CN112597815A - Synthetic aperture radar image ship detection method based on Group-G0 model - Google Patents

Synthetic aperture radar image ship detection method based on Group-G0 model Download PDF

Info

Publication number
CN112597815A
CN112597815A CN202011418137.0A CN202011418137A CN112597815A CN 112597815 A CN112597815 A CN 112597815A CN 202011418137 A CN202011418137 A CN 202011418137A CN 112597815 A CN112597815 A CN 112597815A
Authority
CN
China
Prior art keywords
ship
network
sar
loss
image
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
CN202011418137.0A
Other languages
Chinese (zh)
Inventor
文载道
王琳
陆昱廷
王小旭
潘泉
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.)
Northwestern Polytechnical University
Original Assignee
Northwestern Polytechnical 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 Northwestern Polytechnical University filed Critical Northwestern Polytechnical University
Priority to CN202011418137.0A priority Critical patent/CN112597815A/en
Publication of CN112597815A publication Critical patent/CN112597815A/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
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • 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
    • 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)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The invention discloses a synthetic aperture radar image ship detection method based on a Group-G0 model, which comprises the steps of obtaining a plurality of SAR images containing ship targets for training; wherein each ship in the SAR image is marked with a marking frame; training an SAR ship detection network through an SAR image containing a ship target; acquiring an SAR image to be detected, and extracting a characteristic map of the SAR image to be detected by using an SAR ship detection network; sequentially using an RPN (resilient packet network) and a Fast RCNN (Fast-forward neural network) to detect the characteristic graph to obtain an SAR image with a ship detection frame; according to the SAR ship detection network and the method, the SAR image which is marked with the mark frame and contains the ship target is used for training the SAR ship detection network, the network parameters of the SAR ship detection network are jointly optimized by combining the detection results of the model branch and the detection branch, the characteristic incidence relation of the region level information and the pixel level information is concerned, the optimization can be completed only by few training samples, the recall rate of ship detection can be improved, and the occurrence of missing detection is reduced.

Description

Synthetic aperture radar image ship detection method based on Group-G0 model
Technical Field
The invention belongs to the technical field of radar image ship detection, and particularly relates to a synthetic aperture radar image ship detection method based on a Group-G0 model.
Background
China is a maritime big country, ship target detection is an important link for implementing ocean monitoring, and sea area security is of great importance. A large number of ships move in and out in China sea area every day, so that ship target detection has important strategic significance on ocean monitoring, military reconnaissance, information acquisition and the like. The synthetic aperture radar is a high-resolution imaging radar, is an active microwave imaging sensor, is not influenced by environmental factors such as weather, light, cloud and the like in an imaging process compared with a traditional passive imaging sensor (such as an infrared sensor and an optical sensor), can work under all weather and all day time conditions to detect hidden objects, and is widely applied to the civil and military fields in recent years.
With the development of deep learning technology and computing power, the target detection algorithm based on deep learning has made a rapid progress. In recent years, a target detection algorithm based on deep learning is gradually applied to the field of SAR image ship detection, and has achieved great success. However, the detection method needs more training samples to train the detection model, so that the ship detection recall rate is low, and the detection omission phenomenon frequently occurs.
Disclosure of Invention
The invention aims to provide a synthetic aperture radar image ship detection method based on a Group-G0 model, so as to improve the recall rate of ship detection and reduce the occurrence of missed detection.
The invention adopts the following technical scheme: a synthetic aperture radar image ship detection method based on a Group-G0 model comprises the following steps:
acquiring a plurality of SAR images containing ship targets for training; wherein each ship in the SAR image is marked with a marking frame;
training an SAR ship detection network through an SAR image containing a ship target;
acquiring an SAR image to be detected, and extracting a characteristic map of the SAR image to be detected by using an SAR ship detection network;
and sequentially using the RPN and Fast RCNN to detect the characteristic diagram to obtain the SAR image with the ship detection frame.
Further, training the SAR ship detection network through the SAR image containing the ship target comprises the following steps:
carrying out multi-scale feature extraction on the SAR image containing the ship target through an SAR ship detection network to obtain a feature map;
classifying and regressing the characteristic diagram by sequentially using an RPN network and a Fast RCNN network;
and optimizing the parameters of the SAR ship detection network by combining a second loss function according to the classification and regression results to obtain the trained SAR ship detection network.
