing a template to match and position sand waves, wherein the process is to move the template on the search area and calculate the similarity between the template and the target waveform.

1) Using a set of Gaussian templates T _m To simulate the wave characteristics of sand waves, as shown in the following formula:

wherein ,μ_t Is a mathematical expectation of a Gaussian function, u _d Controlling the moving step length of the template in the abscissa; sigma (sigma) _t Is the variance of the Gaussian curve, controls the width of the curve, w _d Controlling the degree of dispersion of the template; b _t Determining the start point of the Gaussian curve s _d Controlling the template to be inThe movement step on the ordinate.

wherein , and T_m ＝[t ₁ ,t ₂ ,t _m ,…t _M ] _1×M The j-th sample data of the i-th ping is used for the template matching part and the template vector, respectively, and M is input sonar data +.>Is a length of (c). The present invention tries to calculate the echo intensity waveform by cross-correlation coefficients +. >And template T _m Overall similarity between->

3) The invention is based on the Taylor formula ln (1+x) =x-x ² /2+x ³ /3+....+(-1) ⁿ x ⁿ/n and taking the reciprocal of the root mean square error, logarithmically balancing the weight between the cross-correlation coefficient and the root mean square error, and providing a morphological similarity comprehensive template matching criterion based on a Taylor formula for combining the root mean square error and the cross-correlation coefficient:

the evaluation of the space similarity is enhanced, so that Euclidean distance matching of the Gaussian template and the target waveform on the same curve length is considered, and the similarity of geometric forms is also considered. The most ideal matching result is the root mean square error between the target waveform and the templateVery small, cross-correlation coefficient->Large, we have therefore constructed firstHowever, it was found that when a portion of the points in the template are close to or even coincident with the echo intensity waveform,will be very large, the matching algorithm will have too strong capacity to evaluate the spatial distance, the cross-correlation coefficient +.>The effect of (2) is almost ignored, resulting in an imbalance between the two, resulting in a template being selected with partial overlap but with a very different overall trend or even opposite trend between the template and the echo intensity waveform. The project is characterized by logarithmic function->Error of root mean square->Constraint is performed to balance the capabilities of the root mean square error and the cross correlation coefficient in the template matching algorithm. Furthermore, when an abnormal value occurs in the echo intensity waveform, the root mean square error +. >When very large, add>Will be small, according to the formula->It can be seen that the matching criteria of the project design are +.>Will be similar to and />For the values at this time are smaller +.>Ratio of change of (2)Is itself more sensitive and thus can effectively avoid logarithmic operation pairs +.>At the same time, sufficient attention is paid to outliers with large root mean square errors.

The partial result of the template matching in the sand wave experiment (sand wave profile waveform detection) is shown in fig. 8, and it can be seen from the graph that the form similarity comprehensive template matching criterion provided by the invention has good matching performance, and can realize matching of sand waves with different scales.

(2) Sand waves are depicted and analyzed by locating peaks and troughs in the sounding profile using an envelope demodulation method. The upper and lower envelopes of the sand wave profile are formed into reconstruction boundaries containing the maximum and minimum signals of echo intensities. Because the topography of the sand wave can generate a vibration signal with lower echo intensity, the envelope demodulation process reflects the amplitude change and direction of the sand wave signal by the extremum of the waveform, and the vibration with low signal-to-noise ratio is detected from a large amount of noise through filtering, so that the sand wave type is determined. The sand wave envelope is shown in fig. 9 and the background envelope is shown in fig. 10, wherein the upper curve represents the upper envelope of the waveform, the middle curve represents the echo intensity waveform, and the lower curve represents the lower envelope of the waveform. The main fluctuation direction of the sand wave waveform can be obtained more clearly through extracting the waveform envelope, so that the noise influence can be eliminated, the sampling rate can be reduced, and the pressure of the storage space of the system can be relieved.

(3) The rate of change of the spectral sign of the sand wave waveform, i.e., the number of times that the wave waveform is reversed or reversed from positive to negative within a given period, is checked using the Zero Crossing Rate (ZCR). The zero crossing rate is used to provide an indirect clue to the sand wave frequency and to identify the topology in the subsea sand wave by counting the number of zero crossings of the echo intensity, the zero crossing rate Z' can be expressed as:

wherein , and />The j-th sample of the ipping sonar signal>Echo intensities at K and (K-1), K being echo intensity +.>Length of->Is->Mean value of sgn [ · ]]Is a sign function. As shown in FIG. 11, (a) is the case without sand wave, and (b) (c) (d) is the case with sand wave waveform and zero crossing point distributionThe sand wave exists; statistics of zero crossing rate of waveforms as shown in fig. 12, the zero crossing rates of the regions where sand waves exist and the regions where sand waves do not exist are statistically analyzed in (a) and (b), respectively, to explore the zero crossing rate distribution of the waveforms of echo intensities with sand waves and without sand waves. In a simulation experiment, the invention discovers that in sand wave data acquired by the Bay of the 12 th month of the 2019, the zero crossing rate of the sand wave is kept between 0.3 and 0.4, and the echo intensity of the sand wave shows a strong periodicity rule. The zero crossing rate of the background waveform is higher, the fluctuation rate is fast between 0.4 and 0.6, and the echo intensity is very chaotic, and is mainly shown as interference noise distribution. It can be seen that the zero crossing rate of the sand-free waveform is similar to the ambient noise and greater than the zero crossing rate of the sand wave. Thus, low zero crossing rate can be a significant feature of a sand wave as distinguished from a background waveform without a sand wave.

wherein ,respectively representing the upper and lower envelopes of the echo intensities. /> and />Are respectively->Andis a variance of (c). The cross-correlation coefficient comparison is shown in FIG. 13, wherein a) and (b) are the cross-correlation coefficient statistics of the sand wave waveform and the background waveform, respectively, and the upper and lower envelope cross-correlation coefficients of the sand wave can be found to be relatively highLarge, both will exhibit synchronous ripple conditions. The cross-correlation coefficient of the waveform of the background area is small, the upper envelope and the lower envelope are almost independently changed, no synchronous trend exists, and the envelope shows randomness, which is consistent with the fact that the envelope mainly consists of noise interference. And the present invention found that when the cross correlation coefficient of the upper and lower envelopes is greater than 0.7, the probability of existence of sand waves is 90%.

(5) Setting the wave crest to be the highest point along the sand wave, and the wavelength S of the sand wave _L Is defined as the horizontal trough-to-trough distance. Wave height S _H Is defined as the vertical height difference between the peak and the adjacent trough of the sand wave. And define the projection length of the ascending slope as S _u The projection length of the descending slope is S _d . The sand wave profile parameter diagram is shown in fig. 14.

1) Actual wavelength S of sand wave _L The estimation of (2) is as follows:

wherein ,is the number of echo intensities occupied by sand waves in the side-scan sonar image, +.>Representing the maximum value of single-side echo intensity points in side-scan sonar images, R _H Representing the actual single-sided operating range of the side-scan sonar at the seafloor. The parameters recorded by the morphological similarity comprehensive template matching criterion provided by the project are used for carrying out statistical analysis mapping on the actual wavelength of the sand wave, as shown in fig. 15. We can find that the wavelength of the simulation experiment area is mainly between 10m-30m, wherein the wavelength is more than 10m and less than 15m and accounts for 25.9%; 59.9% of wavelengths greater than 15m and less than 30 m; and the wavelength is less than 10m and accounts for 14.2 percent. We can find that 85.8% of the sand wavelengths are all greater than 10 m.

wherein ,M₁ Is the average wavelength coefficient, f ₁ Is the average wavelength index.

Upper limit of wave heightRelationship with wavelength:

The statistical analysis shows that the wavelength of sand wave in the Bay area of Guozhou is mainly more than 15m, the wave height is about 0.86m, and the upper limit of the wave height is mainly 2.23m.

