CN115187855A - Seabed substrate sonar image classification method - Google Patents

Seabed substrate sonar image classification method Download PDF

Info

Publication number
CN115187855A
CN115187855A CN202210905490.4A CN202210905490A CN115187855A CN 115187855 A CN115187855 A CN 115187855A CN 202210905490 A CN202210905490 A CN 202210905490A CN 115187855 A CN115187855 A CN 115187855A
Authority
CN
China
Prior art keywords
image
gray level
characteristic
scan sonar
degrees
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.)
Pending
Application number
CN202210905490.4A
Other languages
Chinese (zh)
Inventor
赵玉新
郑良锋
朱可心
邓雄
赵廷
吴昌哲
姜南
何永旭
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Engineering University
Original Assignee
Harbin Engineering University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Harbin Engineering University filed Critical Harbin Engineering University
Priority to CN202210905490.4A priority Critical patent/CN115187855A/en
Publication of CN115187855A publication Critical patent/CN115187855A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/05—Underwater scenes
    • 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
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/30—Noise filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Multimedia (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Molecular Biology (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Measurement Of Velocity Or Position Using Acoustic Or Ultrasonic Waves (AREA)

Abstract

The invention discloses a submarine substrate sonar image classification method, which comprises the steps of obtaining a side scan sonar image data set, wherein the data set comprises an image to be classified and an image with a true value label; calculating a gray level co-occurrence matrix, and performing statistical feature extraction on the side scan sonar image by using the gray level co-occurrence matrix; performing fractal dimension feature extraction on the side scan sonar image based on a pixel coverage method; extracting a channel energy characteristic value of the side scan sonar image; and performing joint representation on the extracted characteristic parameters of the statistical characteristic, the fractal dimension characteristic and the channel energy characteristic to form a joint characteristic, inputting the joint characteristic parameters of the image to be classified into GoogLeNet trained by using the image with the true value label to classify the seabed substrate, and outputting a network output result as the image substrate category. The method combines the advantages of various feature extraction methods, and improves the accuracy of classification of the seabed sediment by adopting the improved GoogLeNet.

Description

Seabed substrate sonar image classification method
Technical Field
The invention belongs to the field of submarine surveying and mapping and artificial intelligence, and relates to a submarine substrate sonar image classification method, in particular to an improved GoogLeNet submarine substrate sonar image classification method based on multiple features.
Background
In the field of ocean development aiming at military affairs and economy, investigation and research of seabed sediments are of great significance. At present, most of the seabed data are obtained through sonar telemetering and are mainly presented in the form of sonar images. In the classification process of the seabed bottom materials, feature extraction is a main link of sonar image processing. On one hand, the characteristics of the seabed sediment are better extracted to a certain extent, so that the defect of single processing method of the complex data of the marine environment can be made up; on the other hand, the characteristics of the seabed sediment are extracted to the maximum degree, so that the accuracy of sediment classification is improved. The single texture feature extraction method has certain limitations, which may cause the loss of key features. Therefore, the research on the multi-feature extraction of the seabed sediment has important significance.
Seabed sediment characteristic extraction is based on sonar images. For different seabed substrates, the intensity of the echoes received by the side-scan sonar is different, and the gray value of each pixel point corresponding to the side-scan sonar image is also different, so that the light and shade change of the side-scan sonar image can reflect the type of the seabed substrate to a great extent. The classification target of the seabed sediment image is to classify pixels or regions in the image into a certain category, and quickly identify and judge the target range so as to further achieve the purposes of inversion and restoration of seabed landforms. At present, on one hand, classification of seabed sediments can be carried out by combining acoustic parameters, such as acoustic impedance and acoustic absorption coefficient, with empirical formulas of physical properties of the sediments; on the other hand, classification is carried out on various characteristic parameters based on the substrate sonar images, the basic idea is that characteristic information is extracted from image data through a classical algorithm or a new theory, the characteristic information reflects the images at different angles, the characteristic information is input into a classifier, classification learning is carried out on the characteristics according to respective learning strategies of classification models, higher classification accuracy is obtained through continuous training, and finally classification of the images is achieved.
The traditional bottom material classification method relies on the original scattering data of the sea bottom, and calculates indexes to classify the types of the regional bottom materials, such as fast Fourier transform spectrum characteristics, singular value decomposition coefficients and wavelet transform entropy. In addition, the method is judged by using a substrate physical characteristic empirical formula, and the method needs to carry out experimental research on various contents such as a seabed high-frequency sound scattering theory, data acquisition and pretreatment, evaluation of a seabed classification system, verification of a classification result and the like for a long time. Because the seabed environment is very complicated, the traditional method has strong limitation and the classification precision is slightly low.
With the continuous development of artificial intelligence technology, various artificial intelligence methods are used for substrate classification. Such as SVM (support vector machine), neural networks, etc. The corresponding feature extraction method to be used includes principal component analysis, wavelet transform, gray matrix method, and the like. Efficient feature extraction and new classification methods are currently the direction of research that is gaining attention.
Most of the existing seabed sediment characteristic extraction methods are only based on one image characteristic, and a single texture characteristic extraction method has certain limitation, cannot fully express the characteristics of the seabed sediment, and the used classification method has low generalization capability.
Disclosure of Invention
Aiming at the prior art, the invention aims to provide an improved GoogleLeNet seabed substrate sonar image classification method based on multiple features, combines the advantages of multiple feature extraction methods, and improves the accuracy of seabed substrate classification by adopting the improved GoogleLeNet.
In order to solve the technical problem, the invention provides a submarine substrate sonar image classification method, which comprises the following steps:
step 1: acquiring a side-scan sonar image data set, wherein the data set comprises an image to be classified and an image with a truth value label;
step 2: calculating a gray level co-occurrence matrix, and performing statistical feature extraction on the side scan sonar image by using the gray level co-occurrence matrix;
and step 3: performing fractal dimension feature extraction on the side scan sonar image based on a pixel coverage method;
and 4, step 4: extracting a channel energy characteristic value of the side scan sonar image;
and 5: and performing joint representation on the extracted statistical characteristics, fractal dimension characteristics and characteristic parameters of the channel energy characteristics to form joint characteristics, inputting the joint characteristic parameters of the images to be classified into GoogLeNet trained by using the images with truth labels to classify the seabed substrate, and outputting a result as the image substrate category by a network.
Further, step 1 carries out preprocessing on the acquired side scan sonar image data set, wherein the preprocessing comprises the following steps: and performing noise smoothing and enhancement processing on the side-scan sonar image by using a sequencing self-adaptive median filtering and histogram equalization method.
Further, the calculating the gray level co-occurrence matrix in step 1 includes:
given a distance d and a direction theta, the gray level of a pixel point on a straight line with the direction theta is i, if the gray level of a pixel point with the distance d from the pixel point is j, the frequency of the gray level pair appearing at the same time is the value of the (i, j) th array element of a gray level co-occurrence matrix P (i, j, d, theta), a gray level co-occurrence matrix is generated in the directions of 0 degrees, 45 degrees, 90 degrees and 135 degrees, the gray level co-occurrence matrix in each direction is NxN order, and N is the gray level of an image, and the method specifically comprises the following steps:
P(i,j,d,0°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)|n 1 -n 2 ＝0,
|m 1 -m 2 |＝d,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
P(i,j,d,45°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)||m 1 -m 2 |＝d,
|n 1 -n 2 |＝d,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
P(i,j,d,90°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)||n 1 -n 2 |＝d,
m 1 -m 2 ＝0,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
P(i,j,d,135°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)||m 1 -m 2 |＝d,
|n 1 -n 2 |＝d,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
wherein Z represents the number of elements in the statistical set, M and N represent the number of rows and columns of the image, respectively, (M) 1 ,n 1 ) And (m) 2 ,n 2 ) And representing the coordinates of two pixel point pairs, I is a function of the gray value, and I and j respectively represent the corresponding gray values.
Further, utilizing the gray level co-occurrence matrix to perform statistical feature extraction on the side-scan sonar image comprises:
selecting an angular second moment, entropy, correlation and inertia moment as statistics to represent the characteristics of the seabed sediment, and respectively obtaining 4 characteristic values of the image in the directions of 0 degree, 45 degrees, 90 degrees and 135 degrees;
the angular second moment is specifically:
Figure BDA0003772335350000031
the entropy is specifically:
Figure BDA0003772335350000032
the correlation is specifically:
Figure BDA0003772335350000033
Figure BDA0003772335350000034
Figure BDA0003772335350000035
the moment of inertia is specifically:
Figure BDA0003772335350000037
wherein the value of theta is 0 degrees, 45 degrees, 90 degrees and 135 degrees.
Further, the step 3 of extracting the fractal dimension features of the side-scan sonar image based on a pixel coverage method comprises the following steps:
firstly, binarizing pixel points of a side-scan sonar image to form a digital matrix, and equally dividing the digital matrix into n small blocks, wherein the fractal dimension D specifically comprises the following steps:
Figure BDA0003772335350000036
and fitting data points (log n and log delta) in a dual-logarithmic coordinate system by using a least square method, and calculating to obtain a slope and a negative pseudo-fractal dimension of the slope.
Further, the step 4 of extracting the channel energy characteristic value of the side-scan sonar image comprises the following steps:
firstly, performing wavelet decomposition on a side-scan sonar image, binarizing the side-scan sonar gray level image, calculating by using a wavelet decomposition function to obtain an image threshold, setting the number of decomposition layers to be 4, and decomposing the image by selecting a db3 wavelet basis;
then, the wavelet energy is calculated:
Figure BDA0003772335350000041
in the formula: f (x, y) is a wavelet coefficient of each decomposed sub-image, x and y represent horizontal and vertical coordinates of pixel points, and M and N respectively represent the row number and column number of the image;
calculating the energy percentage of each channel of the high-frequency subgraph and the low-frequency subgraph as a channel energy characteristic value, specifically:
Figure BDA0003772335350000042
Figure BDA0003772335350000043
in the formula: n is the current decomposition layer number, L is the low-frequency channel, norm (L) represents the energy value of the low-frequency channel, norm (X) n ) And the energy value of the high-frequency channel of the nth layer is shown.
Further, the improved google lenet comprises a fully-connected layer, a 5 × 5 convolutional layer, a max pooling layer, a local response normalization layer, a 3 × 3 convolutional layer, a local response normalization layer, a max pooling layer, 4 continuous inclusion layers, a global average pooling layer, a fully-connected layer and softmax which are connected in sequence.
Further, the training in step 5 comprises:
dividing the image with the truth label into a training set and a verification set, inputting the combined feature vector of the training set and the verification set into GoogLeNet, using the cross entropy as a loss function, carrying out network training by using a back propagation algorithm, calculating the verification accuracy on the verification set, and finishing the training when the verification accuracy is not increased any more.
The invention has the beneficial effects that: the invention relates to a multi-feature fusion extraction and classification method for submarine substrate texture aiming at a side-scan sonar image. The method combines a plurality of feature extraction methods, comprehensively captures the texture feature change rule of the side-scan sonar image, each feature extraction method has respective advantages, and the limitation of a single method is complemented by the combined use of different methods. The GoogLeNet is improved by using the fusion feature input for training, so that the classification accuracy can be effectively improved.
Compared with the traditional single feature extraction and classification method, the method has the remarkable characteristics that: according to the invention, the statistical characteristic, the fractal dimension characteristic and the channel energy characteristic extraction method are respectively applied to the sonar images which are smoothed and enhanced by the sorting self-adaptive median filtering and histogram equalization method to extract the image characteristics, the combined characteristic parameters extracted by the three methods are input into the improved GoogleNet to classify the seabed substrate, the detail characteristics of the image can be effectively mastered, and the classification accuracy of the image is improved. The feature extraction method integrates the characteristics of various feature extraction methods, makes up the defect of using a single feature extraction method, and comprehensively captures the change rule of the texture features of the side-scan sonar image. Aiming at the characteristics of sonar images, the GoogleLeNet structure is improved, so that the method adapts to sonar images of the seabed substrate, accelerates the network convergence speed, and can effectively improve the accuracy of seabed substrate classification.
Compared with the prior art, the invention has the beneficial effects that:
(1) Before feature extraction, the image is preprocessed by filtering, histogram equalization enhancement and the like, so that the noise interference of the image is reduced, and the significance of image feature information is ensured.
(2) The invention uses three feature extraction methods, adopts a multi-feature fusion mode, and makes up for the defect that a single feature extraction method may have incomplete extraction features. The classification accuracy of the classifier can be effectively improved.
(3) The invention uses the improved GoogLeNet for classification. The improved GoogLeNet removes a full connection layer in the conventional convolutional neural network, reduces a large number of parameters, saves the calculation time and relieves the overfitting problem. In addition, an inclusion framework is introduced, input multi-scale information can be obtained, convolution and pooling of different scales are fused, classification accuracy is improved, and calculation amount can be saved. Aiming at the characteristic dimensionality of the sonar image, the invention adjusts the layer number of the network, so that the network can process a smaller sonar image. Meanwhile, redundant auxiliary softmax is removed, and the network convergence speed is improved. The classification by using the improved GoogLeNet can effectively improve the classification accuracy of the sonar images of the seabed substrate.
By adopting the method provided by the invention to classify the submarine substrate sonar images, the difference between different types of images can be effectively extracted, and the classification accuracy is higher than that of the existing method.
Drawings
FIG. 1 is a flow chart of a method for extracting the multi-feature of the seabed sediment based on sonar images;
FIG. 2 is a flow chart of a sorting adaptive median filtering algorithm in sonar image preprocessing;
fig. 3 is a schematic diagram of an inclusion architecture in a network;
fig. 4 is a diagram of the structure of the google lenet network used in the present invention.
Detailed Description
The invention is further described with reference to the drawings and examples.
The method comprises the steps of preprocessing a sonar image, then obtaining a sonar image combined feature by using various feature extraction methods, and then inputting the obtained image texture combined feature into GoogleLeNet to classify the seabed substrate. By using the method, the change rule of the texture features of the side-scan sonar image can be effectively captured, and the accuracy of side-scan sonar image classification is improved. With reference to fig. 1, the invention relates to a submarine substrate sonar image classification method, which comprises the following steps:
the method comprises the following steps: preprocessing a sonar image;
and respectively carrying out noise smoothing and enhancement processing on the side-scan sonar image by utilizing a sequencing self-adaptive median filtering method and a histogram equalization method. The purpose is to disturb the noise in the image and enhance the details.
Step two: performing statistical feature extraction on the side-scan sonar image by using a gray level co-occurrence matrix;
and calculating the occurrence times of pixel pairs with certain intensity under certain displacement to construct a gray level co-occurrence matrix, and calculating 4 characteristic statistics according to the matrix.
Step three: performing fractal dimension feature extraction on the side scan sonar image based on a pixel coverage method;
and obtaining a digital matrix through the side scan sonar image, dividing the digital matrix into blocks to obtain the box number, and calculating the physical box-counting dimension of the seabed sediment image.
Step four: extracting the channel energy characteristic value of the side-scan sonar image;
firstly, calculating the threshold value of a side-scan sonar image, then decomposing the image, calculating wavelet energy, calculating and summarizing the percentage distribution characteristics of different seabed sediment energies, and simultaneously judging whether the energy can be used as an effective characteristic parameter.
Step five: classification of seafloor substrates was performed using a modified google net.
And performing joint representation on the characteristic parameters extracted by the statistical characteristic, the fractal dimension characteristic and the channel energy characteristic to form a joint characteristic. Dividing the processed data into a training set and a testing set, using the training set to obtain a trained network, inputting the characteristic parameters to be classified into GoogLeNet to classify the seabed substrate, and outputting the result of the network, namely the image substrate category.
Examples are given below with specific parameters:
the method comprises the following steps: sonar image preprocessing
Acquiring a side-scan sonar image data set, wherein the data set comprises an image to be classified and an image with true-value labels; images with true value labels are used as a training sonar image and a verification sonar image, the images with true value labels are required to be used for the two types of sonar images, the images are called a training data set and a verification data set, and the ratio of the training data set to the verification data set is 4:1.
Due to interference from various aspects such as a complex marine environment, sound wave propagation characteristics, sonar equipment limitation and the like, imaging sonar systems are often affected by various types of noise. For a side-scan sonar system, the seafloor reverberation generated by the sound wave scattering from the rough seafloor surface and various scatterers in the sea water is essentially a random signal, and the reverberation noise can be approximately described by speckle noise and is determined by multiple researches: the statistical properties of this speckle noise are described by rayleigh distribution rather closely.
The speckle noise is multiplicative noise, and is modeled for side-scan sonar image noise according to a multiplicative noise model:
u(x,y)＝u 0 (x,y)·N(x,y)
u 0 (x, y) is an image not contaminated by noise, N (x, y) is speckle noise subject to Rayleigh distribution, and u (x, y) is a noisy image. The above formula describes speckle noise which greatly interferes with a sonar image, and in addition to the speckle noise, additive gaussian noise in the marine environment also affects the image, but the degree of the effect is not as good as that of the speckle noise.
The statistical properties of speckle noise are generally described by a probability density function, where the probability density of speckle noise N, subject to rayleigh distribution, is:
Figure BDA0003772335350000071
alpha is an attenuation parameter of the Rayleigh distribution and alpha > 0. In summary, median filtering is used which is better at removing speckle noise.
Step 1.1: the image is first smoothed by a rank-ordered adaptive median filter.
Compared with the traditional median filtering, the method adjusts the size of a filtering window by the noise density, and differentiates noise points and edge points: and performing median filtering on the noise, and keeping the gray value unchanged for the edge points. The method has smaller damage degree to the edge characteristics, and is more beneficial to subsequent characteristic enhancement and characteristic extraction processing.
The sorting self-adaptive median filtering takes the image statistical characteristics defined by a rectangular filtering window Sxy with the size of m multiplied by n as the basis, has good filtering effect on the pulse noise with high probability, and can keep details when smoothing non-pulse noise. Within the filter window there are the following variables: zxy is the gray scale value at point (x, y), zmax and Zmin are the maximum and minimum gray scale values, zmed is the gray scale median, and Smax is the maximum size allowed by Sxy. The sequencing self-adaptive median filtering algorithm mainly comprises two parts, namely a process A and a process B, and the specific process comprises the following steps:
1) And a process A:
A 1 ＝Zmed-Z min
A 2 ＝Zmed-Z max
if A 1 > 0 and A 2 If the window size is less than 0, the process B is switched to, otherwise, the window size is increased; if the window size does not exceed Smax, then process A is repeated, otherwise Zmed is output.
2) And a process B:
B 1 ＝Zxy-Zmin
B 2 ＝Zxy-Zmax
if B is present 1 > 0 and B 2 If < 0, it outputs Zxy, otherwise it outputs Zmed.
As described above, in the algorithm, the noise point and the edge point are distinguished according to the maximum Zmax and Zmin of the gray value, if Zmed is between Zmax and Zmin, zmed is regarded as edge information, otherwise, it is noise; if Zxy is between Zmax and Zmin, zxy is an edge signal, otherwise it is noise. When both the Zmed and Zxy are edge information, the gray value Zxy of the point is preferentially output; when Zmed is noise, output Zxy; when Zxy is noise, zmed is output. The rank adaptive median filtering ignores gray maxima and minima at each window, and these ignored pixels are often noise.
Step 1.2: histogram equalization is applied to enhance the sonar image.
The gray scale statistical histogram of the image describes statistical information of each gray scale of the image, reflects the frequency or probability of each gray scale of an image, can give a general description of gray scale distribution, and can be described by the following one-dimensional discrete function:
Figure BDA0003772335350000081
S k is the K-th gray scale value, n, of the image f (x, y) k Is f (x, y) has a gray value S k N is the total number of image pixels, P (S) k ) Representing the probability of occurrence of the kth gray level of the original image. With n k As an independent variable, with P (S) k ) The functional relationship obtained for the dependent variable is the histogram of the image. Histogram equalization, also called gray level equalization, refers to that all levels of gray levels of an image are uniformly output by approximately the same number of pixels through a certain gray level mapping, so that the frequency of the occurrence of brightness values in an extended range tends to be equalized, and thus the dynamic range of the gray level values of the pixels is increased, and the contrast of the image is increased. The histogram equalization follows the criterion that the higher contrast can be displayed only by fully utilizing all gray level ranges of the continuous tone image, so that the bright and dark parts of the processed image show greater contrast, and the edge details are clearer, thereby achieving the effects of enhancing the visual experience of the image and highlighting the characteristic information. The gray level conversion process can be described by the following equation:
Figure BDA0003772335350000082
D A is the original gray value, D B Is the converted gray value.
Step two: statistical feature extraction for side-scan sonar images using gray level co-occurrence matrix
The gray level co-occurrence matrix of the seabed sediment image is calculated by the following method: given a distance d and a direction theta, on a straight line with the direction theta, the gray level of one pixel point is i, and if the gray level of one pixel point with the distance d from the pixel point is j, the frequency of the simultaneous occurrence of such gray level pairs is the value of the (i, j) th array element of the gray level co-occurrence matrix P (i, j, d, theta). If the gray scale of the image is N, the gray co-occurrence matrix in each direction is N × N, that is, N × N kinds of possibilities of pixel-to-gray value combinations in a specific direction and a specific distance are available. The gray level co-occurrence matrix is generated by taking the directions of 0 degrees, 45 degrees, 90 degrees and 135 degrees, namely:
P(i,j,d,0°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)|n 1 -n 2 ＝0,
|m 1 -m 2 |＝d,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
P(i,j,d,45°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)||m 1 -m 2 |＝d,
|n 1 -n 2 |＝d,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
P(i,j,d,90°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)||n 1 -n 2 |＝d,
m 1 -m 2 ＝0,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
P(i,j,d,135°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)||m 1 -m 2 |＝d,
|n 1 -n 2 |＝d,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
wherein Z represents the number of elements in the statistical set, M and N represent the number of rows and columns of the image, respectively, (M) 1 ,n 1 ) And (m) 2 ,n 2 ) And representing the coordinates of two pixel point pairs, wherein I is a function of taking a gray value, I and j respectively represent the corresponding gray values, and the gray co-occurrence matrix records the number of the pixel pairs with the same gray value combination.
In an image, along a specific direction theta, the gray values of two pixel points separated by a specific interval d are i and j respectively. The features of the seafloor substrate are represented by 4 statistics (angular second moment, entropy, correlation, moment of inertia). The calculation formula for the 4 statistics is as follows:
(1) Second moment of angle
Figure BDA0003772335350000091
(2) Entropy of the entropy
Figure BDA0003772335350000101
(3) Correlation
Figure BDA0003772335350000102
Figure BDA0003772335350000103
Figure BDA0003772335350000104
(4) Moment of inertia
Figure BDA0003772335350000105
The feature values of 4 original images in each direction can be obtained, and the step corresponds to the first 20 bits in the combined feature
Step three: fractal dimension feature extraction is carried out on side-scan sonar image based on pixel coverage method
Firstly, binarizing pixel points of a side-scan sonar image to form a digital matrix, equally dividing the digital matrix into n small blocks, wherein the fractal dimension D can be represented by the following formula:
Figure BDA0003772335350000106
in order to calculate the physical box-counting dimension of the seabed sediment image intuitively, fitting data points (log n and log delta) in a dual-logarithmic coordinate system by using a least square method, calculating a slope, wherein the negative value of the slope is the solved dimension, and the solved fractal dimension can form a 10-bit vector through coding.
Step four: channel energy characteristic value extraction is carried out on side scan sonar images
The image is first wavelet decomposed. Firstly, binarizing a side-scan sonar gray level image, calculating an image threshold value by using a wavelet decomposition function, and setting the number of decomposition layers to be 4 layers in order to prevent the occurrence of information redundancy; due to the orthogonality, the compactness and the low complexity of the wavelet function, the db3 wavelet basis is selected for decomposing the image.
After the image decomposition is completed, the wavelet energy needs to be calculated, and the formula for calculating the energy is as follows:
Figure BDA0003772335350000111
in the formula: f (x, y) is the wavelet coefficient of each decomposed sub-image, x, y represent the horizontal and vertical coordinates of the pixel points, and M, N represent the number of rows and columns of the image respectively.
And finally, calculating and summarizing the percentage distribution characteristics of different seabed sediment energies, wherein the energy percentage formula of each channel of the high-frequency subgraph and the low-frequency subgraph is as follows:
Figure BDA0003772335350000112
Figure BDA0003772335350000113
in the formula: n is the current decomposition layer number, L is the low-frequency channel, norm (L) represents the energy value of the low-frequency channel, norm (X) n ) And the energy value of the high-frequency channel of the nth layer is shown. E is the energy value of the image as a whole. Finally, the energy percentage can be encoded as a 20-dimensional vector, and the final three features form a 1 × 50 tensor, forming the combined features.
Step five: classification of seafloor substrates using improved google lenet
And performing joint representation on the characteristic parameters extracted from the statistical characteristics, the fractal dimension characteristics and the channel energy characteristics to obtain a tensor with the length of 1 × 50. This tensor can be directly input into the google lenet of the present invention for classification of seafloor substrates.
The improved google lenet network architecture used by the present invention is shown in fig. 4. First, the input conversion layer converts the input 50-dimensional tensor input full connection layer into 250 dimensions, and further into a 50 × 50 eigenmap. This profile is passed through a pre-layer that reduces the overfitting using local corresponding normalization. Then, 4 consecutive inclusion layers are input, and the inclusion architecture details are shown in fig. 3, which can concatenate the outputs of various convolution windows. And finally, outputting the image category after replacing the full connection layer by using the global average pooling plus softmax.
The improved network reduces the number of layers of the network, and adjusts the size of a convolution kernel, so that the improved network can process sonar image characteristic data with lower scale. Meanwhile, redundant auxiliary softmax is removed, and the convergence speed is improved. Meanwhile, the original GoogLeNet processes 3-channel images, and the convolution structure is adjusted to process a single-channel feature map.
In order to enable the network to have the capability of classifying submarine substrate sonar images, joint feature vectors are respectively calculated by the training set and the verification set in the step one and then input into the network, cross entropy (cross entropy) is used as a loss function, network training is carried out by using a back propagation algorithm, and verification accuracy is calculated on the verification set. Training is completed when the validation accuracy no longer increases.
The trained network can be used for carrying out submarine substrate sonar image classification. The image characteristics are input, and the classifier automatically outputs the substrate category to which the image belongs.

Claims (8)

1. A submarine substrate sonar image classification method is characterized by comprising the following steps:
step 1: acquiring a side-scan sonar image data set, wherein the data set comprises an image to be classified and an image with a truth value label;
step 2: calculating a gray level co-occurrence matrix, and performing statistical feature extraction on the side-scan sonar image by using the gray level co-occurrence matrix;
and step 3: performing fractal dimension feature extraction on the side scan sonar image based on a pixel coverage method;
and 4, step 4: extracting a channel energy characteristic value of the side scan sonar image;
and 5: and performing joint representation on the extracted characteristic parameters of the statistical characteristic, the fractal dimension characteristic and the channel energy characteristic to form a joint characteristic, inputting the joint characteristic parameters of the image to be classified into GoogLeNet trained by using the image with the true value label to classify the seabed substrate, and outputting a network output result as the image substrate category.
2. The method for classifying the sonar images of the seabed substrate according to claim 1, characterized by comprising the following steps: step 1, preprocessing the acquired side scan sonar image data set, wherein the preprocessing comprises the following steps: and performing noise smoothing and enhancement processing on the side-scan sonar image by using a sequencing self-adaptive median filtering and histogram equalization method.
3. The method for classifying the sonar images of the seabed substrate according to claim 1, is characterized in that: step 1, the calculating the gray level co-occurrence matrix comprises:
given a distance d and a direction theta, the gray level of a pixel point on a straight line with the direction theta is i, if the gray level of a pixel point with the distance d from the pixel point is j, the frequency of the simultaneous occurrence of the gray level pair is the value of the (i, j) th array element of a gray level co-occurrence matrix P (i, j, d, theta), the gray level co-occurrence matrix is generated in the directions of 0 degrees, 45 degrees, 90 degrees and 135 degrees, the gray level co-occurrence matrix in each direction is NxN order, N is the gray level of an image, and the method specifically comprises the following steps:
P(i,j,d,0°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)|n 1 -n 2 ＝0,|m 1 -m 2 |＝d,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
P(i,j,d,45°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)||m 1 -m 2 |＝d,|n 1 -n 2 |＝d,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
P(i,j,d,90°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)||n 1 -n 2 |＝d,m 1 -m 2 ＝0,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
P(i,j,d,135°)＝Z{((m 1 ,n 1 ),(m 2 ,n 2 ))∈(M,N)×(M,N)||m 1 -m 2 |＝d,|n 1 -n 2 |＝d,I(m 1 ,n 1 )＝i,I(m 2 ,n 2 )＝j}
wherein Z represents the number of elements in the statistical set, M and N represent the number of rows and columns of the image, respectively, (M) 1 ,n 1 ) And (m) 2 ,n 2 ) And representing the coordinates of two pixel point pairs, I is a function of the gray value, and I and j respectively represent the corresponding gray values.
4. The method for classifying the sonar images of the seabed substrate according to claim 3, is characterized in that: the method for extracting the statistical characteristics of the side-scan sonar image by utilizing the gray level co-occurrence matrix comprises the following steps:
selecting an angular second moment, entropy, correlation and inertia moment as statistics to represent the characteristics of the seabed sediment, and respectively obtaining 4 characteristic values of the image in the directions of 0 degree, 45 degrees, 90 degrees and 135 degrees;
the angular second moment is specifically:
Figure FDA0003772335340000021
the entropy is specifically:
Figure FDA0003772335340000022
the correlation is specifically:
Figure FDA0003772335340000023
Figure FDA0003772335340000024
Figure FDA0003772335340000025
the moment of inertia is specifically:
Figure FDA0003772335340000026
wherein the value of theta is 0 degrees, 45 degrees, 90 degrees and 135 degrees.
5. The method for classifying the sonar images of the seabed substrate according to claim 1, is characterized in that: step 3, the fractal dimension feature extraction of the side scan sonar image based on the pixel point coverage method comprises the following steps:
firstly, binarizing pixel points of a side-scan sonar image to form a digital matrix, and equally dividing the digital matrix into n small blocks, wherein the fractal dimension D specifically comprises the following steps:
Figure FDA0003772335340000027
wherein, delta represents the side length of each small block, and a least square method is utilized to fit data points (logn, log delta) in a dual logarithmic coordinate system, and the slope and the negative pseudo-fractal dimension of the slope are obtained through calculation.
6. The method for classifying the sonar images of the seabed substrate according to claim 1, is characterized in that: step 4, the channel energy characteristic value extraction of the side scan sonar image comprises the following steps:
firstly, performing wavelet decomposition on a side-scan sonar image, binarizing the side-scan sonar gray level image, calculating by using a wavelet decomposition function to obtain an image threshold, setting the number of decomposition layers to be 4, and decomposing the image by selecting a db3 wavelet basis;
then, the wavelet energy is calculated:
Figure FDA0003772335340000031
in the formula: f (x, y) is a wavelet coefficient of each decomposed sub-image, x and y represent horizontal and vertical coordinates of pixel points, and M and N respectively represent the row number and column number of the image;
calculating the energy percentage of each channel of the high-frequency subgraph and the low-frequency subgraph as a channel energy characteristic value, specifically:
Figure FDA0003772335340000032
Figure FDA0003772335340000033
in the formula: n is the current decomposition layer number, L is the low-frequency channel, norm (L) represents the energy value of the low-frequency channel, norm (X) n ) And the energy value of the high-frequency channel of the nth layer is shown.
7. The method for classifying the sonar images of the seabed substrate according to claim 1, is characterized in that: the improved GoogLeNet comprises a full connection layer, a 5 x 5 convolution layer, a maximum pooling layer, a local response normalization layer, a 3 x 3 convolution layer, a local response normalization layer, a maximum pooling layer, 4 continuous inclusion layers, a global average pooling layer, a full connection layer and softmax which are connected in sequence.
8. The method for classifying the sonar images of the seabed substrate according to claim 1, characterized by comprising the following steps: step 5 the training comprises:
dividing the image with the truth label into a training set and a verification set, inputting the joint feature vector of the training set and the verification set into GoogLeNet, taking the cross entropy as a loss function, carrying out network training by using a back propagation algorithm, calculating the verification accuracy on the verification set, and finishing the training when the verification accuracy is not increased any more.
CN202210905490.4A 2022-07-29 2022-07-29 Seabed substrate sonar image classification method Pending CN115187855A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210905490.4A CN115187855A (en) 2022-07-29 2022-07-29 Seabed substrate sonar image classification method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210905490.4A CN115187855A (en) 2022-07-29 2022-07-29 Seabed substrate sonar image classification method

Publications (1)

Publication Number Publication Date
CN115187855A true CN115187855A (en) 2022-10-14

Family

ID=83521088

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210905490.4A Pending CN115187855A (en) 2022-07-29 2022-07-29 Seabed substrate sonar image classification method

Country Status (1)

Country Link
CN (1) CN115187855A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117953314A (en) * 2024-03-26 2024-04-30 自然资源部第三海洋研究所 Multi-dimensional feature optimization ocean substrate classification method and system

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117953314A (en) * 2024-03-26 2024-04-30 自然资源部第三海洋研究所 Multi-dimensional feature optimization ocean substrate classification method and system

Similar Documents

Publication Publication Date Title
CN111242862B (en) Multi-scale fusion parallel dense residual convolution neural network image denoising method
CN103049892B (en) Non-local image denoising method based on similar block matrix rank minimization
CN112184577B (en) Single image defogging method based on multiscale self-attention generation countermeasure network
CN113688941B (en) Small sample sonar image classification recognition optimization method based on generation countermeasure network
CN110288550B (en) Single-image defogging method for generating countermeasure network based on priori knowledge guiding condition
CN108550121A (en) A kind of sediment sonar image processing method based on medium filtering and wavelet transformation
CN111723701A (en) Underwater target identification method
CN112819732A (en) B-scan image denoising method for ground penetrating radar
CN111145145B (en) Image surface defect detection method based on MobileNet
CN114972107A (en) Low-illumination image enhancement method based on multi-scale stacked attention network
CN115661649B (en) BP neural network-based shipborne microwave radar image oil spill detection method and system
CN110675410A (en) Side-scan sonar sunken ship target unsupervised detection method based on selective search algorithm
CN108985304A (en) It is a kind of based on the Structure of the deposits extraction method for shallowly cuing open data
CN110490816A (en) A kind of underwater Heterogeneous Information data noise reduction
CN115861359B (en) Self-adaptive segmentation and extraction method for water surface floating garbage image
CN114764801A (en) Weak and small ship target fusion detection method and device based on multi-vision significant features
CN116468995A (en) Sonar image classification method combining SLIC super-pixel and graph annotation meaning network
CN115187855A (en) Seabed substrate sonar image classification method
He et al. GPR image denoising with NSST-UNET and an improved BM3D
CN109034070B (en) Blind separation method and device for replacement aliasing image
Zhou et al. Deep denoising method for side scan sonar images without high-quality reference data
CN110751667A (en) Method for detecting infrared dim small target under complex background based on human visual system
CN111627030A (en) Rapid and efficient sea-land accurate segmentation method for visible light remote sensing image
CN116452450A (en) Polarized image defogging method based on 3D convolution
CN115689958A (en) Synthetic radar image denoising method based on deep learning

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