Further, obtaining the feature map further includes:
generating a hidden variable characteristic diagram with the same size as the SAR image by adopting an encoder consisting of a convolutional neural network;
carrying out binarization on the hidden variable characteristic graph to generate a ship type mask graph and a non-ship type mask graph;
multiplying the ship-type occultation map and the non-ship-type occultation map with the SAR image respectively, and inputting the multiplied images into a decoder consisting of a convolutional neural network;
decoding the multiplied image through a decoder to obtain a G0 distribution parameter value of each pixel point;
and alternately optimizing the SAR ship detection network based on the classification and regression result, the second loss function, the G0 distribution parameter value and the first loss function to obtain the trained SAR ship detection network.
Further, classifying and regressing the feature map by using the RPN network and the Fast RCNN network in sequence comprises:
predicting a plurality of candidate target areas which are possibly foreground in the SAR image by adopting an RPN (resilient packet network) based on the characteristic diagram;
and classifying and regressing the candidate target area based on the Fast RCNN network to obtain a target frame of a ship in the SAR image.
Further, the first loss function is:
Figure BDA0002820918550000031
wherein x isiA certain pixel point in the SAR image; z is a radical ofkIs a pixel point xiThe corresponding category hidden variable has the value range of {0, 1}, z11 and z2When equal to 0, it represents pixel point xiIs a non-ship type pixel point, z10 and z1When 1, the pixel point x is representediThe pixel points are ship-like pixel points; n is a radical ofkThe sample number N of two types of pixel points in the SAR image1The number of ship-like pixel points N in the SAR image2The number of non-ship type pixel points alpha in the SAR imageiDenotes a shape parameter, gammaiScale parameters are represented, and Gamma is a Gamma function.
Further, the second loss function is:
Figure BDA0002820918550000032
wherein, Loss _2 represents a second Loss function, which comprises two parts of Loss functions of Loss _ RPN and Loss _ FastRCCNN of the RPN network, Loss _ RPN _ cla represents the classification Loss of the RPN network, Loss _ RPN _ loc represents the regression Loss of the RPN network, Loss _ Fast _ RCNN _ cla represents the classification Loss of the Fast RCNN network, and Loss _ Fast _ RCNN _ cla represents the regression Loss of the Fast RCNN network.
The invention has the beneficial effects that: according to the SAR ship detection network and the method, the SAR image which is marked with the mark frame and contains the ship target is used for training the SAR ship detection network, the network parameters of the SAR ship detection network are jointly optimized by combining the detection results of the model branch and the detection branch, the characteristic incidence relation of the region level information and the pixel level information is concerned, the optimization can be completed only by few training samples, the recall rate of ship detection can be improved, and the occurrence of missing detection is reduced.
Drawings
FIG. 1 is a flowchart of a synthetic aperture radar image ship detection method based on a Group-G0 model according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a training process of the SAR ship detection network in the embodiment of the present invention;
FIG. 3 is a network architecture diagram of a SAR ship detection network in an embodiment of the present invention;
FIG. 4 is a diagram of a network architecture of model branches in an embodiment of the present invention;
FIG. 5 is a comparison chart of the effects of the method of the present invention and the FPN method in the verification embodiment of the present invention.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
In recent years, a large number of researchers have studied SAR image ship detection, which is represented by a Constant False Alarm Rate (CFAR) detection method and a target detection method based on deep learning.
The CFAR ship detection method is a classic target detection algorithm and is widely applied due to the advantages of simplicity and strong self-adaptability. The CFAR detection method comprises two stages of modeling and adaptive threshold acquisition.
Firstly, sea clutter in an image is modeled, and common distribution models such as Gaussian (Gaussian) distribution, Gamma (Gamma) distribution, K distribution, W-sharp distribution, G0 distribution and the like are adopted, wherein the G0 distribution has good fitting effect on homogeneous, non-homogeneous and extremely-heterogeneous sea clutter and is widely applied by students.
And then, sliding a window with a fixed size in the image, modeling pixel points in the window area to obtain statistical model distribution of the area, and then setting a constant false alarm rate to obtain a target pixel threshold value of each area.
Although CFAR and the improved method thereof make good progress in the SAR image application field, the method constructs a statistical distribution model based on pixel points, and judges whether the target is the target or not pixel point by pixel point according to a threshold value, and the defect of semantic feature constraint based on the regional level often causes that the ship detection false alarm rate is overhigh.
With the continuous development of the field of artificial intelligence, the deep learning algorithm is gradually applied to SAR image ship detection tasks. Among them, Convolutional Neural Network (CNN) is widely used in the field of deep learning by virtue of its powerful characterization capability and feature extraction capability. The CNN-based ship detection method aims to obtain the category (i.e., ship class or non-ship class) and position coordinates of a target (i.e., each pixel point), and is generally divided into a one-stage detection method and a two-stage detection method.
The first-stage ship detection algorithm is represented by a YOLO series algorithm, the image is divided into grids, and target detection is realized by predicting confidence coefficient of target occurrence in the grids and relative coordinates of grid units; the two-stage detection algorithm is represented by Fast RCNN, and after extracting features of a picture, candidate Region object frames are extracted in advance through an RPN (Region suggestion Network), and then classified and regressed through the Fast RCNN Network to perform object detection. Although the two types of classical algorithms have succeeded in the application of ship target detection, the detection algorithms based on the CNN only extract the characteristics of the region level according to the manually marked labels, and lack the constraint based on the pixel point statistical model, so that on one hand, the detection results are missed and the recall rate is low, and on the other hand, the model learned by the neural network lacks the interpretability, the generalization ability is weak and the robustness is poor.
Although the traditional CFAR method and the target detection method based on deep learning have achieved certain success in the application of ship target detection. However, the performance of the CFAR detection algorithm depends on the accuracy of background clutter modeling, and although some learners propose complex distributions (such as K distribution, G0 distribution, and the like) to solve the difficulty of non-homogeneous sea clutter modeling, the complexity of the statistical model expression makes the traditional parameter estimation method (such as maximum likelihood estimation, EM estimation, and the like) unable to obtain an analytic solution, which becomes one of the reasons for restricting the CFAR detection algorithm. In addition, the CFAR algorithm models the pixel points in a certain area in the SAR image, and then calculates the pixel threshold for distinguishing the target and the background of the area according to the preset false alarm rate of the root domain to detect the ship, so that the CFAR algorithm focuses on the mining of the statistical distribution characteristics of the pixel points, and the target detection result is often caused to generate a large amount of false alarms due to the lack of complementary information of the area characteristics. In the deep learning based target detection algorithm, feature acquisition is based on CNN. On one hand, no matter one-stage detection or two-stage detection is carried out, the characteristic extraction process utilizes a learnable filter to extract the characteristics of the region level, the inherent statistical distribution characteristic information of the pixel level is ignored to a certain extent, the small-scale target is easy to miss detection, and the recall rate is too low. On the other hand, the learning process of the model based on the CNN depends on the number of manually labeled labels, and if the number of labels is too small, the process is likely to fall into overfitting.
Therefore, the SAR image ship detection method adopts a (pixel-region) double-current network architecture with a model and region feature jointly optimized, multi-scale region feature extraction is carried out on the SAR image, meanwhile, modeling is carried out on SAR image pixel points by utilizing an SAR image statistical model based on Group-G0 distribution, and parameter distribution of the pixel point statistical model is adaptively learned by utilizing an encoder-decoder structure, so that the SAR image ship detection under model and data dual drive is realized.
The embodiment of the invention discloses a synthetic aperture radar image ship detection method based on a Group-G0 model, which comprises the following steps:
s110, acquiring a plurality of SAR images containing ship targets for training; wherein each ship in the SAR image is marked with a marking frame; s120, training an SAR ship detection network through an SAR image containing a ship target; s130, acquiring an SAR image to be detected, and extracting a characteristic diagram of the SAR image to be detected by using an SAR ship detection network; s140, detecting the characteristic graph by sequentially using the RPN and Fast RCNN to obtain the SAR image with the ship detection frame.
According to the SAR ship detection network and the method, the SAR image which is marked with the mark frame and contains the ship target is used for training the SAR ship detection network, the network parameters of the SAR ship detection network are jointly optimized by combining the detection results of the model branch and the detection branch, the characteristic incidence relation of the region level information and the pixel level information is concerned, the optimization can be completed only by few training samples, the recall rate of ship detection can be improved, and the occurrence of missing detection is reduced.
In the embodiment of the present invention, as shown in fig. 2 and fig. 3, after the bottom layer feature extraction is implemented by using the feature pyramid network, the bottom layer feature extraction is sent to the model branch and the detection branch. As shown in fig. 4, parameters of a statistical model obeyed by each pixel point are adaptively learned by using model branches, and the regional level characteristics of the SAR image are mined by using the detection branch depth. And sharing and interacting the bottom layer characteristics of the model branch and the detection branch, jointly optimizing and solving the characteristic subspace of the target, and realizing SAR image target detection by utilizing the target characteristics of the optimization solution.
The model branch focuses on parameter estimation of a pixel statistical model, and an encoder is responsible for learning hidden variable characteristics from bottom layer characteristics, wherein the hidden variables are used for describing category information of image pixels, namely the hidden variables control the generation process of the image. For example, changing the information of the hidden variable (category) changes the category of the pixel point in the input image. And the decoding is responsible for reconstructing the input image statistical distribution parameters by the hidden variables obtained by encoding.
The statistical distribution model of the SAR image comprises Gaussian distribution, gamma distribution, Wishart distribution, G0 distribution and the like, and because G0 distribution has strong inclusion compared with other distributions and has good fitting effect on homogeneous, non-homogeneous and extremely-heterogeneous areas in the SAR image, the invention adopts G0-based distribution modeling. Based on ship detection tasks, the method hopes to focus on a statistical model of a target, and a single G0 cannot accurately model an image, so that an implicit variable z is introduced, pixels of the whole SAR image are divided into two types, and the two types of pixels are distributed according to two groups of G0.
The SAR image pixel level is considered to have semantic features, namely each pixel point has category information, for example, the pixel category of a ship region is ship, the pixel category of a sea island region is land, and the like. For the SAR image target detection task, only whether each pixel point is a ship or not is concerned, so that category information is introduced in the image modeling process as a hidden variable of a model, and the category is set to be ship and background (namely non-ship). For simplicity, the G0 distribution with the introduced category hidden variables is named as a Group-G0 distribution model, and the modeling and parameter estimation process of the SAR image based on the Group-G0 is described in detail below.
Specifically, the G0 statistical distribution model is as follows:
Figure BDA0002820918550000081
wherein n is an equivalent view and can be obtained through SAR image prior knowledge, and alphai、γiIs a pixel point xiAnd carrying out parameter self-adaptive estimation on the corresponding scale parameter and the corresponding shape parameter through a neural network.
In addition, because samples with the same category have approximately the same statistical distribution characteristics, in the modeling process of the whole SAR image, the pixel information of the same category is modeled into a G0 statistical distribution cluster, and the pixels in each G0 distribution cluster are independent of each other, and the specific distribution is as follows:
Figure BDA0002820918550000082
wherein x isiA certain pixel point in the SAR image; z is a radical ofkIs a pixel point xiThe corresponding category hidden variable has the value range of {0, 1}, z11 and z2When equal to 0, it represents pixel point xiIs a non-ship type pixel point, z10 and z1When 1, the pixel point x is representediThe pixel points are ship-like pixel points; n is a radical ofkThe number of samples of two types of pixel points in the SAR image is shown.
N1The number of ship-like pixel points N in the SAR image2The number of non-ship type pixel points alpha in the SAR imageiDenotes a shape parameter, gammaiRepresenting scale parameters, Gamma representing a Gamma function。
In the model branch, an encoder formed by a convolutional neural network is adopted to generate a hidden variable characteristic diagram with the same size as the SAR image; carrying out binarization on the hidden variable characteristic graph to generate a ship type mask graph and a non-ship type mask graph; multiplying the ship-type occultation map and the non-ship-type occultation map with the SAR image respectively, and inputting the multiplied images into a decoder consisting of a convolutional neural network; decoding the multiplied image through a decoder to obtain a G0 distribution parameter value of each pixel point; and alternately optimizing the SAR ship detection network based on the classification and regression result, the second loss function, the G0 distribution parameter value and the first loss function to obtain the trained SAR ship detection network.