R＝S _u /S _d (29)

the asymmetric morphology index R is calculated and its distribution is mapped into fig. 16. We found that the sand wave fraction with R < 0.5 West-biased was 63.4%. The fraction of R < 1.5, which is normally distributed, is 26.4%. The proportion of R > 1.5 to the east is 10.2 percent. Thus, we can find that the steep slope of most sand waves in the gulf of gum face west, the gentle slope faces east, and infer that sand waves may have a tendency to migrate west. Our AUV was deployed a second time around the same study area of the gulf of yurt 1, month 1, and submarine sand wave data was collected over a duration of more than one year to verify the predicted direction of sand wave migration.

Geometric analysis of sand waves:

the form of the sand wave reflects the seabed power, seabed sediment, sand wave size and relative motion strength, and is a key factor for researching the seabed sand wave. The method designs a morphological similarity comprehensive template matching criterion for positioning the sand waves. The upper and lower envelopes of the waveform are extracted to analyze the main direction of fluctuation of the acquired waveform. The zero crossing rate analysis of the side scan sonar data found that there was a sand wave region. And analyzing the upper envelope and the lower envelope of the echo intensity waveform by adopting the cross-correlation coefficient, and judging the possibility of existence of the sand wave. And obtaining morphological parameters such as wavelength, wave height, asymmetry index and the like of the sand wave. And obtaining the migration trend of the sand wave through statistical analysis.

Claims

1. A side-scan sonar submarine sand wave detection method based on a multi-scale convolution and pooling strategy is characterized by comprising the following steps:

s1: collecting submarine side-scan sonar data and attitude positioning data;

s2: preprocessing the submarine side-scan sonar data collected in the step S1;

s5: identifying the side-scan sonar testing data after S2 pretreatment by using the trained submarine sand wave identification network of S4, and outputting an identification result;

in the step S3, a submarine sand wave identification network based on an improved convolutional neural network is constructed, and the submarine sand wave identification network specifically comprises the following steps:

(1) Using a cumulative learning bilateral branch network BBN as a main network for submarine sand wave identification, wherein the main network comprises convolution learning branches and rebalancing branches; for the problem of training sample imbalance, the convolution learning branches are used for sampling training samples (u _c ,y _c ) Training samples (u) are obtained using a reverse sampling mechanism _r ,y _r ) As an input to the rebalancing branch, focus learning attention on the sand wave side-scan image data with a small number of training samples;

(2) Based on a cumulative learning bilateral branch network BBN as a main body, a pyramid convolution module, a strip pooling module and a seesaw loss module are introduced to jointly construct a submarine sand wave identification network;

The pyramid convolution module is introduced, and the pyramid convolution module is specifically as follows:

(1) Introducing a pyramid convolution module into a cumulative learning bilateral branch network BBN; the convolution learning branch and the rebalancing branch have the same structure, and all parameters except the last residual block are shared; analyzing input by utilizing a local multi-scale context aggregation module and a global multi-scale context aggregation module so as to enrich the feature capturing function of the accumulated learning bilateral branch network;

(2) For each branch network, firstly extracting a feature map through a convolution layer, adding a network receptive field through a residual block, then carrying out cavity convolution and pyramid convolution, fusing the captured features of various convolution kernel scales through pyramid convolution, changing the depth of a convolution kernel along each level to form full depth and connectivity, and obtaining the feature map of an initial dimension again through improved residual block and convolution;

(3) Repeating the steps, namely respectively obtaining a characteristic diagram by a convolution learning branch and a rebalancing branch through multistage improvement residual blocks, wherein the convolution scale is adjusted by adding super-parameter void ratio on the basis of standard convolution in the void convolution process, and the equivalent convolution kernel size after introducing void convolution is k' =k+ (k-1) x (d-1), wherein k is the size of the original convolution kernel, and d is the super-parameter void ratio;

the horizontal pooling output dimension is (f) _h 1, c)

The vertical pooling output dimension is (1, f _w C)

The two output features of the strip pool are expanded through convolution operation and respectively transformed into the size f _H ×f _W Horizontal bar expansion of (2)And vertical bar expansion->By y=y ^H +y ^V Will y ^H and y^V Combining; by z=scale (P _ij Sigma (f (y)) is calculated to have global prior feature output so as to create feature capture suitable for the long-strip-shaped morphology of the sand wave, assist in seabed sand wave feature optimization and parameter learning, wherein Scale (·, ·) is the multiplication and addition element by element, sigma is an activation function, and f is a 1×1 convolution;

the introduced seesaw loss module is specifically as follows:

(1) Introducing seesaw loss to a cumulative learning bilateral branch network BBN; introducing a distribution function mu (k) independent of the input samples and a smoothing parameter eta, regularizing LSR to y' = (1-eta) y+eta mu (k) by sample labels; according to the seesaw loss, dynamically rebalancing positive and negative sample gradients of each category through the proportion of the number of samples in the training process:

for samples of class m, the negative sample gradient applied to class n is:

penalty increase for class n (σ _n /σ _m ) ^q Multiple, wherein q is a hyper-parameter controlling the ratio；C _mn When=1, C _mn By only applying the relieving factor M _mn The method comprises the steps of carrying out a first treatment on the surface of the Thereby avoiding misclassification of category n due to the overwhelming penalty mitigation for category n.

2. The side-scan sonar submarine sand wave detection method as defined in claim 1, wherein in S1,

(2) High H by side-scan sonar detection _a Sonic pitch R _s Single-side detection range R _h The method comprises the steps of carrying out a first treatment on the surface of the Labeling the side-scan sonar image I, dividing the side-scan sonar image I into two types of sand waves and no sand waves, wherein the corresponding label is y, and each s is arranged according to the longitudinal interval _v Ping and lateral spacing per s _h Overlapping and blocking the echo intensity data points to form an image block P with the size of l _ij ,i＝1,...,(H-l)/s _v +1,j＝1,...,(W-l)/s _h +1 and respectively occupy the proportion lambda according to the training set, the test set and the verification set ₁ 、λ ₂ 、λ ₃ Distribution of sample data (lambda) ₁ +λ ₂ +λ ₃ ＝1)。

3. The side-scan sonar submarine sand wave detection method of claim 1, wherein in S2, the preprocessing comprises:

(1) Carrying out blind area removal, noise elimination, intensity correction, speed correction and time-varying gain correction data preprocessing on the side-scan sonar image I respectively;

(2) Adaptive filtering of echo intensities using ensemble empirical mode decomposition EEMD to obtain echo intensity data

wherein i is E [1, H]，Δt _i The time difference between the reception of the ipping data and the reception of the (i + 1) th ping data for the sonar sensor,is the instantaneous speed of the AUV at the ipping, < >>Is the instantaneous voyage speed of AUV at the i+1th ping, W is the width of echo intensity in the ipping, < ->Is the AUV actual sailing distance, W/2R _s Representing the mapping coefficient between the width of the side-scan sonar image and the actual detection range of the seabed perpendicular to the navigation track direction, deltat _i Is the time difference between the ipping and (i+1) th ping of the echo received by the side scan sonar;

wherein ,x_ij Is originalEcho intensity X _i Is set for the j-th echo intensity value of (c),is the echo intensity value after filtering and velocity correction,/->Is the echo intensity value after time-varying gain correction; a is an echo intensity correction coefficient, and B is an echo intensity correction bias.

4. A side-scan sonar submarine sand wave detection method as recited in claim 3, wherein said step (2) comprises:

1) Adding white noise with a given amplitude to the echo intensity;

2) Performing Empirical Mode Decomposition (EMD) on the signal with echo intensities and added white noise to obtain an Intrinsic Mode Function (IMF) component and a residual component;

where S is the number of trials and r is the residual component of EEMD;

6) Finally, the denoising method of EEMD comprises the following steps:

5. The method for detecting side-scan sonar submarine sand wave according to claim 1, wherein in the step (1), a cumulative learning bilateral branch learning strategy is specifically:

wherein ,N_m 、N _j The number of samples of class m and class j, N _max All samples, C is the total number of categories;

2) Through an accumulated learning strategy, adjusting a parameter alpha to control weight and classification loss occupied by two branch characteristics, transferring learning key points of bilateral branches, and focusing sand wave side scanning image data with a small number of training samples;

z＝αW _c ^T f _c +(1-α)W _r ^T f _r (11)

the BBN loss function in the whole learning process is a seesaw loss function:

L＝αL _s (z _c )+(1-α)L _s (z _r ) (12)

6. The side-scan sonar submarine sand wave detection method of claim 1, further comprising S6: the analysis of the seabed sand wave morphology, drawing a sand wave waveform by extracting an echo intensity waveform of side-scan sonar image data, analyzing from a sand wave section angle, and researching the geometric morphology of the sand wave, wherein the method specifically comprises the following steps:

(1) Segmenting the acquired echo intensities to form a wavelet shape serving as a search area, introducing a template matching positioning sand wave, moving the template on the search area, and calculating the similarity between the template and the target waveform;

(2) Drawing an envelope curve for positioning wave crests and wave troughs in the echo intensity profile by an envelope demodulation method;

(3) Checking the rate of change of the spectrum sign of the sand wave waveform by using the zero-crossing rate ZCR, namely the number of times of negative or reverse from positive in a given period;

(4) The cross-correlation coefficient is used for exploring the obvious correlation characteristics of the sand wave profile waveform by analyzing the upper envelope and the lower envelope of the sand wave profile waveform:

wherein , respectively representing the upper envelope and the lower envelope of the echo intensity; /> and />Are respectively-> and />Is a variance of (2);

(5) Let sand wave wavelength S _L For the horizontal distance from trough to trough, the wave height S _H Defining the projection length of the ascending slope as S for the vertical height difference between the wave crest and the adjacent wave trough of the sand wave _u The projection length of the descending slope is S _d The method specifically comprises the following steps:

1) Sand wave actual wavelength S _L The estimation is as follows:

upper limit of wave heightThe relation with wavelength is:

wherein ,M₂ For maximum wavelength coefficient, f ₂ Is the maximum wavelength index;

R＝S _u /S _d (23)。

7. the method for detecting side-scan sonar submarine sand waves as defined in claim 6, wherein the step (1) specifically comprises:

wherein ,μ_t Is a mathematical expectation of a Gaussian function, u _d Controlling the moving step length of the template in the abscissa; sigma (sigma) _t Is the variance of the Gaussian function, controls the width of the function, w _d Controlling the degree of dispersion of the template; b _t Determines the start point of the Gaussian function s _d Controlling the moving step length of the template on the ordinate;

wherein , and T_m ＝[t ₁ ,t ₂ ,t _m ,…t _M ] _1×M The j-th sample data, which is the iping, is used for template matching of the sub-waveform and the template vector, respectively, M is the input echo intensity data +.>Is a length of (2);

3) According to taylor formula ln (1+x) =x-x ² /2+x ³ /3+....+(-1) ⁿ x ⁿ/n and taking the reciprocal of the root mean square error and logarithmically, and providing a morphological similarity comprehensive template matching criterion based on a Taylor formula and combining the root mean square error and the cross-correlation coefficient:

first constructThen go through logarithmic function->For root mean square errorConstraint is performed to balance the capabilities of the root mean square error and the cross correlation coefficient in the template matching algorithm.

8. The method for detecting side-scan sonar submarine sand waves as defined in claim 6, wherein in the step (2), the upper and lower envelopes of the sand wave section are formed into a reconstruction boundary containing the maximum and minimum signals of the echo intensities; in the step (3), the zero crossing rate is used to provide an indirect clue of the sand wave frequency, and the topological structure in the submarine sand wave is identified by calculating the number of zero crossings of the echo intensity, and the zero crossing rate Z' is expressed as: