CN113516626A - Side-scan sonar seabed sand wave detection method based on multi-scale convolution and pooling strategy - Google Patents
Side-scan sonar seabed sand wave detection method based on multi-scale convolution and pooling strategy Download PDFInfo
- Publication number
- CN113516626A CN113516626A CN202110497412.0A CN202110497412A CN113516626A CN 113516626 A CN113516626 A CN 113516626A CN 202110497412 A CN202110497412 A CN 202110497412A CN 113516626 A CN113516626 A CN 113516626A
- Authority
- CN
- China
- Prior art keywords
- sand
- wave
- sand wave
- convolution
- echo intensity
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 239000004576 sand Substances 0.000 title claims abstract description 254
- 238000011176 pooling Methods 0.000 title claims abstract description 66
- 238000001514 detection method Methods 0.000 title claims abstract description 34
- 230000002146 bilateral effect Effects 0.000 claims abstract description 35
- 238000004458 analytical method Methods 0.000 claims abstract description 14
- 238000011835 investigation Methods 0.000 claims abstract description 10
- 238000005457 optimization Methods 0.000 claims abstract description 7
- 238000000034 method Methods 0.000 claims description 63
- 238000012937 correction Methods 0.000 claims description 32
- 230000006870 function Effects 0.000 claims description 32
- 238000012549 training Methods 0.000 claims description 32
- 230000008569 process Effects 0.000 claims description 24
- 230000002776 aggregation Effects 0.000 claims description 16
- 238000004220 aggregation Methods 0.000 claims description 16
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 15
- 230000001186 cumulative effect Effects 0.000 claims description 12
- 230000000116 mitigating effect Effects 0.000 claims description 12
- 238000005070 sampling Methods 0.000 claims description 11
- 238000012360 testing method Methods 0.000 claims description 11
- 238000000354 decomposition reaction Methods 0.000 claims description 9
- 238000007781 pre-processing Methods 0.000 claims description 9
- 238000004422 calculation algorithm Methods 0.000 claims description 8
- 238000001914 filtration Methods 0.000 claims description 8
- 238000013527 convolutional neural network Methods 0.000 claims description 7
- 238000005315 distribution function Methods 0.000 claims description 7
- 238000002474 experimental method Methods 0.000 claims description 7
- 230000001174 ascending effect Effects 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 6
- 230000008030 elimination Effects 0.000 claims description 6
- 238000003379 elimination reaction Methods 0.000 claims description 6
- 238000009499 grossing Methods 0.000 claims description 6
- 230000006872 improvement Effects 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 6
- 238000009825 accumulation Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 4
- 230000006978 adaptation Effects 0.000 claims description 4
- 238000002372 labelling Methods 0.000 claims description 4
- 239000013535 sea water Substances 0.000 claims description 4
- 238000001228 spectrum Methods 0.000 claims description 4
- 230000003044 adaptive effect Effects 0.000 claims description 3
- 230000005540 biological transmission Effects 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 2
- 239000011800 void material Substances 0.000 claims 1
- 238000009826 distribution Methods 0.000 abstract description 11
- 230000000877 morphologic effect Effects 0.000 abstract description 10
- 238000011160 research Methods 0.000 abstract description 8
- 238000013508 migration Methods 0.000 abstract description 6
- 230000005012 migration Effects 0.000 abstract description 6
- 239000013049 sediment Substances 0.000 description 8
- 238000010586 diagram Methods 0.000 description 7
- 238000007619 statistical method Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 239000003292 glue Substances 0.000 description 3
- 238000012876 topography Methods 0.000 description 3
- 238000012800 visualization Methods 0.000 description 3
- 230000002159 abnormal effect Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 239000002245 particle Substances 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 238000003860 storage Methods 0.000 description 2
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000011109 contamination Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000001125 extrusion Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 229910052500 inorganic mineral Inorganic materials 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 239000011707 mineral Substances 0.000 description 1
- 238000012821 model calculation Methods 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 230000002195 synergetic effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10132—Ultrasound image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Pure & Applied Mathematics (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Mathematical Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Computational Linguistics (AREA)
- Computational Mathematics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Biomedical Technology (AREA)
- Mathematical Optimization (AREA)
- Databases & Information Systems (AREA)
- Quality & Reliability (AREA)
- Algebra (AREA)
- Probability & Statistics with Applications (AREA)
- Operations Research (AREA)
- Measurement Of Velocity Or Position Using Acoustic Or Ultrasonic Waves (AREA)
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 seabed sand wave image is extracted based on an accumulative learning bilateral branch network, a pyramid convolution module and strip pooling are introduced according to the self characteristics of the seabed sand wave to perform characteristic self-adaptive optimization and global information enhancement, the problems that the number distribution of seabed sand wave samples is unbalanced, the morphological characteristics of sand waves in a complex seabed environment are various, and the interference of ocean reverberation noise is serious are effectively solved, the seabed sand wave is positioned by using echo intensity as a research object through waveform matching, the morphological characteristics of the seabed sand wave such as wavelength, wave height and asymmetric index and the migration trend of the seabed sand wave are analyzed, and therefore a foundation is laid for further investigation and analysis of the seabed sand wave.
Description
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
The ocean exploration is mainly used for exploring various geophysical field characteristics, geological structures and mineral resources on the seabed and exploring the distribution characteristics of seabed placer and sand waves. Among them, the seabed sand wave is a topographic form formed under various environmental factors such as sediment particle size, tidal characteristics and related residual flow, and is generally distributed in various sea areas. Submarine sand wave identification is an important mode for investigating submarine sediments, and is an important research content in the fields of marine sediment bathymetry, marine geomorphology, marine geophysics, marine geological investigation and the like. The method for identifying seabed sand waves generally comprises a numerical value algorithm, a model calculation method, remote sensing detection and a field investigation method. Currently, an acoustic field investigation method based on an Autonomous Underwater Vehicle (AUV) or a cabled Underwater Vehicle (ROV) carrying a side-scan sonar is widely applied to the research of submarine sand waves.
The side-scan sonar is usually installed on two sides of a carrying device, a transducer array emits sound waves according to a certain frequency, the sound waves return when encountering seabed sediments such as sand waves and the like, a receiving array of the side-scan sonar receives sonar signals obtained by reflection, diffraction or scattering from the bottom frequency by frequency and splices echo intensities into side-scan sonar images, different information of seabed characteristics is provided according to different frequencies, complicated topography and landform of the seabed is reflected, and seabed sand wave investigation in a large range can be realized. The method for carrying out seabed sand wave investigation through the side-scan sonar is widely applied, provides a research foundation for further seabed sand wave simulation and prediction analysis, and faces a lot of challenges at present.
Firstly, the quality of the collected side-scan sonar image is affected due to factors such as ocean reverberation noise, noise influence of detection equipment, energy loss of sound waves in seawater, limited single-transmission information amount of the sound waves and the like, and the sand wave edge is unclear. These problems all cause that the traditional image recognition method is seriously hindered when the sand wave target characteristics are manually extracted, all characteristics are difficult to be considered, the shape of the seabed sand wave is fitted, and the processing speed and the accuracy of sand wave recognition are influenced.
Secondly, in data collected by the side-scan sonar, the number of images in a background area is far larger than that of the sand waves, sample imbalance is presented, so that the deep learning network puts more attention on improvement of the accuracy of the samples in the background area, and excessive negative sample gradients are applied to the sand wave area images by the background area images, so that the positive sample gradients in the sand wave area are submerged, overfitting is caused, and the identification accuracy of the sand wave data is seriously influenced.
Thirdly, the submarine topography features are diversified due to the difference of hydrological environments such as tidal flows, residual flows and submarine sediment types in different sea areas, the geometric morphology features of sand waves, the sand wave categories and the submarine environments have multilayer and complex nonlinear relations, and the shapes of the sand waves under different submarine topography are greatly different, so that the sand waves need to be identified comprehensively according to the surrounding environment, and the large sensing field and the strong generalization performance and reasoning capability need to be possessed.
Finally, no matter the sand waves are linear type sand waves, the crescent sand waves have the common strip shape, the ripples are distributed discretely at intervals, the context flexibility of sand wave feature capture is limited by the regular square pooling cores, the sand wave features are not beneficial to being effectively extracted, and the traditional square pooling cores are likely to introduce noise and other unnecessary connections to interfere the identification of the seabed sand waves.
Disclosure of Invention
In view of the above problems, the present invention provides a method for detecting a side-scan sonar subsea sand wave based on a multi-scale convolution and pooling strategy, including a method for identifying a side-scan sonar subsea sand wave and a method for morphological analysis.
In order to achieve the purpose, the invention adopts the following specific technical scheme:
a side-scan sonar seabed sand wave detection method based on a multi-scale convolution and pooling strategy comprises the following steps:
s1: collecting seabed side-scan sonar data and attitude positioning data;
s2: preprocessing the collected seabed side-scan sonar data of S1;
s3: constructing a submarine sand wave identification network based on an improved convolutional neural network;
s4: training the submarine sand wave recognition network obtained by S3;
s5: and (5) recognizing the side-scan sonar test data preprocessed by the S2 by using the seabed sand wave recognition network trained by the S4, and outputting a recognition result.
Further, in the step S1,
(1) data are acquired by using a side scan sonar seabed detection means: including longitude lon, latitude lat, speed of the underwater vehicleCourse h, pitch angle p, roll angle r, and raw echo intensity data XiH, i 1.. H; collectively mapping original echo intensity data X into a side-scan sonar image through preprocessing, defining the image as I, and setting the size as H multiplied by W, wherein H is height and W is width;
(2) high H determination by side-scan sonaraSonic slope distance RsSingle sided detection range Rh(ii) a Labeling the side-scan sonar image I, wherein the side-scan sonar image I is divided into two types of sand waves and sand-free waves, the corresponding label is y, and every s is arranged according to the longitudinal intervalvOne ping and one transverse interval per shThe echo intensity data points are overlapped and blocked to form an image block P with the size of l x lij,i=1,...,(H-l)/sv+1,j=1,...,(W-l)/sh+1, and according to the training set, the test set,The verification sets respectively occupy the proportion lambda1、λ2、λ3Sample data (λ) is allocated1+λ2+λ3=1);
Further, in S2: the pretreatment comprises the following steps:
(1) and respectively carrying out data preprocessing of blind area removal, noise elimination, intensity correction, speed correction and time-varying gain correction on the side-scan sonar image I.
(2) Using Ensemble Empirical Mode Decomposition (EEMD) to perform adaptive filtering on echo intensity to obtain echo intensity data
(3) Calculating correction proportionality coefficient by speed correction to obtain echo intensity dataSetting the range distance delta h of the side-scan sonar image corresponding to the ith ping dataiExpressed as:
wherein i ∈ [1, H)],ΔtiThe time difference between the reception of the fourth ping data and the reception of the (i +1) th ping data for the sonar transducer transmission,is the instantaneous speed of the AUV navigating at the ith ping,is the instantaneous cruising speed of the AUV at the i +1 th ping, W is the width of the echo intensity in the ith ping,is the actual sailing distance of AUV, W/2RsRepresenting the space between the width of the side scan sonar image and the actual detection range of the sea bottom perpendicular to the navigation track directionOf the mapping coefficient, Δ tiIs the time difference between the ith ping and the (i +1) th ping in which the side scan sonar receives the echo.
(4) Through the time-varying gain correction method, the partial energy loss caused by the propagation of the sound wave emitted by the side-scan sonar in the seawater is compensated:
wherein ,xijIs the original echo intensity XiThe jth echo intensity value of (a) is,is the echo intensity value after filtering and speed correction,are the echo intensity values after time varying gain correction. A is the echo intensity correction coefficient, and B is the echo intensity correction offset.
Further, the step (1) comprises:
1) adding white noise of a given amplitude to the echo intensity;
2) performing Empirical Mode Decomposition (EMD) on the signal with the echo intensity and the added white noise to obtain an Intrinsic Mode Function (IMF) component and a residual component;
3) this operation was repeated using a given number of trials; the amount of IMF was constant in each experiment; all of the tests can be written as,
wherein ,XiIs the ith original signal in the echo intensity; n issIs white noise added in the s trial; IMFspIs the p-th IMF module in the s-th experiment; r issIs the remaining component in the s trial;
4) calculating the integral average value of all tests, and decomposing the calculated echo intensity of sonar by using the set empirical modeiThe final decomposition result of (a) is expressed as:
wherein S is the number of trials and r is the residual component of EEMD;
5) the first batch p is selected1IMF and denoising EEMD through a general threshold formula, as follows:
wherein ,σpIs the standard deviation of noise, N is IMFpLength of (d);
in determining the general threshold TpThen, selecting a soft threshold function to filter the noise component coefficient in the high-frequency IMF;
6) finally, the EEMD denoising method comprises the following steps:
the echo intensity data is subjected to noise elimination to obtainHistogram equalization is used as an intensity correction to further highlight target features.
Further, in S3, a submarine sand wave identification network based on the improved convolutional neural network is constructed, specifically:
(1) using cumulative learning Bilateral Branch Network (Bilateral-Branch Network)BBN) as the backbone network for ocean bottom sand wave identification, including the convolutional learning branch and the rebalancing branch. For the training sample imbalance problem, the convolution learning branch uniformly samples the training sample (u)c,yc) Obtaining training samples (u) using an inverse sampling mechanismr,yr) As input to the rebalance branch, the learning attention is focused on the sand wave side scan image data with a small number of training samples;
(2) on the basis of taking an accumulative learning bilateral branch network BBN as a backbone, a pyramid convolution module, a strip-shaped pooling module and a seesaw 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 an inverse sampling mechanismr,yr) As input to the rebalancing branch, the sampling probability P of class m is first calculated by the inverse samplerm:
wherein ,Nm、NjNumber of samples of class m and class j, respectively, NmaxIs the number of all samples and C is the total number of categories.
2) And adjusting a parameter alpha to control the weight and classification loss occupied by the characteristics of the two branches by accumulating a learning strategy, transferring the learning key points of the bilateral branches, and focusing the sand wave side scanning image data with less training samples.
Firstly, learning a general mode, and then gradually paying attention to tail data; after the convolution learning branch and the rebalancing branch are processed, the characteristic vector f is output through the corresponding full-connection layerc and frThen the data is used as input to be sent to the accumulation learning; if the total number of training iterations is TmaxIf the current training iteration number is T, the adjustment parameter α is:
alpha controls the output feature vector f of two branchesc and frWeight of (2), feature vector α f after weightingcAnd (1-. alpha.) frIs sent to the classifier of the corresponding branchAnd a characterizerIn the method, output fusion is performed in an element addition mode, and a fused C-dimensional output prediction characteristic z is represented as:
the BBN loss function in the whole learning process is a seesaw loss function:
L=αLs(zc)+(1-α)Ls(zr) (12)
wherein ,zc,zr={z1,z2,...,zCIs category prediction, Ls(. cndot.) represents a seesaw loss function.
Further, the pyramid convolution module is introduced as follows:
(1) the pyramid convolution module is introduced into the cumulative learning bilateral branch network BBN. The convolutional learning branch and the rebalancing branch have the same structure and share parameters except the last residual block. Analyzing the input by using a Local multi-scale context aggregation module (Local multi-scale context aggregation module) and a Global multi-scale context aggregation module (Global multi-scale context aggregation module) so as to enrich the function of capturing the cumulative learning bilateral branch network characteristics;
(2) for each branch network, firstly extracting a feature map through a convolutional layer, increasing a network receptive field through a Residual Block (Residual Block), then performing a hole convolution (scaled conv) and a pyramid convolution (PyConv), fusing features captured by a plurality of convolutional kernel scales through the pyramid convolution, wherein the depth of a convolutional kernel is changed along each level to form full depth and connectivity, and obtaining the feature map of an initial dimension through an improved Residual Block and convolution again.
(3) And repeating the steps, and respectively carrying out multistage improvement on the convolution learning branch and the rebalancing branch to obtain the feature map. In the hole convolution process, the convolution scale is adjusted by increasing the hyperparametric 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 hyperparametric hole rate.
The introduction of the strip-shaped pooling module is as follows:
(1) introducing a strip pooling module to an accumulative learning bilateral branch network BBN, and utilizing a narrow pooling kernel to better fit and identify a sand wave form; set to the dimension fh×fwInput feature map ofc 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) establishing an improved strip-shaped pooling strategy, and extracting effective characteristic expression symbols from the angles of row by row and column by column in parallel through equal weight combination of average pooling and maximum pooling:
Expanding two strip-shaped pooled output characteristics through convolution operation, and respectively converting the two output characteristics into the output characteristics with the size of fH×fWHorizontal strip-shaped expansionAnd vertical strip expansionBy y ═ yH+yVWill yH and yVCombining; by z ═ Scale (P)ijSigma (f (y)) calculation with global prior feature output to create feature capture suitable for the strip morphology of sand waves, assisted ocean bottom sand wave feature optimization and parameter learning, where Scale (·, ·) is element-by-element multiplication addition, sigma is the activation function, and f is 1 × 1 convolution.
The introduction of the seesaw loss module is as follows:
(1) introducing seesaw loss to an accumulated learning bilateral branch network BBN; introducing a distribution function mu (k) independent of the input samples and a smoothing parameter eta, and optimizing the distribution function mu (k) into y' ((1-eta)) y + eta mu (k) through sample labeling regularization (LSR); dynamically rebalancing the positive and negative sample gradients of each class according to the seesaw loss through the sample number ratio in the training process:
wherein ,ym' is to artificial label ymThe class label after the LSR is processed, so that the method has stronger fault tolerance to sample marking errors; z is a radical ofm={z1,z2,...,zCThe category is predicted;
for the class m samples, the negative sample gradient applied to the class n is:
wherein ,SmnAs an adjustable balance factor between different types of samples, adjusting the punishment applied to the nth type by the mth type; by a mitigation factor MmnAnd a compensation factor CmnDetermination of Smn=MmnCmn;
When class m occurs more frequently than class n, the mitigation factor is:
according to the ratio N of the number of samples between class N and class mn/NmReduce the penalty on the tail class n; the exponent p is a hyper-parameter of the adaptation amplitude;
if the prediction probability of class n is greater than class m, the back-off factor is:
penalty increase for class n (σ)n/σm)qMultiplying, wherein q is a hyper-parameter of the control proportion; cmnWhen 1, CmnApplying only the mitigation factor Mmn(ii) a Thereby avoiding misclassification of class n due to reduced overwhelming penalty for class n.
The invention introduces the improved process analysis of the pyramid convolution module, the strip-shaped pooling module and the seesaw loss module:
(1) introducing a pyramid convolution module:
for the feature extraction of the convolutional neural network, the global information of the image cannot be effectively utilized due to the limitation of the receptive field, and a discontinuous phenomenon may be caused during the extraction of the target feature. Seabed sand waves are a product of combined action of seabed and tidal current, bed slope factors, sediment particle sizes and material types can all affect the seabed sand waves, in a complex scene such as seabed sand wave identification, the receptive field is not enough to capture the correlation of different positions in the scene, useful details can be lost, and sand wave identification can be misjudged. The invention introduces a pyramid convolution module to the cumulative 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 capture function of the wide global information enrichment cumulative learning bilateral branch network; and further optimizing the pyramid convolution module, increasing the receptive field of the network by combining with the cavity convolution, and enhancing the characteristics by the optimization module combining the cavity convolution and the space pyramid convolution, thereby realizing the improvement of the sand wave characteristic identification and extraction capability and effectively obtaining the global prior information of the seabed sand waves.
(2) Introducing a strip-shaped pooling module:
the use of a conventional large square pooled nuclear window for a bottom sand wave in the form of a long strip will inevitably contain contamination information from irrelevant areas. The invention introduces a strip pooling module to the accumulation learning bilateral branch convolution network, better fits and identifies the form of the sand waves by using a narrow pooling kernel, avoids the interference of an irrelevant area on the identification of the sand waves at the seabed, and avoids the influence of noise on the identification of the sand waves at the seabed to a certain extent. And the longer pooling core is convenient for capturing the relation among the sand wave stripes in the isolated area, and reduces the probability that the sand wave is misjudged as the background. The use of such long and narrow pooled cores can be suitable for identification of ocean bottom sand waves in side-scan sonar data, while aggregating information in both global and local context.
(3) Introducing a seesaw loss module:
and (4) performing cross entropy loss calculation by adopting Softmax on the cumulative learning bilateral branch network. However, in subsea sand identification background samples are exponentially more dominant than sand samples, which are likely to act as negative samples of the sand class. The learning process of the classifier produces a bias. Objects of the sand category are more likely to be misclassified as background categories. According to the method, seesaw loss is introduced to replace an original Softmax cross entropy loss function, the ratio of the number of accumulated training samples among different classes and 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 relieved through the synergistic action of the relieving factor and the compensating factor, and the misjudgment of the background sample as the submarine sand wave caused by the lightening punishment is avoided. In addition, the geometric form of the seabed sand waves is complex and changeable, and the requirements on the speciality and the detail of manual marking are high. According to the method, Seesaw Loss is optimized through sample label regularization, noise is added to the labels to realize the constraint of an accumulative learning bilateral branch network model, the problem of overfitting caused by insufficient submarine sand wave labels is solved, meanwhile, the generalization performance of the network can be effectively improved, and the network is enabled to be suitable for sand wave identification application in complex submarine landforms and different sea areas.
A side-scan sonar undersea sand wave detection method based on an improved convolutional neural network comprises S6, in addition to the side-scan sonar undersea sand wave identification method S1-S5: analyzing the submarine sand wave form, drawing a sand wave form by extracting an echo intensity form of the side-scan sonar image data, analyzing from a sand wave section angle, and researching the geometric form analysis of the sand wave; the method specifically comprises the following steps:
(1) the collected echo intensity waveform is used as a search area and introduced into a template to match and position the sand wave, and the process is to move the template on the search area and calculate the similarity between the template and a target waveform;
(2) describing and analyzing the sand waves by positioning peaks and valleys in the sounding profile by using an envelope demodulation method;
(3) using zero-crossing rate to check the rate of change of the sign of the sand wave waveform spectrum, namely the number of times of positive direction negative direction or negative direction in a given period;
(4) the cross-correlation coefficient is used for exploring the remarkable characteristics of the sand wave by analyzing the upper envelope and the lower envelope of the sand wave profile waveform:
wherein ,respectively representing the upper and lower echo intensitiesEnveloping;andare respectivelyAndthe variance of (a);
(5) setting the wave crest as the highest point upward along the sand wave, the wavelength S of the sand waveLIs defined as the horizontal distance from trough to trough, the wave height SHIs 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 SuProjection length of descending slope is Sd(ii) a The method specifically comprises the following steps:
1) actual wavelength S of sand waveLThe estimation of (A) is:
wherein ,is the degree of echo intensity occupied by the sand waves in the side-scan sonar image,representing the maximum value of the number of single-side echo intensity points in the side-scan sonar image, RHRepresenting the actual unilateral operation range of the side-scan sonar at the sea bottom;
2) analyzing the balance relation between the observed sand wave height and the observed wavelength by using a formula proposed by Fleming in field investigation, and estimating the sand wave height value as follows:
wherein ,M1Is an average wavelength coefficient, f1Is the average wavelength index;
wherein ,M2Is the maximum wavelength coefficient, f2Is the maximum wavelength index.
3) The projection length S of the ascending slope recorded in the process of matching and positioning the wave crest of the templateuAnd the projection length S of the descending slopedThe asymmetric morphology index R is estimated by the formula:
R=Su/Sd (23)。
the detection method comprises the following steps:
further, the step (1) specifically includes:
1) using a set of Gaussian templates TmTo simulate the fluctuation characteristics of sand waves:
wherein ,μtIs a mathematical expectation of a Gaussian curve, udControlling the moving step length of the template on the abscissa; sigmatIs the variance of a Gaussian curve, controlling the width of the curve, wdControlling the discrete degree of the template; btDetermines the starting point, s, of the Gaussian curvedControlling the moving step length of the template on the ordinate;
2) by root mean square errorMeasuring the similarity between the echo intensity waveform and the template to screen out the bestMatching:
wherein , and Tm=[t1,t2,tm,…tM]1×MThe jth sample data of the ith is used for a part of template matching and a template vector, and M is input sonar dataLength of (d);
3) according to Taylor formula ln (1+ x) ═ x-x2/2+x3/3+....+(-1)nxn/n and taking reciprocal of the root mean square error, carrying out logarithm treatment, balancing the weight between the cross correlation coefficient and the root mean square error, and providing a morphology similarity comprehensive template matching criterion which combines the root mean square error and the cross correlation coefficient based on a Taylor formula:
the most ideal match result is the root mean square error between the target waveform and the templateVery small, cross-correlation coefficientIs very large, therefore at first constructThen passes through a logarithmic functionFor root mean square errorAnd (4) carrying out constraint to balance the capability of the root mean square error and the cross-correlation coefficient in the template matching algorithm.
Furthermore, in the step (2), the upper and lower envelopes of the sand wave profile are formed into a reconstruction boundary containing the maximum and minimum signals of the echo intensity. The geomorphology of the sand wave can generate a vibration signal with lower echo intensity frequency, the amplitude change and the direction of the sand wave signal are reflected by the extreme value of the waveform in the envelope demodulation process, 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 more clearly obtained by extracting the waveform envelope, so that the noise influence can be eliminated, and meanwhile, the sampling rate can be reduced and the pressure of a system storage space can be reduced.
Further, in the step (3), a zero-crossing rate is used to provide an indirect clue of the sand wave frequency, and the topological structure in the seabed sand wave is identified by calculating the number of times of zero crossing of the echo intensity, wherein the zero-crossing rate Z' is expressed as:
wherein ,andrespectively the j sample of the ipning sonar signalEcho intensities at the K and (K-1), K being the echo intensityThe length of (a) of (b),is thatMean value of (g), sgn [ ·]Is a sign function.
The invention has the advantages and beneficial effects that:
the method is mainly used for carrying out seabed sand wave identification based on deep learning and is applied to seabed sand wave detection by combining waveform matching analysis. Deep semantic information of a seabed sand wave image is extracted based on an accumulative learning bilateral branch network, a pyramid convolution module and strip pooling are introduced according to the self characteristics of the seabed sand wave to perform characteristic self-adaptive optimization and global information enhancement, the problems that the number distribution of seabed sand wave samples is unbalanced, the morphological characteristics of sand waves in a complex seabed environment are various, and the interference of ocean reverberation noise is serious are effectively solved, the seabed sand wave is positioned by using echo intensity as a research object through waveform matching, the morphological characteristics of the seabed sand wave such as wavelength, wave height and asymmetric index and the migration trend of the seabed sand wave are analyzed, and therefore a foundation is laid for further investigation and analysis of the seabed sand wave.
The morphological parameters of the seabed sand wave such as wavelength, wave height, asymmetric index and the like are the morphology and topological indexes of sand wave dynamics, are closely related to the tidal current size and direction of the seabed, sediment types, sand wave migration, ocean engineering construction of seabed pipelines, oil platforms and the like, and are essential parameters for sand wave research. The invention promotes the development and application of the deep learning method in marine geological survey, further carries out statistical analysis on the morphological parameters of the sand waves and predicts the migration of the sand waves.
Drawings
FIG. 1 is an overall flow chart of the present invention.
Fig. 2 is the AUV trace and manually labeled sand area of example 1.
Fig. 3 is an image of the ocean bottom sand wave to be recognized in example 1.
Fig. 4 is a block diagram of a subsea sand wave identification network in example 1.
Fig. 5 is a diagram of the pyramid convolution structure in example 1.
FIG. 6 is a diagram of the striped pooling scheme in example 1.
Fig. 7 is a bottom sand wave image recognition visualization of the section in example 1.
Fig. 8 is a sand wave profile waveform detection diagram in example 1.
Fig. 9 is a sand envelope diagram in example 1.
Fig. 10 is a background envelope diagram in example 1.
Fig. 11 is a sand waveform and zero crossing point profile of example 1.
FIG. 12 is a statistical plot of the zero crossing rate 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 profile parameter plot for example 1.
Fig. 15 is a wavelength distribution map of sand waves in example 1.
Fig. 16 is a sand wave asymmetric exponential distribution map in example 1.
Detailed Description
In order to make the objects, embodiments and advantages of the present invention clearer, the present invention is further described in detail below by way of specific examples with reference to the accompanying drawings.
Example 1: submarine sand waves of the gulf of Qingdao, China in 12 months in 2019 are used as objects for identification and morphological analysis.
The specific flowchart of this embodiment is shown in fig. 1.
In this embodiment, an AUV system (as shown in fig. 2) deployed from the central Qingdao gulf in 2019 and 12 months in the year shown in fig. 3 is specifically used to perform seabed detection sand wave picture as a picture for identification and morphological analysis.
The following steps should be described in detail with reference to the accompanying drawings and specific results and should be only steps outlined in the summary.
Firstly, seabed detection is carried out by depending on an Autonomous Underwater Vehicle (AUV), and the Autonomous Underwater Vehicle is mainly configured as a side scan sonar (Marine Sonic Sea Sca) and a water surfaceLower video camera (deep Power)&Light HD Multi SeaCam), Doppler velocimeter (DVL Teledyne RDI), processor (Intel atom eBOX530-820-FL), inertial navigation unit (KVH 1750IMU), digital compass (OceanServer OS5000), depth meter (measurementSpecial services 300), and underwater lighting (deep Power)&Light SLS-5100), acoustic communication (Teledyne BenthosATM-90), and the like, wherein the acquisition task of detection data is mainly undertaken by mainly utilizing a side scan sonar and an underwater video camera. And obtaining the longitude l of the underwater vehicle1Latitude l2Speed ofCourse h, pitch angle p, roll angle r and raw echo intensity data X. The raw echo intensity data is pre-processed and mapped into a side scan sonar image, and defined as I, and the size is H multiplied by W, wherein H is height and W is width.
Step two, determining the high H by side-scan sonar detectionaSonic slope distance RsSingle sided detection range RhConcretely, set forth is Ha=10m,Rs=200m,Rh199.75 m. Labeling the side-scan sonar image I by using labelme, taking the existence of sand waves and the absence of sand waves as corresponding labels y, and overlapping and partitioning the side-scan sonar image I according to a longitudinal interval e ping and a transverse interval f echo intensity points to form an image block u with the size k ij1, (H-k)/e + 1, j 1, (W-k)/f + 1, specifically 100 echo intensity points apart, with an image block of size 224 x 224 superimposed as u, training set: and (3) test set: the verification set is 0.6:0.2: 0.2.
Thirdly, carrying out data preprocessing of blind area removal, noise elimination, intensity correction, speed correction and time-varying gain correction on the original echo intensity data X to obtain a side-scan sonar image I, wherein the preprocessing work comprises the following steps:
(1) using Ensemble Empirical Mode Decomposition (EEMD) to perform adaptive filtering on echo intensity to obtain echo intensity data
(2) Calculating correction proportionality coefficient by speed correction to obtain echo intensity dataSetting the range distance delta h of the side-scan sonar image corresponding to the ith ping dataiExpressed as:
wherein i ∈ [1, H)],ΔtiThe time difference between the reception of the fourth ping data and the reception of the (i +1) th ping data for the sonar transducer transmission,is the instantaneous speed of the AUV at the eighth ping,is the instantaneous cruising speed of the AUV at the i +1 th ping, W is the width of the echo intensity in the ith ping,is the actual sailing distance of AUV, W/2RsThe mapping coefficient, delta t, between the width of the side scan sonar image and the actual detection range of the sea bottom perpendicular to the navigation track direction is showniIs the time difference between the ith ping and the (i +1) th ping in which the side scan sonar receives the echo.
(3) Through the time-varying gain correction method, the partial energy loss caused by the propagation of the sound wave emitted by the side-scan sonar in the seawater is compensated:
wherein ,xijIs the original echo intensity XiThe jth echo intensity value of (a) is,after filtering and velocity correctionThe intensity value of the echo of (a),are the echo intensity values after time varying gain correction. A is the echo intensity correction coefficient, and B is the echo intensity correction offset.
4. The side-scan sonar subsea sand wave identification method of claim 3, wherein step (2) comprises:
1) adding white noise of a given amplitude to the echo intensity;
2) performing Empirical Mode Decomposition (EMD) on the signal with the echo intensity and the added white noise to obtain an Intrinsic Mode Function (IMF) component and a residual component;
3) this operation was repeated using a given number of trials; the amount of IMF was constant in each experiment; all of the tests can be written as,
wherein ,XiIs the ith original signal in the echo intensity; n issIs white noise added in the s trial; IMFspIs the p-th IMF module in the s-th experiment; r issIs the remaining component in the s trial;
4) calculating the integral average value of all tests, and decomposing the calculated echo intensity of sonar by using the set empirical modeiThe final decomposition result of (a) is expressed as:
wherein S is the number of trials and r is the residual component of EEMD;
5) the first batch p is selected1IMF and denoising EEMD through a general threshold formula, as follows:
wherein ,σpIs the standard deviation of noise, N is IMFpLength of (d);
in determining the general threshold TpThen, selecting a soft threshold function to filter the noise component coefficient in the high-frequency IMF;
6) finally, the EEMD denoising method comprises the following steps:
the echo intensity data is subjected to noise elimination to obtainHistogram equalization is used as an intensity correction to further highlight target features.
Fourthly, constructing a side-scan sonar submarine sand wave identification network based on a convolutional neural network, wherein a block diagram of the submarine sand wave identification network is shown in fig. 4, and specifically comprises the following steps:
(1) a cumulative learning Bilateral Branch Network (BBN) is used as a backbone network for submarine sand wave identification, and comprises a convolution learning branch and a rebalance branch. For the training sample imbalance problem, the convolution learning branch uniformly samples the training sample (u)c,yc) Obtaining training samples (u) using an inverse sampling mechanismr,yr) As input to the rebalance branch, the learning attention is focused on the sand wave side scan image data with a small number of training samples;
1) obtaining samples (u) using an inverse sampling mechanismr,yr) As input to the rebalancing branch, the sampling probability P of class m is first calculated by the inverse samplerm:
wherein ,Nm、NjNumber of samples of class m and class j, respectively, NmaxIs the number of all samples and C is the total number of categories.
2) And adjusting a parameter alpha to control the weight and classification loss occupied by the characteristics of the two branches by accumulating a learning strategy, transferring the learning key points of the bilateral branches, and focusing the sand wave side scanning image data with less training samples.
Firstly, learning a general mode, and then gradually paying attention to tail data; after the convolution learning branch and the rebalancing branch are processed, the characteristic vector f is output through the corresponding full-connection layerc and frThen the data is used as input to be sent to the accumulation learning; if the total number of training iterations is TmaxIf the current training iteration number is T, the adjustment parameter α is:
alpha controls the output feature vector f of two branchesc and frWeight of (2), feature vector α f after weightingcAnd (1-. alpha.) frIs sent to the classifier of the corresponding branchAnd a characterizerIn the method, output fusion is performed in an element addition mode, and a fused C-dimensional output prediction characteristic z is represented as:
the BBN loss function in the whole learning process is a seesaw loss function:
L=αLs(zc)+(1-α)Ls(zr) (12)
wherein ,zc,zr={z1,z2,...,zCIs category prediction, Ls(. cndot.) represents a seesaw loss function.
(2) On the basis of taking an accumulative learning bilateral branch network BBN as a backbone, a pyramid convolution module, a strip-shaped pooling module and a seesaw loss module are introduced to jointly construct a submarine sand wave identification network.
(2) The method is characterized in that a pyramid convolution module, a bar-shaped pooling module and a seesaw loss are introduced to form a submarine sand wave identification network on the basis of taking an accumulative learning bilateral branch network as a backbone:
(1) the pyramid convolution module is introduced into the cumulative learning bilateral branch network BBN. The convolutional learning branch and the rebalancing branch have the same structure and share parameters except the last residual block. Analyzing the input by using a Local multi-scale context aggregation module (Local multi-scale context aggregation module) and a Global multi-scale context aggregation module (Global multi-scale context aggregation module) so as to enrich the function of capturing the cumulative learning bilateral branch network characteristics;
(2) for each branch network, firstly extracting a feature map through a convolutional layer, increasing a network receptive field through a Residual Block (Residual Block), then performing a hole convolution (scaled conv) and a pyramid convolution (PyConv), fusing features captured by a plurality of convolutional kernel scales through the pyramid convolution, wherein the depth of a convolutional kernel is changed along each level to form full depth and connectivity, and obtaining the feature map of an initial dimension through an improved Residual Block and convolution again.
(3) And repeating the steps, and respectively carrying out multistage improvement on the convolution learning branch and the rebalancing branch to obtain the feature map. In the hole convolution process, the convolution scale is adjusted by increasing the hyperparametric 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 hyperparametric hole rate.
The introduction of the strip-shaped pooling module is as follows:
(1) introducing a strip pooling module to an accumulative learning bilateral branch network BBN, and utilizing a narrow pooling kernel to better fit and identify a sand wave form; set to the dimension fh×fwInput feature map ofc 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) establishing an improved strip-shaped pooling strategy, and extracting effective characteristic expression symbols from the angles of row by row and column by column in parallel through equal weight combination of average pooling and maximum pooling:
Expanding two strip-shaped pooled output characteristics through convolution operation, and respectively converting the two output characteristics into the output characteristics with the size of fH×fWHorizontal strip-shaped expansionAnd vertical strip expansionBy y ═ yH+yVWill yH and yVCombining; by z ═ Scale (P)ijσ (f (y)) computing with global prior feature outputAnd (3) capturing the characteristics of the strip shape suitable for the sand wave, and assisting the optimization of the characteristics of the sand wave and parameter learning of the seabed, wherein Scale (·,) is multiplied and added element by element, σ is an activation function, and f is 1 × 1 convolution.
The introduction of the seesaw loss module is as follows:
(1) introducing seesaw loss to an accumulated learning bilateral branch network BBN; introducing a distribution function mu (k) independent of input samples and a Smoothing parameter eta, and optimizing the distribution function mu (k) into y' ((1-eta)) y + eta mu (k) through sample Smoothing Regularization (LSR); dynamically rebalancing the positive and negative sample gradients of each class according to the seesaw loss through the sample number ratio in the training process:
wherein ,ym' is to artificial label ymThe class label after the LSR is processed, so that the method has stronger fault tolerance to sample marking errors; z is a radical ofm={z1,z2,...,zCThe category is predicted;
for the class m samples, the negative sample gradient applied to the class n is:
wherein ,SmnAs an adjustable balance factor between different types of samples, adjusting the punishment applied to the nth type by the mth type; by a mitigation factor MmnAnd a compensation factor CmnDetermination of Smn=MmnCmn;
When class m occurs more frequently than class n, the mitigation factor is:
according to samples between class n and class mNumber ratio Nn/NmReduce the penalty on the tail class n; the exponent p is a hyper-parameter of the adaptation amplitude;
if the prediction probability of class n is greater than class m, the back-off factor is:
penalty increase for class n (σ)n/σm)qMultiplying, wherein q is a hyper-parameter of the control proportion; cmnWhen 1, CmnApplying only the mitigation factor Mmn(ii) a Thereby avoiding misclassification of class n due to reduced overwhelming penalty for class n.
(2) On the basis of taking an accumulative learning bilateral branch network BBN as a backbone, a pyramid convolution module, a strip-shaped pooling module and a seesaw loss module are introduced to jointly construct a submarine sand wave identification network.
1) The pyramid convolution module is introduced into the cumulative learning bilateral branch network BBN. The convolutional learning branch and the rebalancing branch have the same structure and share parameters except the last residual block. Analyzing the input by using a Local multi-scale context aggregation module (Local multi-scale context aggregation module) and a Global multi-scale context aggregation module (Global multi-scale context aggregation module) so as to enrich the function of capturing the cumulative learning bilateral branch network characteristics;
2) for each branch network, firstly extracting a feature map through a convolutional layer, increasing a network receptive field through a Residual Block (Residual Block), then performing a hole convolution (scaled conv) and a pyramid convolution (PyConv), fusing features captured by a plurality of convolutional kernel scales through the pyramid convolution, wherein the depth of a convolutional kernel is changed along each level to form full depth and connectivity, and obtaining the feature map of an initial dimension through an improved Residual Block and convolution again.
3) And repeating the steps, and respectively carrying out multistage improvement on the convolution learning branch and the rebalancing branch to obtain the feature map. In the hole convolution process, the convolution scale is adjusted by increasing the hyperparametric 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 hyperparametric hole rate.
The introduction of the strip-shaped pooling module is as follows:
1) introducing a strip pooling module to an accumulative learning bilateral branch network BBN, and utilizing a narrow pooling kernel to better fit and identify a sand wave form; set to the dimension fh×fwInput feature map ofc 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) establishing an improved strip-shaped pooling strategy, and extracting effective characteristic expression symbols from the angles of row by row and column by column in parallel through equal weight combination of average pooling and maximum pooling:
Expanding two strip-shaped pooled output characteristics through convolution operation, and respectively converting the two output characteristics into the output characteristics with the size of fH×fWHorizontal strip-shaped expansionAnd vertical strip expansionBy y ═ yH+yVWill yH and yVCombining; by z ═ Scale (P)ijSigma (f (y)) calculation with global prior feature output to create feature capture suitable for the strip morphology of sand waves, assisted ocean bottom sand wave feature optimization and parameter learning, where Scale (·, ·) is element-by-element multiplication addition, sigma is the activation function, and f is 1 × 1 convolution.
The introduction of the seesaw loss module is as follows:
1) introducing seesaw loss to an accumulated learning bilateral branch network BBN; introducing a distribution function mu (k) independent of input samples and a Smoothing parameter eta, and optimizing the distribution function mu (k) into y' ((1-eta)) y + eta mu (k) through sample Smoothing Regularization (LSR); dynamically rebalancing the positive and negative sample gradients of each class according to the seesaw loss through the sample number ratio in the training process:
wherein ,ym' is to artificial label ymThe class label after the LSR is processed, so that the method has stronger fault tolerance to sample marking errors; z is a radical ofm={z1,z2,...,zCThe category is predicted;
for the class m samples, the negative sample gradient applied to the class n is:
wherein ,SmnAs an adjustable balance factor between different types of samples, adjusting the punishment applied to the nth type by the mth type; by a mitigation factor MmnAnd a compensation factor CmnDetermination of Smn=MmnCmn;
When class m occurs more frequently than class n, the mitigation factor is:
according to the ratio N of the number of samples between class N and class mn/NmReduce the penalty on the tail class n; the exponent p is a hyper-parameter of the adaptation amplitude;
if the prediction probability of class n is greater than class m, the back-off factor is:
penalty increase for class n (σ)n/σm)qMultiplying, wherein q is a hyper-parameter of the control proportion; cmnWhen 1, CmnApplying only the mitigation factor Mmn(ii) a Thereby avoiding misclassification of class n due to reduced overwhelming penalty for class n.
The visualization of the experimental result of the submarine sand wave recognition is shown in fig. 7, and part of submarine sand wave image recognition effect is shown. Fig. (a) is the original sonar image from which the water body was removed, and fig. (b) is the group Truth of fig. (a). Fig. (c) and (d) are raster image visualizations of Resnet50 and the side-scan sonar image recognition result of the present algorithm, respectively. The graphs (e) and (f) are the grid projection of the Resnet50 and the recognition result of the algorithm of the present invention into the original sonar image (c), respectively. By comparing the Resnet50 algorithm with the recognition result of the invention, it can be obviously seen that the invention realizes better recognition effect. In the recognition effect of 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 for an area where the sand wave is gentle.
The sand wave waveform analysis method comprises the following steps:
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 morphology analysis of the sand waves comprises the following steps:
(1) the collected echo intensity waveform is used as a search area and is introduced into a template matching positioning sand wave, and the process is to move the template on the search area and calculate the similarity between the template and a target waveform.
1) Using a set of Gaussian templates TmThe fluctuation characteristics of the sand waves are simulated, and the following formula is shown:
wherein ,μtIs a mathematical expectation of a Gaussian function, udControlling the moving step length of the template on the abscissa; sigmatIs the variance of a Gaussian curve, controlling the width of the curve, wdControlling the discrete degree of the template; btDetermines the starting point, s, of the Gaussian curvedThe moving step length of the template on the ordinate is controlled.
2) By root mean square errorMeasuring the similarity between the echo intensity waveform and the template, and screening out the best match:
wherein , and Tm=[t1,t2,tm,…tM]1×MThe jth sample data of the ith ping is used for the template matching part and the template vector, and M is the input sonar dataLength of (d). The present invention attempts to compute the echo intensity waveform from the cross-correlation coefficientsAnd a template TmOverall similarity between them
3) The invention is based on the Taylor formula ln (1+ x) ═ x-x2/2+x3/3+....+(-1)nxn/n and taking reciprocal of the root mean square error, carrying out logarithm treatment, balancing the weight between the cross correlation coefficient and the root mean square error, and providing a morphology similarity comprehensive template matching criterion which combines the root mean square error and the cross correlation coefficient based on a Taylor formula:
by enhancing the evaluation of the spatial similarity, the Euclidean distance matching of the Gaussian template and the target waveform on the same curve length is considered, and the similarity on the geometric form is also considered. The most ideal match result is the root mean square error between the target waveform and the templateVery small, cross-correlation coefficientIs very big, therefore we firstly constructIt was found, however, that when the partial points in the template are close to or even coincident with the echo intensity waveform,the size of the space is very large, the evaluation capability of a matching algorithm on the space distance is too strong, and the cross-correlation coefficientThe effect of (A) is almost neglected, resulting in an imbalance between the twoA template is selected that results in partial coincidence but overall large differences in trends between the template and the echo intensity waveform. The item passes through a logarithmic functionFor root mean square errorAnd (4) carrying out constraint to balance the capability of the root mean square error and the cross-correlation coefficient in the template matching algorithm. In addition, when the echo intensity waveform has abnormal value, the root mean square errorIn the very large of the time,will be very small according to the formulaIt can be seen that the matching criteria of the project designWill approximate to and For which the time value is smallerRatio of change ofIs more sensitive per se, so that the logarithm operation pair can be effectively avoidedThe extrusion is simultaneously carried out on abnormal values with larger root mean square errorsSufficient care is taken.
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 figure that the form similarity comprehensive template matching criterion provided by the invention has good matching performance and can realize the matching of sand waves with different scales.
(2) The sand waves are delineated and analyzed by the location of peaks and valleys in the sounding profile using an envelope demodulation method. The upper and lower envelopes of the sand wave profile form a reconstruction boundary containing the maximum and minimum signals of the echo intensity. The geomorphology of the sand wave can generate a vibration signal with lower echo intensity frequency, the amplitude change and the direction of the sand wave signal are reflected by the extreme value of the waveform in the envelope demodulation process, 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 envelope is shown in fig. 9 and the background envelope is shown in fig. 10, where 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 more clearly obtained by extracting the waveform envelope, so that the noise influence can be eliminated, and meanwhile, the sampling rate can be reduced and the pressure of a system storage space can be reduced.
(3) The Zero Crossing Rate (ZCR) is used to check the rate of change of sign of the sand waveform spectrum, i.e. the number of times that the spectrum changes from positive to negative or negative in a given period. The zero crossing rate is used to provide an indirect clue to the sand frequency and identify the topology in the seabed sand by counting the number of times the echo intensity crosses zero, and the zero crossing rate Z' can be expressed as:
wherein ,andrespectively the j sample of the ipning sonar signalEcho intensities at the K and (K-1), K being the echo intensityThe length of (a) of (b),is thatMean value of (g), sgn [ ·]Is a sign function. The sand waveform and the zero crossing point distribution, as shown in fig. 11, (a) is the case where there is no sand wave, (b) (c) (d) is the case where there is a sand wave; the zero-crossing rate statistics of the waveform, as shown in fig. 12, statistical analysis is performed on the zero-crossing rates of the areas with and without sand waves in (a) and (b), respectively, to explore the zero-crossing rate distribution of the echo intensity waveforms with and without sand waves. In a simulation experiment, the research of the invention finds that in the sand wave data collected in the gulf of glue 12 months in 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 periodic rule. The zero-crossing rate of the background waveform is higher, and is between 0.4 and 0.6, the fluctuation rate is fast, the echo intensity looks disordered and is mainly represented as interference noise distribution. It can be seen that the zero crossing rate of the waveform without sand waves is similar to the ambient noise and is greater than the zero crossing rate of the sand waves. Therefore, a low zero crossing rate can be used as a significant feature for distinguishing the sand wave from the background waveform without the sand wave.
(4) The cross-correlation coefficient is used for exploring the remarkable characteristics of the sand wave by analyzing the upper envelope and the lower envelope of the sand wave profile waveform:
wherein ,representing the upper and lower envelopes of the echo intensity, respectively.Andare respectivelyAndthe variance of (c). Cross-correlation coefficient comparison is shown in fig. 13, where a) and (b) are the statistical data of the cross-correlation coefficients of the sand waveform and the background waveform, respectively, it can be found that the cross-correlation coefficient of the upper envelope and the lower envelope of the sand waveform is relatively large, and the two are in a synchronous fluctuation state. And 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 synchronization trend exists, and the envelope curve can show randomness, which is consistent with the fact that the envelope curve is mainly composed 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 the presence of sand waves is 90%.
(5) Setting the wave crest as the highest point upward along the sand wave, the wavelength S of the sand waveLDefined as the horizontal distance from trough to trough. Wave height SHIs defined as the vertical height difference between the crest and adjacent trough of a sand wave. And defining the projection length of the ascending slope as SuProjection length of descending slope is Sd. The sand profile parameter map is shown in fig. 14.
1) Actual wavelength S of sand waveLThe estimation of (A) is:
wherein ,is the degree of echo intensity occupied by the sand waves in the side-scan sonar image,side sweep of the representativeMaximum value of single-sided echo intensity point number in sonar image, RHRepresenting the actual single-sided operating range of the side-scan sonar at the seafloor. The parameters recorded by the form similarity comprehensive template matching criterion proposed 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 wavelengths of the simulated experimental region are mainly between 10m and 30m, with 25.9% of wavelengths greater than 10m and less than 15 m; 59.9% of the wavelength of more than 15m and less than 30 m; and 14.2% of the wavelength less than 10 m. We can find that 85.8% of the sand wavelengths are all greater than 10 m.
2) Analyzing the balance relation between the observed sand wave height and the observed wavelength by using a formula proposed by Fleming in field investigation, and estimating the sand wave height value as follows:
wherein ,M1Is an average wavelength coefficient, f1Is the average wavelength index.
wherein ,M2Is the maximum wavelength coefficient, f2Is the maximum wavelength index.
Statistical analysis shows that the wave length of sand waves in the Bay area of Guzhou is mainly more than 15m, the wave height is about 0.86m, and the upper limit of the wave height is mainly 2.23 m.
3) The projection length S of the ascending slope recorded in the process of matching and positioning the wave crest of the templateuAnd the projection length S of the descending slopedThe asymmetric morphology index R is estimated by the formula:
R=Su/Sd (29)
the asymmetric morphology index R was calculated and its distribution was mapped into fig. 16. We found that the fraction of sand waves with R < 0.5 westward shift was 63.4%. The proportion of R < 1.5 > 0.5 in the normal distribution is 26.4%. R > 1.5 accounts for 10.2% of the eastern part. Thus, we can find that most of the sand waves in the gulf of glue have a steep slope toward the west and a gentle slope toward the east, and conclude that the sand waves may have a tendency to migrate toward the west. Our AUVs were deployed a second time around the same research area in gulf of glue in 1 month 2021 and subsea sand wave data were collected over a duration of more than a year to verify the predicted direction of sand wave migration.
Geometric morphology analysis on sand waves:
the form of the sand wave reflects the power of the sea bottom, the sediment of the sea bottom, the size of the sand wave and the relative movement strength, and is a key factor for researching the sand wave of the sea bottom. We design the matching criterion of the form similarity comprehensive template to locate the sand waves. And extracting the upper envelope and the lower envelope of the waveform to analyze and obtain the main fluctuation direction of the waveform. And (4) carrying out zero-crossing rate analysis on the side-scan sonar data to find that a sand wave area exists. 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 sand waves. And acquiring morphological parameters such as the wavelength, the wave height, the asymmetric index and the like of the sand wave. The migration trend of the sand waves is obtained through statistical analysis.
Claims (10)
1. A side-scan sonar seabed sand wave detection method based on a multi-scale convolution and pooling strategy is characterized by comprising the following steps:
s1: collecting seabed side-scan sonar data and attitude positioning data;
s2: preprocessing the collected seabed side-scan sonar data of S1;
s3: constructing a submarine sand wave identification network based on an improved convolutional neural network;
s4: training the submarine sand wave recognition network obtained by S3;
s5: and (5) recognizing the side-scan sonar test data preprocessed by the S2 by using the seabed sand wave recognition network trained by the S4, and outputting a recognition result.
2. The side-scan sonar subsea sand wave detection method of claim 1, wherein, in S1,
(1) data are acquired by using a side scan sonar seabed detection means: the method comprises the longitude lon, the latitude lat, the speed v, the heading h, the pitch angle p, the roll angle r and original echo intensity data X of the underwater vehicleiH, i 1.. H; collectively mapping original echo intensity data X into a side-scan sonar image through preprocessing, defining the image as I, and setting the size as H multiplied by W, wherein H is height and W is width;
(2) high H determination by side-scan sonaraSonic slope distance RsSingle sided detection range Rh(ii) a Labeling the side-scan sonar image I, wherein the side-scan sonar image I is divided into two types of sand waves and sand-free waves, the corresponding label is y, and every s is arranged according to the longitudinal intervalvOne ping and one transverse interval per shThe echo intensity data points are overlapped and blocked to form an image block P with the size of l x lij,i=1,...,(H-l)/sv+1,j=1,...,(W-l)/sh+1, and occupying ratio lambda according to training set, test set and verification set respectively1、λ2、λ3Sample data (λ) is allocated1+λ2+λ3=1)。
3. The side-scan sonar subsea sand wave detection method of claim 1, wherein in S2, the preprocessing comprises:
(1) respectively carrying out data preprocessing of blind area removal, noise elimination, intensity correction, speed correction and time-varying gain correction on the side-scan sonar image I;
(2) using Ensemble Empirical Mode Decomposition (EEMD) to perform adaptive filtering on echo intensity to obtain echo intensity data
(3) Calculating correction proportionality coefficient by speed correction to obtain echo intensity dataSetting the range distance delta h of the side-scan sonar image corresponding to the ith ping dataiExpressed as:
wherein i ∈ [1, H)],ΔtiFor the time difference between the transmission and reception of the ith ping data and the reception of the (i +1) th ping data by the sonar transducer,is the instantaneous speed of the AUV navigating at the ith ping,is the instantaneous cruising speed of the AUV at the i +1 th ping, W is the width of the echo intensity in the ith ping,is the actual sailing distance of AUV, W/2RsThe mapping coefficient, delta t, between the width of the side scan sonar image and the actual detection range of the sea bottom perpendicular to the navigation track direction is showniIs the time difference between the ith ping and the (i +1) th ping where the side scan sonar receives the echo;
(4) through the time-varying gain correction method, the partial energy loss caused by the propagation of the sound wave emitted by the side-scan sonar in the seawater is compensated:
wherein ,xijIs the original echo intensity XiThe jth echo intensity value of (a) is,is the echo intensity value after filtering and speed correction,is the echo intensity value after time-varying gain correction; a is the echo intensity correction coefficient, and B is the echo intensity correction offset.
4. The side-scan sonar subsea sand wave detection method of claim 3, wherein step (2) comprises:
1) adding white noise of a given amplitude to the echo intensity;
2) performing Empirical Mode Decomposition (EMD) on the signal with the echo intensity and the added white noise to obtain an Intrinsic Mode Function (IMF) component and a residual component;
3) this operation was repeated using a given number of trials; the amount of IMF was constant in each experiment; all of the tests can be written as,
wherein ,XiIs the ith original signal in the echo intensity; n issIs white noise added in the s trial; IMFspIs the p-th IMF module in the s-th experiment; r issIs the remaining component in the s trial;
4) calculating the integral average value of all tests, and decomposing the calculated echo intensity of sonar by using the set empirical modeiThe final decomposition result of (a) is expressed as:
wherein S is the number of trials and r is the residual component of EEMD;
5) the first batch p is selected1IMF and denoising EEMD through a general threshold formula, as follows:
wherein ,σpIs the standard deviation of noise, N is IMFpLength of (d);
in determining the general threshold TpThen, selecting a soft threshold function to filter the noise component coefficient in the high-frequency IMF;
6) finally, the EEMD denoising method comprises the following steps:
5. The side-scan sonar subsea sand wave detection method of claim 1, wherein in S3, a subsea sand wave identification network based on an improved convolutional neural network is constructed, specifically:
(1) using an accumulative learning bilateral branch network BBN as a backbone network for submarine sand wave identification, wherein the backbone network comprises a convolution learning branch and a rebalance branch; for the training sample imbalance problem, the convolution learning branch uniformly samples the training sample (u)c,yc) Obtaining training samples (u) using an inverse sampling mechanismr,yr) As input to the rebalance branch, the learning attention is focused on the sand wave side scan image data with a small number of training samples;
(2) on the basis of taking an accumulative learning bilateral branch network BBN as a backbone, a pyramid convolution module, a strip-shaped pooling module and a seesaw loss module are introduced to jointly construct a submarine sand wave identification network.
6. The side-scan sonar subsea sand wave identification method of claim 5, wherein in step (1), the cumulative learning bilateral branch learning strategy specifically is:
1) obtaining samples (u) using an inverse sampling mechanismr,yr) As input to the rebalancing branch, the sampling probability P of class m is first calculated by the inverse samplerm:
wherein ,Nm、NjNumber of samples of class m and class j, respectively, NmaxIs the number of all samples, C is the total number of categories;
2) adjusting a parameter alpha to control the weight and classification loss occupied by the characteristics of the two branches through an accumulated learning strategy, transferring the learning key points of the bilateral branches, and focusing the sand wave side scanning image data with less training samples;
firstly, learning a general mode, and then gradually paying attention to tail data; after the convolution learning branch and the rebalancing branch are processed, the characteristic vector f is output through the corresponding full-connection layerc and frThen the data is used as input to be sent to the accumulation learning; if the total number of training iterations is TmaxIf the current training iteration number is T, the adjustment parameter α is:
alpha controls the output feature vector f of two branchesc and frWeight of (2), feature vector α f after weightingcAnd (1-. alpha.) frIs sent to the classifier of the corresponding branchAnd a characterizerIn the method, output fusion is performed in an element addition mode, and a fused C-dimensional output prediction characteristic z is represented as:
the BBN loss function in the whole learning process is a seesaw loss function:
L=αLs(zc)+(1-α)Ls(zr) (12)
wherein ,zc,zr={z1,z2,...,zCIs category prediction, Ls(. cndot.) represents a seesaw loss function.
7. The method for side-scan sonar seafloor sand wave detection of claim 5, wherein the pyramid convolution module is introduced as follows:
(1) introducing a pyramid convolution module into an accumulative learning bilateral branch network BBN; the convolution learning branch and the rebalancing branch have the same structure and share parameters except the last residual block; analyzing the input by using a local multi-scale context aggregation module and a global multi-scale context aggregation module so as to enrich the function of capturing the accumulated learning bilateral branch network characteristics;
(2) for each branch network, firstly extracting a feature map through a convolution layer, increasing a network receptive field through a residual block, then performing void convolution and pyramid convolution, fusing features captured by multiple convolution kernel scales through pyramid convolution, wherein the depth of a convolution kernel is changed along each level to form full depth and connectivity, and obtaining the feature map of an initial dimension again through improved residual blocks and convolution;
(3) and repeating the steps, and respectively carrying out multistage improvement on the convolution learning branch and the rebalancing branch to obtain the feature map. In the hole convolution process, the convolution scale is adjusted by increasing the hyperparametric hole rate on the basis of standard convolution, and the size of an equivalent convolution kernel after the hole convolution is introduced is k' ═ k + (k-1) x (d-1), wherein k is the size of an original convolution kernel, and d is the hyperparametric hole rate;
the introduction of the strip-shaped pooling module is as follows:
(1) introducing a strip pooling module to an accumulative learning bilateral branch network BBN, and utilizing a narrow pooling kernel to better fit and identify a sand wave form; set to the dimension fh×fwInput feature map ofc 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) establishing an improved strip-shaped pooling strategy, and extracting effective characteristic expression symbols from the angles of row by row and column by column in parallel through equal weight combination of average pooling and maximum pooling:
Expanding two strip-shaped pooled output characteristics through convolution operation, and respectively converting the two output characteristics into the output characteristics with the size of fH×fWHorizontal strip-shaped expansionAnd vertical strip expansionBy y ═ yH+yVWill yH and yVCombining; by z ═ Scale (P)ijSigma (f (y)) computing has global prior feature output to create feature capture suitable for the strip shape of the sand wave, and assists seabed sand wave feature optimization and parameter learning, wherein Scale (·, ·) is multiplied and added element by element, sigma is an activation function, and f is 1 × 1 convolution;
the introduction of the seesaw loss module is as follows:
(1) introducing seesaw loss to an accumulated learning bilateral branch network BBN; introducing a distribution function μ (k) independent of the input samples and a smoothing parameter η, and optimizing by sample label regularization LSR to y' ═ 1- η) y + η μ (k); dynamically rebalancing the positive and negative sample gradients of each class according to the seesaw loss through the sample number ratio in the training process:
wherein ,ym' is to artificial label ymThe class label after the LSR is processed, so that the method has stronger fault tolerance to sample marking errors; z is a radical ofm={z1,z2,...,zCThe category is predicted;
for the class m samples, the negative sample gradient applied to the class n is:
wherein ,SmnAs an adjustable balance factor between different types of samples, adjusting the punishment applied to the nth type by the mth type; by a mitigation factor MmnAnd a compensation factor CmnDetermination of Smn=MmnCmn;
When class m occurs more frequently than class n, the mitigation factor is:
according to the ratio N of the number of samples between class N and class mn/NmReduce the penalty on the tail class n; the exponent p is a hyper-parameter of the adaptation amplitude;
if the prediction probability of class n is greater than class m, the back-off factor is:
penalty increase for class n (σ)n/σm)qMultiplying, wherein q is a hyper-parameter of the control proportion; cmnWhen 1, CmnApplying only the mitigation factor Mmn(ii) a Thereby avoiding misclassification of class n due to reduced overwhelming penalty for class n.
8. The side-scan sonar subsea sand wave detection method of claim 1, comprising, in addition to the side-scan sonar subsea sand wave identification method of claim 1, steps S1-S5, S6: the analysis of seabed sand wave form, drawing the sand wave form by extracting the echo intensity wave form of the side scan sonar image data, analyzing from the sand wave section angle, researching the geometric form of the sand wave, specifically comprising:
(1) segmenting the acquired echo intensity to form sub-waveforms serving as search areas, introducing templates to match and position the sand waves, moving the templates on the search areas, and calculating the similarity between the templates and target waveforms;
(2) drawing an envelope curve for positioning the wave crest and the wave trough in the echo intensity profile by an envelope demodulation method;
(3) using zero crossing rate ZCR to check the frequency of the change of the sign of the sand wave waveform spectrum, namely the number of times of positive direction negative direction or negative direction in a given period;
(4) the cross-correlation coefficient is used for exploring the remarkable 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 and lower envelopes of the echo intensity;andare respectivelyAndthe variance of (a);
(5) setting the sand wave wavelength SLThe horizontal distance from wave trough to wave trough, wave height SHDefining 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 waveuProjection length of descending slope is SdThe method specifically comprises the following steps:
1) actual wavelength S of sand waveLThe estimation is as follows:
wherein ,is the degree of echo intensity occupied by the sand waves in the side-scan sonar image,representing the maximum value of the number of single-side echo intensity points in the side-scan sonar image, RHRepresenting the actual unilateral operation range of the side-scan sonar at the sea bottom;
2) analyzing the balance relation between the observed sand wave height and the observed wavelength by using a formula proposed by Fleming in field investigation, and estimating the sand wave height value as follows:
wherein ,M1Is an average wavelength coefficient, f1Is the average wavelength index;
wherein ,M2Is the maximum wavelength coefficient, f2Is the maximum wavelength index;
3) the projection length S of the ascending slope recorded in the process of matching and positioning the wave crest of the templateuAnd the projection length S of the descending slopedThe asymmetric morphology index R is estimated by the formula:
R=Su/Sd (23)。
9. the side-scan sonar subsea sand wave detection method of claim 8, wherein step (1) specifically comprises:
1) using a set of Gaussian templates TmThe fluctuation characteristics of the sand waves are simulated, and the following formula is shown:
wherein ,μtIs a mathematical expectation of a Gaussian function, udControlling the moving step length of the template on the abscissa; sigmatIs the variance of a Gaussian function, controlling the width of the function, wdControlling the discrete degree of the template; btDetermining the starting point, s, of the Gaussian functiondControlling the moving step length of the template on the ordinate;
2) by root mean square errorMeasuring the similarity between the echo intensity waveform and the template, and screening out the best match:
wherein , and Tm=[t1,t2,tm,…tM]1×MThe jth sample data of ith ping is used for the sub-waveform and template vector of template matching, M is input echo intensity dataLength of (d);
3) according to Taylor formula ln (1+ x) ═ x-x2/2+x3/3+....+(-1)nxn/n and taking reciprocal of the root mean square error and carrying out logarithm treatment, and providing a morphology similarity comprehensive template matching criterion combining the root mean square error and the cross-correlation coefficient based on a Taylor formula:
10. The side-scan sonar seafloor sand wave detection method of claim 8, wherein in step (2), the upper and lower envelopes of the sand wave profile are formed into reconstructed boundaries containing maximum and minimum signals of echo intensities; in the step (3), an indirect clue of sand wave frequency is provided by using a zero crossing rate, and a topological structure in the seabed sand wave is identified by calculating the number of times of zero crossing of echo intensity, wherein the zero crossing rate Z' is expressed as:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110497412.0A CN113516626B (en) | 2021-05-07 | 2021-05-07 | Side-scan sonar submarine sand wave detection method based on multi-scale convolution and pooling strategy |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110497412.0A CN113516626B (en) | 2021-05-07 | 2021-05-07 | Side-scan sonar submarine sand wave detection method based on multi-scale convolution and pooling strategy |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113516626A true CN113516626A (en) | 2021-10-19 |
CN113516626B CN113516626B (en) | 2023-09-26 |
Family
ID=78063861
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110497412.0A Active CN113516626B (en) | 2021-05-07 | 2021-05-07 | Side-scan sonar submarine sand wave detection method based on multi-scale convolution and pooling strategy |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113516626B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114462445A (en) * | 2021-12-24 | 2022-05-10 | 扬州大学 | Feature extraction method for marine multidimensional asymmetric signal with given observation dimension |
CN115754107A (en) * | 2022-11-08 | 2023-03-07 | 福建省龙德新能源有限公司 | Automatic sampling analysis system and method for preparing lithium hexafluorophosphate |
CN117455910A (en) * | 2023-12-22 | 2024-01-26 | 广州金和精密机电设备有限公司 | Winding identification method and winding equipment based on machine vision |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103345759A (en) * | 2013-07-24 | 2013-10-09 | 国家***第二海洋研究所 | Accurate detection method for submarine large complex sandwave landforms |
CN111539314A (en) * | 2020-04-21 | 2020-08-14 | 上海海事大学 | Cloud and fog shielding-oriented sea surface target significance detection method |
-
2021
- 2021-05-07 CN CN202110497412.0A patent/CN113516626B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103345759A (en) * | 2013-07-24 | 2013-10-09 | 国家***第二海洋研究所 | Accurate detection method for submarine large complex sandwave landforms |
CN111539314A (en) * | 2020-04-21 | 2020-08-14 | 上海海事大学 | Cloud and fog shielding-oriented sea surface target significance detection method |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114462445A (en) * | 2021-12-24 | 2022-05-10 | 扬州大学 | Feature extraction method for marine multidimensional asymmetric signal with given observation dimension |
CN114462445B (en) * | 2021-12-24 | 2024-03-29 | 扬州大学 | Feature extraction method for ocean multidimensional asymmetric signals with given observation dimensions |
CN115754107A (en) * | 2022-11-08 | 2023-03-07 | 福建省龙德新能源有限公司 | Automatic sampling analysis system and method for preparing lithium hexafluorophosphate |
CN117455910A (en) * | 2023-12-22 | 2024-01-26 | 广州金和精密机电设备有限公司 | Winding identification method and winding equipment based on machine vision |
CN117455910B (en) * | 2023-12-22 | 2024-03-26 | 广州金和精密机电设备有限公司 | Winding identification method and winding equipment based on machine vision |
Also Published As
Publication number | Publication date |
---|---|
CN113516626B (en) | 2023-09-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109709603B (en) | Seismic horizon identification and tracking method and system | |
CN113516626B (en) | Side-scan sonar submarine sand wave detection method based on multi-scale convolution and pooling strategy | |
Reed et al. | The fusion of large scale classified side-scan sonar image mosaics | |
CN109100710A (en) | A kind of Underwater targets recognition based on convolutional neural networks | |
Połap et al. | Side-scan sonar analysis using ROI analysis and deep neural networks | |
CN112883564B (en) | Water body temperature prediction method and prediction system based on random forest | |
CN114595732B (en) | Radar radiation source sorting method based on depth clustering | |
Nasim et al. | Seismic facies analysis: a deep domain adaptation approach | |
CN115114949A (en) | Intelligent ship target identification method and system based on underwater acoustic signals | |
CN110675410A (en) | Side-scan sonar sunken ship target unsupervised detection method based on selective search algorithm | |
CN112613504A (en) | Sonar underwater target detection method | |
Leng et al. | A novel bathymetry signal photon extraction algorithm for photon-counting LiDAR based on adaptive elliptical neighborhood | |
CN117056680A (en) | Data noise reduction and signal detection method, device and system and storage medium | |
CN115547347A (en) | Whale acoustic signal identification method and system based on multi-scale time-frequency feature extraction | |
White et al. | More than a whistle: Automated detection of marine sound sources with a convolutional neural network | |
Ji et al. | Full-waveform classification and segmentation-based signal detection of single-wavelength bathymetric LiDAR | |
CN114282576A (en) | Radar signal modulation format identification method and device based on time-frequency analysis and denoising | |
CN116660996A (en) | Drifting type shallow sea local earth sound parameter prediction method based on deep learning | |
Alevizos et al. | Quantification of the fine-scale distribution of Mn-nodules: Insights from AUV multi-beam and optical imagery data fusion | |
Quintana et al. | Towards automatic recognition of mining targets using an autonomous robot | |
Sengupta et al. | Seagrassdetect: A novel method for the detection of seagrass from unlabelled underwater videos | |
Barbaresco et al. | Doppler spectrum segmentation of radar sea clutter by mean-shift and information geometry metric | |
CN115187855A (en) | Seabed substrate sonar image classification method | |
Wang et al. | Seafloor classification based on deep-sea multibeam data—Application to the southwest Indian Ridge at 50.47° E | |
Doray et al. | A geostatistical method for assessing biomassof tuna aggregations around moored fish aggregating devices with star acoustic surveys |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |