CN112862748B - Multi-dimensional domain feature combined SAR ship intelligent detection method - Google Patents
Multi-dimensional domain feature combined SAR ship intelligent detection method Download PDFInfo
- Publication number
- CN112862748B CN112862748B CN202011562298.7A CN202011562298A CN112862748B CN 112862748 B CN112862748 B CN 112862748B CN 202011562298 A CN202011562298 A CN 202011562298A CN 112862748 B CN112862748 B CN 112862748B
- Authority
- CN
- China
- Prior art keywords
- feature
- sar
- vessel
- domain
- target
- 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.)
- Active
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 54
- 230000009466 transformation Effects 0.000 claims abstract description 27
- 230000004927 fusion Effects 0.000 claims abstract description 16
- 239000013598 vector Substances 0.000 claims description 15
- 238000001228 spectrum Methods 0.000 claims description 12
- 238000000605 extraction Methods 0.000 claims description 9
- 238000000034 method Methods 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 7
- 238000013528 artificial neural network Methods 0.000 claims description 6
- 238000013519 translation Methods 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 5
- 238000013527 convolutional neural network Methods 0.000 claims description 3
- 238000007499 fusion processing Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 230000006399 behavior Effects 0.000 abstract description 10
- 238000013135 deep learning Methods 0.000 abstract description 5
- 230000000694 effects Effects 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 4
- 230000000295 complement effect Effects 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 238000011161 development Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008092 positive effect Effects 0.000 description 1
- 239000013589 supplement Substances 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
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- 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
- G06T2207/10032—Satellite or aerial image; Remote sensing
- G06T2207/10044—Radar 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
- G06T2207/20048—Transform domain processing
- G06T2207/20056—Discrete and fast Fourier transform, [DFT, FFT]
-
- 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]
-
- 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
- G06T2207/20212—Image combination
- G06T2207/20221—Image fusion; Image merging
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Quality & Reliability (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention belongs to the technical field of SAR ship target detection, and discloses a multi-dimensional domain feature combined SAR ship intelligent detection method, wherein the existing Synthetic Aperture Radar (SAR) ship detection method based on deep learning only utilizes space feature information of a ship target, and can not obtain satisfactory detection effects for multi-scale targets and rotating targets under a complex background. To overcome these problems, the present invention employs a feature pyramid network to learn multi-scale features of a vessel target space; and secondly, obtaining the rotation invariant feature of the SAR vessel target on the frequency domain by utilizing polar coordinate Fourier transformation. On this basis, a new space-frequency feature fusion network is proposed to obtain a more compact cross-domain feature representation. The invention effectively overcomes the scale transformation and rotation behavior of the SAR vessel target and improves the detection performance of the SAR vessel detection algorithm.
Description
Technical Field
The invention belongs to the technical field of SAR vessel target detection, and particularly relates to an intelligent detection method for a multi-dimensional domain feature combined SAR vessel.
Background
At present, the synthetic aperture radar can generate high-resolution microwave images of an observation scene under all-weather conditions throughout the day, and is widely applied to many civil and military fields. In recent years, SAR ship detection has been increasingly valued for its important value in practical applications such as offshore management, offshore traffic control, marine environmental protection, and the like. With the rapid development of SAR imaging technology, the observation level of a ship target is higher and higher, so that SAR ship detection based on fine information and high-level characteristics is possible.
The traditional SAR vessel target detection method mostly depends on manual participation, and target characteristics need to be selected according to manual experience; and the problems of motion blur, speckle noise and the like of the SAR image are caused, so that the traditional ship detection method is difficult to extract effective characteristics, and the target detection performance of the SAR ship is limited.
In recent years, with the development of artificial intelligence and SAR imaging technologies, more and more scholars apply a deep learning technology to the field of SAR ship target detection, and the deep learning technology can autonomously learn the refined characteristics and high-level semantic information of the ship target, so that the artificial participation degree is effectively reduced, and the ship target detection precision is improved. However, only the space characteristic information of the SAR vessel target is utilized, and the complementary characteristics of other dimension fields of the vessel target are ignored, so that the vessel target detection reliability and the classification recognition accuracy are limited.
Through the above analysis, the problems and defects existing in the prior art are as follows: the existing Shan Weiyu SAR ship detection method only utilizes the space characteristic information of ship targets, and particularly cannot obtain satisfactory detection effects under multi-scale or rotating targets and complex backgrounds.
The difficulty of solving the problems and the defects is as follows: 1. how to accurately extract the frequency domain rotation invariant feature of the vessel target in deep learning; 2. how to accurately align the spatial features and the frequency domain features so as to ensure the effectiveness of feature fusion; 3. how to effectively realize cross-domain fusion of multidimensional domain features and obtain better feature expression.
The meaning of solving the problems and the defects is as follows: 1. the SAR ship detection network integrating multidimensional domain features and deep learning is provided; 2. the rotation problem of the ship target is solved by utilizing the frequency domain characteristics; 3. the ship detection performance of the method is superior to that of other mainstream target detection algorithms.
Disclosure of Invention
Aiming at the problems existing in the prior art, the invention provides an intelligent detection method for a multi-dimensional domain feature combined SAR vessel. According to the invention, by exploring and utilizing the spatial multi-scale characteristics and the frequency domain rotation invariant characteristics of the SAR vessel target, more compact cross-domain characteristic representation is obtained through a spatial-frequency domain characteristic fusion network, so that the detection performance is improved.
The invention is realized in such a way that the intelligent detection method of the multi-dimensional domain feature combined SAR vessel comprises the following steps:
step one, analyzing multi-scale problems and rotation problems existing in the existing SAR vessel targets, acquiring layered space characteristics from bottom to top through a backbone network, and acquiring the multi-scale characteristics by adopting FPN;
step two, the spatial multi-scale characteristics obtained in the step one are obtained through an RPN network, coordinate information and category information of a candidate region of the SAR image are obtained, and the coordinate information and the category information are mapped to an original image and a multi-scale characteristic image respectively;
step three, processing is carried out by combining a time-frequency domain dimension domain aiming at the rotation behavior of the ship; according to the candidate region of the original image obtained in the second step, rotating invariant information of the SAR vessel target is obtained through Fourier-polar coordinate transformation, and the rotating invariant feature of the SAR vessel target frequency domain is obtained through feature extraction;
and fourthly, designing a finer feature fusion network aiming at space-frequency domain feature fusion, and obtaining more compact feature expression in different modes by interactively updating parameters of different networks so as to comprehensively characterize the features of each dimension domain.
Further, in the first step, the neural network Ω is convolved through a spatial channel S The spatial feature phi can be easily obtained from the SAR image S (x, y), which can be expressed as
φ S (x,y)=Ω S (I(x,y),θ s );
Wherein I (x, y) represents an input SAR vessel image, θ s Representing omega S Related parameter sets in (a); upsampling high-level features by FPN and enhancement via cross-connect to yield spatial multi-scale features φ S_MS (x,y)。
Further, in the second step, the proposed area of the image is obtained by sharing the spatial feature to the RPN, and the specific process is as follows:
the first step, a 3×3 convolution check is used to convolve the shared feature map, so as to make the extracted features more robust;
secondly, using a convolution kernel of 1 multiplied by 1 twice to obtain category information and position information of the candidate region;
third, mapping the candidate region position information to the original image and the multi-scale feature map to obtain I respectively Sub (x, y) and phi S_MS_Sub (x,y)。
Further, in the third step, the obtained original image candidate region I Sub (x, y) performing polar coordinate transformation, wherein the transformation relation is as follows:
where r represents the distance from the point to the origin in a rectangular coordinate system, θ is the angle between the x-axis and the line connecting the point and the origin, and is positive in the counterclockwise direction.
Further, the obtained polar coordinate image is subjected to Fourier transformation and the amplitude spectrum is taken
Wherein F is Sub (mu, v) represents candidate region frequency domain information, mu, v are corresponding two-dimensional frequency domain variables respectively, and the amplitude spectrum is taken to obtain rotation invariant information M Sub (u,v)。
Further, the rotation invariant information M obtained by the convolution neural network pair Sub (u, v) feature extraction
φ F_RI_Sub (x,y)=Ω F (M Sub (u,v),θ F );
In phi F_RI_Sub (x, y) represents the rotation invariant feature of the SAR vessel target, Ω F Extracting network for frequency domain characteristics, θ F For its associated parameter set.
Further, in the third step, the rotation behavior of the SAR vessel target is processed, and the rotation invariant feature is extracted, and the specific process is as follows:
the rotation behavior under the rectangular coordinates of the image is represented as translation of the target under the polar coordinates, so that the rotation transformation under the rectangular coordinates is converted into translation transformation by utilizing the polar coordinate transformation; and simultaneously, combining the characteristic of Fourier transformation, and taking the magnitude spectrum of the spectrum information of the SAR ship target so as to adapt to any rotation angle of the SAR ship target, thereby obtaining the rotation invariant information of the SAR ship target.
In the third step, the rotation invariant feature of the SAR vessel target is obtained by extracting the feature of the SAR vessel target through a convolutional neural network.
In the fourth step, a finer feature fusion network is designed, and the spatial-frequency domain features are fused by interactively updating parameters of different networks, which comprises the following specific processes:
expanding the spatial features into one-dimensional vectors, and expanding the frequency domain features through the full connection layer to obtain feature vectors with the same dimension, wherein the obtained output feature size is 1024; the representation is
a S =FC(Flatten(φ S_MS_Sub (x,y)));
a F =Flatten(φ F_RI_Sub (x,y));
Wherein a is S And a F Output feature vectors representing the spatial domain and the frequency domain, respectively, and FC represents the fully connected layer.
Further, the cross-domain fusion process of the obtained one-dimensional spatial feature and the frequency domain feature vector is as follows:
a SF,1 =[FC 1 (a S ),FC 1 (a F )]
a SF,2 =[FC 2 (a F ),FC 2 (a S )]
where [ ·, ] denotes stitching together two feature vectors, the intersection between the network and the feature makes the output feature more compact while preserving the domain features of each dimension.
By combining all the technical schemes, the invention has the advantages and positive effects that: the invention starts from the angle of multi-dimensional domain feature combination, and excavates the space multi-scale feature and the frequency domain rotation invariant feature of the SAR vessel target by further excavating the multi-dimensional domain complementary information of the SAR vessel target, thereby providing a novel multi-dimensional domain feature combination SAR vessel intelligent detection network, effectively overcoming the rotating behavior of the scale transformation of the SAR vessel target and improving the detection performance of the SAR vessel detection algorithm.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following description will briefly explain the drawings needed in the embodiments of the present application, and it is obvious that the drawings described below are only some embodiments of the present application, and that other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a flowchart of a multi-dimensional domain feature combined SAR ship intelligent detection method provided by an embodiment of the invention.
Fig. 2 is a schematic diagram of a multi-dimensional domain feature combined SAR ship intelligent detection network provided by an embodiment of the present invention.
Fig. 3 is a schematic diagram of a spatial multi-scale feature extraction network according to an embodiment of the present invention.
Fig. 4 is a schematic diagram of a proposed network structure according to an embodiment of the present invention.
Fig. 5 is a schematic diagram of image rotation behavior under rectangular coordinates and polar coordinates according to an embodiment of the present invention.
Fig. 6 is a schematic diagram of a frequency domain feature extraction network according to an embodiment of the present invention.
Fig. 7 is a schematic diagram of a spatial-frequency domain feature fusion network according to an embodiment of the present invention.
Fig. 8 and 9 show the results of offshore and onshore vessel target detection, respectively, wherein fig. (a), (b), (c) show the results of real vessel targets, the detection result of the fast RCNN algorithm, and the detection result of the present invention, respectively.
Detailed Description
The present invention will be described in further detail with reference to the following examples in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
Aiming at the problems existing in the prior art, the invention provides an intelligent detection method for a multi-dimensional domain feature combined SAR vessel, and the invention is described in detail below with reference to the accompanying drawings.
As shown in fig. 1, the multi-dimensional domain feature combined SAR ship intelligent detection method provided by the embodiment of the invention includes:
s101: analyzing the multi-scale problem and the rotation problem existing in the existing SAR vessel targets, acquiring layered spatial characteristics from bottom to top through a backbone network, and acquiring the multi-scale characteristics by adopting FPN.
S102: and (3) acquiring coordinate information and category information of the SAR image candidate region through the RPN according to the spatial multi-scale feature acquired in the step (S101), and mapping the coordinate information and the category information to an original image and a multi-scale feature map respectively.
S103: processing is carried out by combining a time-frequency domain and a dimension domain aiming at the rotation behavior of the ship; and (3) according to the candidate region of the original image obtained in the step (S102), obtaining rotation invariant information of the SAR vessel target by carrying out Fourier-polar coordinate transformation on the candidate region, and obtaining the rotation invariant feature of the SAR vessel target frequency domain by carrying out feature extraction on the rotation invariant information.
S104: aiming at space-frequency domain feature fusion, a finer feature fusion network is designed, and by interactively updating parameters of different networks, more compact feature expression is obtained in different modes, so that the features of each dimension domain can be comprehensively represented.
In S101 provided by the embodiment of the present invention, the neural network Ω is convolved through a spatial channel S The spatial feature phi can be easily obtained from the SAR image S (x, y), which can be expressed as
φ S (x,y)=Ω S (I(x,y),θ s );
Wherein I (x, y) represents an input SAR vessel image, θ s Representing omega S Related parameter sets in (a); upsampling high-level features by FPN and enhancement via cross-connect to yield spatial multi-scale features φ S_MS (x,y)。
In S102 provided by the embodiment of the present invention, the spatial feature is shared to the RPN to obtain the proposed area of the image, and the specific process is as follows:
s201: the shared feature map is convolved using a 3 x 3 convolution kernel in order to make the extracted features more robust.
S202: the category information and the position information of the candidate region are obtained using a convolution kernel of 1×1 twice.
S203: mapping the candidate region position information to the original image and the multi-scale feature map to obtain I respectively Sub (x, y) and phi S_MS_Sub (x,y)。
In S203 provided by the embodiment of the present invention, for the obtained original image candidate region I Sub (x, y) performing polar coordinate transformation, wherein the transformation relation is as follows:
where r represents the distance from the point to the origin in a rectangular coordinate system, θ is the angle between the x-axis and the line connecting the point and the origin, and is positive in the counterclockwise direction.
The obtained polar coordinate image is subjected to Fourier transformation and the amplitude spectrum is taken
Wherein F is Sub (mu, v) represents candidate region frequency domain information, mu, v are corresponding two-dimensional frequency domain variables respectively, and the amplitude spectrum is taken to obtain rotation invariant information M Sub (u,v)。
The rotation invariant information M obtained by convolution neural network pair Sub (u, v) feature extraction
φ F_RI_Sub (x,y)=Ω F (M Sub (u,v),θ F );
In phi F_RI_Sub (x, y) represents the rotation invariant feature of the SAR vessel target, Ω F Extracting network for frequency domain characteristics, θ F For its associated parameter set.
In S103 provided by the embodiment of the present invention, the rotation behavior of the SAR vessel target is processed, and the rotation invariant feature is extracted, which specifically includes the following steps:
the rotation behavior under the rectangular coordinates of the image is represented as translation of the target under the polar coordinates, so that the rotation transformation under the rectangular coordinates is converted into translation transformation by utilizing the polar coordinate transformation; and simultaneously, combining the characteristic of Fourier transformation, and taking the magnitude spectrum of the spectrum information of the SAR ship target so as to adapt to any rotation angle of the SAR ship target, thereby obtaining the rotation invariant information of the SAR ship target. And extracting the characteristics of the SAR vessel target through a convolutional neural network to obtain the rotation invariant characteristics of the SAR vessel target.
In S104 provided by the embodiment of the present invention, a finer feature fusion network is designed, and the spatial-frequency domain features are fused by interactively updating parameters of different networks, which is specifically based on the following principle:
in order to ensure the dimension consistency of the spatial features and the frequency features, the spatial features are unfolded into one-dimensional vectors, the size of the obtained output features is 1024 through the full connection layer, and the frequency domain features are unfolded to obtain feature vectors with the same dimension; the representation is
a S =FC(Flatten(φ S_MS_Sub (x,y)));
a F =Flatten(φ F_RI_Sub (x,y));
Wherein a is S And a F Output feature vectors representing the spatial domain and the frequency domain, respectively, and FC represents the fully connected layer.
The cross-domain fusion process of the obtained one-dimensional spatial feature and the frequency domain feature vector comprises the following steps:
a SF,1 =[FC 1 (a S ),FC 1 (a F )]
a SF,2 =[FC 2 (a F ),FC 2 (a S )]
where [ ·, ] denotes stitching together two feature vectors, the intersection between the network and the features makes the output feature more compact while preserving the features of the dimension domain, with which a better, more efficient spatial frequency feature expression is achieved.
The technical scheme of the invention is further described below with reference to specific embodiments.
The intelligent detection network for the multi-dimensional domain feature combined SAR vessel provided by the invention supplements and expands the target space information by exploring and utilizing the multi-dimensional domain complementary feature, and deeply excavates the space multi-scale feature and the frequency domain rotation invariant feature, so that the intelligent detection network for the multi-dimensional domain feature combined SAR vessel is provided, and the detection performance of the SAR vessel is improved.
And (1) analyzing the multi-scale problem and the rotation problem existing in the existing SAR vessel target, and extracting the multi-scale characteristics of the SAR vessel target by using a characteristic pyramid, wherein the specific implementation is shown in figure 3.
And (2) acquiring candidate areas by utilizing an RPN (remote procedure network) according to the multi-scale features obtained in the step (1), and respectively mapping coordinate information onto the feature images and the original images so as to realize the alignment of space frequency information, wherein the specific implementation is shown in figure 4.
And (3) aiming at the rotation behavior of the ship target, obtaining rotation invariant information of the SAR ship target by utilizing Fourier-polar coordinate transformation, and extracting characteristics, wherein the network structure is shown in figure 6.
And (4) aiming at the spatial multi-scale characteristics and the frequency domain rotation invariant characteristics obtained in the step (2) and the step (3), utilizing a cross-fusion mode to perform cross-domain fusion of experimental spatial-frequency domain characteristics, and particularly realizing the cross-domain fusion shown in the figure 7.
The experimental results are shown in the figures, wherein figures 8 and 9 respectively show the offshore and onshore vessel target detection results, and figures (a), (b) and (c) respectively show the real vessel target, the detection result of the fast RCNN algorithm and the detection result of the invention. From the figure, the SAR vessel detection network provided by the invention can effectively reduce false alarm and omission under the background of offshore and onshore, and improve detection performance.
In the description of the present invention, unless otherwise indicated, the meaning of "a plurality" is two or more; the terms "upper," "lower," "left," "right," "inner," "outer," "front," "rear," "head," "tail," and the like are used as an orientation or positional relationship based on that shown in the drawings, merely to facilitate description of the invention and to simplify the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore should not be construed as limiting the invention. Furthermore, the terms "first," "second," "third," and the like are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.
The foregoing is merely illustrative of specific embodiments of the present invention, and the scope of the invention is not limited thereto, but any modifications, equivalents, improvements and alternatives falling within the spirit and principles of the present invention will be apparent to those skilled in the art within the scope of the present invention.
Claims (10)
1. The intelligent detection method for the multi-dimensional domain feature combined SAR vessel is characterized by comprising the following steps of:
analyzing multi-scale problems and rotation problems existing in the existing SAR vessel targets, acquiring layered space characteristics from bottom to top through a backbone network, and acquiring the multi-scale characteristics by adopting a characteristic pyramid FPN;
step two, the spatial multi-scale characteristics obtained in the step one are obtained through an RPN network, coordinate information and category information of candidate areas of SAR images are obtained, and the coordinate information is mapped to an original image and a multi-scale characteristic image respectively;
step three, processing is carried out by combining a time-frequency domain dimension domain aiming at the rotation behavior of the ship; obtaining rotation invariant information of the SAR vessel target by carrying out Fourier-polar coordinate transformation on the candidate region of the original image obtained in the second step, and obtaining the rotation invariant feature of the SAR vessel target frequency domain by feature extraction;
and step four, designing a finer feature fusion network aiming at space-frequency domain feature fusion, and obtaining more compact feature expression by interactively updating parameters of different networks so as to comprehensively characterize the features of each dimension domain.
2. The intelligent detection method for the multi-dimensional domain feature combined SAR vessel according to claim 1, wherein in the first step, the neural network Ω is convolved through a spatial channel S The spatial feature phi can be easily obtained from the SAR image S (x, y), which can be expressed as
φ S (x,y)=Ω S (I(x,y),θ s );
Wherein I (x, y) represents an input SAR vessel image, θ s Representing omega S Related parameter sets in (a); upsampling high-level features via feature pyramid FPN and enhancement via lateral join to obtain spatial multi-scale features φ S_MS (x,y)。
3. The intelligent detection method of multi-dimensional domain feature combined SAR vessel as set forth in claim 1, wherein in the second step, the proposed area of the image is obtained by sharing the spatial feature to the RPN, and the specific process is as follows:
the first step, a 3×3 convolution check is used to convolve the shared feature map, so as to make the extracted features more robust;
secondly, using a convolution kernel of 1 multiplied by 1 twice to obtain category information and position information of the candidate region;
third, mapping the candidate region position information to the original image and the multi-scale feature map to obtain I respectively Sub (x, y) and phi S_MS_Sub (x,y)。
4. The intelligent detection method for multi-dimensional domain feature combined SAR vessel as set forth in claim 3, wherein in said third step, the obtained original image candidate region I Sub (x, y) performing polar coordinate transformation, wherein the transformation relation is as follows:
where r represents the distance from the point (x, y) to the origin in a rectangular coordinate system, θ is the angle between the x-axis and the line connecting the point with the origin, and is positive in the counterclockwise direction.
5. The intelligent detection method for multi-dimensional domain feature combined SAR vessel as set forth in claim 4, wherein the obtained polar coordinate image is subjected to Fourier transform and the amplitude spectrum thereof is taken
Wherein F is Sub (mu, v) represents candidate region frequency domain information, mu, v are corresponding two-dimensional frequency domain variables respectively, and the amplitude spectrum is taken to obtain rotation invariant information M Sub (μ,ν)。
6. The intelligent detection method for the multi-dimensional domain feature combined SAR vessel according to claim 5, wherein the rotation invariant information M is obtained through convolution neural network pair Sub (μV) performing feature extraction
φ F_RI_Sub (x,y)=Ω F (M Sub (u,v),θ F );
In phi F_RI_Sub (x, y) represents the rotation invariant feature of the SAR vessel target, Ω F Extracting network for frequency domain characteristics, θ F For its associated parameter set.
7. The intelligent detection method of the multi-dimensional domain feature combined SAR vessel according to claim 1, wherein in the third step, the rotation behavior of the SAR vessel target is processed, and the rotation invariant feature is extracted, and the specific process is as follows:
the rotation behavior under the rectangular coordinates of the image is represented as translation of the target under the polar coordinates, so that the rotation transformation under the rectangular coordinates is converted into translation transformation by utilizing the polar coordinate transformation; and simultaneously, combining the characteristic of Fourier transformation, and taking the magnitude spectrum of the spectrum information of the SAR ship target so as to adapt to any rotation angle of the SAR ship target, thereby obtaining the rotation invariant information of the SAR ship target.
8. The intelligent detection method of the multi-dimensional domain feature combined SAR vessel according to claim 1, wherein in the third step, the rotation invariant feature of the SAR vessel target is obtained by feature extraction through a convolutional neural network.
9. The intelligent detection method of multi-dimensional domain feature combined SAR vessel as set forth in claim 6, wherein in the fourth step, a finer feature fusion network is designed, and the spatial-frequency domain features are fused by interactively updating parameters of different networks, wherein the specific process is as follows:
expanding the spatial features into one-dimensional vectors, and expanding the frequency domain features through the full connection layer to obtain feature vectors with the same dimension, wherein the obtained output feature size is 1024; the representation is
α S =FC(Flatten(φ S_MS_Sub (x,y)));
α F =Flatten(φ F_RI_Sub (x,y));
Wherein a is S And a F Output feature vectors representing the spatial domain and the frequency domain, respectively, and FC represents the fully connected layer.
10. The intelligent detection method of multi-dimensional domain feature combined SAR vessel according to claim 9, wherein the cross-domain fusion process of the obtained one-dimensional spatial feature and the frequency domain feature vector is:
α SF,1 =[FC 1 (α S ),FC 1 (α F )]
α SF,2 =[FC 2 (α F ),FC 2 (α S )]
where [.] means that the two feature vectors are stitched together and the intersection between the network and the features makes the output feature more compact while maintaining the features of the various dimensions.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011562298.7A CN112862748B (en) | 2020-12-25 | 2020-12-25 | Multi-dimensional domain feature combined SAR ship intelligent detection method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011562298.7A CN112862748B (en) | 2020-12-25 | 2020-12-25 | Multi-dimensional domain feature combined SAR ship intelligent detection method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112862748A CN112862748A (en) | 2021-05-28 |
CN112862748B true CN112862748B (en) | 2023-05-30 |
Family
ID=75997109
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011562298.7A Active CN112862748B (en) | 2020-12-25 | 2020-12-25 | Multi-dimensional domain feature combined SAR ship intelligent detection method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112862748B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115620172B (en) * | 2022-12-21 | 2023-04-11 | 中国电子科技集团公司第十四研究所 | Intelligent comprehensive identification method for marine ship target based on cross-domain multi-feature |
CN116778207B (en) * | 2023-06-30 | 2024-02-09 | 哈尔滨工程大学 | Unsupervised depth multi-scale SAR image change detection method based on spatial frequency domain |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105866778A (en) * | 2016-05-30 | 2016-08-17 | 西安电子科技大学 | Ship detection method based on polarized synthetic aperture radar data |
CN107274401A (en) * | 2017-06-22 | 2017-10-20 | 中国人民解放军海军航空工程学院 | A kind of High Resolution SAR Images Ship Detection of view-based access control model attention mechanism |
KR101925490B1 (en) * | 2018-08-01 | 2018-12-05 | 엘아이지넥스원 주식회사 | Bistatic synthetic aperture radar detection method based on global navigation satellite system |
CN112052899A (en) * | 2020-09-03 | 2020-12-08 | 上海无线电设备研究所 | Single ship target SAR image generation method based on generation countermeasure network |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8384583B2 (en) * | 2010-06-07 | 2013-02-26 | Ellegi S.R.L. | Synthetic-aperture radar system and operating method for monitoring ground and structure displacements suitable for emergency conditions |
-
2020
- 2020-12-25 CN CN202011562298.7A patent/CN112862748B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105866778A (en) * | 2016-05-30 | 2016-08-17 | 西安电子科技大学 | Ship detection method based on polarized synthetic aperture radar data |
CN107274401A (en) * | 2017-06-22 | 2017-10-20 | 中国人民解放军海军航空工程学院 | A kind of High Resolution SAR Images Ship Detection of view-based access control model attention mechanism |
KR101925490B1 (en) * | 2018-08-01 | 2018-12-05 | 엘아이지넥스원 주식회사 | Bistatic synthetic aperture radar detection method based on global navigation satellite system |
CN112052899A (en) * | 2020-09-03 | 2020-12-08 | 上海无线电设备研究所 | Single ship target SAR image generation method based on generation countermeasure network |
Also Published As
Publication number | Publication date |
---|---|
CN112862748A (en) | 2021-05-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Nie et al. | Inshore ship detection based on mask R-CNN | |
CN114202696B (en) | SAR target detection method and device based on context vision and storage medium | |
Cong et al. | RRNet: Relational reasoning network with parallel multiscale attention for salient object detection in optical remote sensing images | |
US11847805B2 (en) | Method for stitching images of capsule endoscope, electronic device and readable storage medium | |
CN110135366B (en) | Shielded pedestrian re-identification method based on multi-scale generation countermeasure network | |
Ozcelik et al. | Rethinking CNN-based pansharpening: Guided colorization of panchromatic images via GANs | |
Sun et al. | Global Mask R-CNN for marine ship instance segmentation | |
CN109584193A (en) | A kind of unmanned plane based on target preextraction is infrared and visible light image fusion method | |
CN112862748B (en) | Multi-dimensional domain feature combined SAR ship intelligent detection method | |
CN110135438B (en) | Improved SURF algorithm based on gradient amplitude precomputation | |
CN116228702A (en) | Camouflage target detection method based on attention mechanism and convolutional neural network | |
Toriya et al. | SAR2OPT: Image alignment between multi-modal images using generative adversarial networks | |
Long et al. | Underwater forward-looking sonar images target detection via speckle reduction and scene prior | |
Tang et al. | Sonar image mosaic based on a new feature matching method | |
Gao et al. | Multiscale curvelet scattering network | |
Sledge et al. | Target detection and segmentation in circular-scan synthetic aperture sonar images using semisupervised convolutional encoder–decoders | |
Wang et al. | Fused adaptive receptive field mechanism and dynamic multiscale dilated convolution for side-scan sonar image segmentation | |
CN117474765B (en) | DEM super-resolution reconstruction system based on reference image texture transfer | |
Quan et al. | Efficient and robust: A cross-modal registration deep wavelet learning method for remote sensing images | |
CN117788296A (en) | Infrared remote sensing image super-resolution reconstruction method based on heterogeneous combined depth network | |
CN113763274A (en) | Multi-source image matching method combining local phase sharpness orientation description | |
CN113886510A (en) | Terminal interaction method, device, equipment and storage medium | |
CN117893744A (en) | Remote sensing image segmentation method based on improved boundary guide context aggregation network | |
CN111275680A (en) | SAR image change detection method based on Gabor convolution network | |
Deka et al. | Joint sparse representation-based single image super-resolution for remote sensing applications |
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 |