The detailed structure of the model branch parameter estimation network is shown in fig. 4 and mainly comprises an encoder and a decoder, wherein the encoder encodes the SAR image into a class diagram with semantic information through a convolutional neural network, and the decoding process decodes the SAR image into a statistical distribution parameter diagram related to G0 distribution through the network according to an encoding result.
The encoding process of the model is first introduced. The input image size is H multiplied by W multiplied by 1, and multi-scale features P2, P3, P4 and P5 are obtained through a feature pyramid network based on Resnet 50. And after upsampling, each scale feature is overlapped and fused with the adjacent feature on the upper layer until the size of the original image is upsampled. Specifically, after the P5 features are extracted from the convolutional layer, the P5 features are sent to the deconvolution layer to be up-sampled to the size of P4 after batch normalization operation, the feature graph obtained by up-sampling is added with P4 to obtain a fused feature graph P4 ', P4' and the steps are continuously repeated until the feature graph with the same channel number as that of the original graph as 1 is coded, a coding probability graph is obtained through an activation function Sigmoid, and the probability of each pixel point represents the probability value that the pixel point is a ship. Two class mask maps m1, m2, both H × W × 1, are obtained by binarizing the probability map. Wherein the probability threshold is set to 0.5, those greater than 0.5 are warships, those less than 0.5 are non-warships, and m1 and m2 are in a complementary relationship. By this point, the encoding process ends.
The decoding process takes the result of multiplying the original image by the mask images m1 and m2 obtained by the encoder as the input of the mask network, and predicts the parameter image of G0 distribution of each pixel point through the decoding network, wherein the size of the parameter image is H multiplied by W multiplied by 2, and the number of channels is 2, and the parameters alpha and gamma of G0 distribution are respectively represented. The specific details of the decoding network are set up with two layers of convolution and normalization, as shown in fig. 4.
In the invention, a neural network is adopted to adaptively estimate the shape parameter alpha and the scale parameter gamma of each pixel point in the image. Specifically, after the SAR image is input into the network, the backbone network extracts the underlying shared features, and then sends the extracted features into the model branch and the detection branch, respectively. After the model branches are sent, firstly, a hidden variable feature map with the same size as the original image is obtained through an encoder formed by a convolutional network, namely a category map consisting of z corresponding to each pixel point x, binarization is carried out on the category map to obtain mask maps m1 and m2 of a ship and a non-ship, wherein the mask represents high-level semantic information of an input image and reflects whether each pixel point is a target or not. In the decoding process, after the mask images m1 and m2 obtained by coding are multiplied by the trained SAR image respectively, two images are obtained, then the two images are input into a decoder consisting of a convolution network, G0 statistical characteristic parameters obeyed by the input image are reconstructed, log alpha is obtained, and therefore estimation of the G0 statistical distribution parameters of the pixel points is completed.
The shape parameter alpha and the scale parameter gamma are used for optimizing the decoding process to further influence the bottom pyramid characteristics, the first loss function comprises the shape parameter alpha and the scale parameter gamma, the bottom characteristics are continuously optimized in the process of returning through the loss function, and the bottom characteristics are continuously optimized in the process of estimating the pixel-level statistical distribution parameters, so that the bottom characteristics comprise pixel-level information.
The detection branch focuses on the region level target feature extraction. Specifically, multi-scale feature extraction is carried out on an SAR image containing a ship target through an SAR ship detection network to obtain a feature map; classifying and regressing the characteristic diagram by sequentially using an RPN network and a Fast RCNN network; and optimizing the parameters of the SAR ship detection network by combining a second loss function according to the classification and regression results to obtain the trained SAR ship detection network. In this embodiment, the SAR ship detection network specifically employs a feature pyramid network.
The invention adopts a two-stage detection algorithm FPN (feature Pyramid Networks for Object detection), which mainly comprises two sub-Networks, namely an RPN (region pro-polar network) network and a Fast RCNN network.
Specifically, based on a characteristic diagram extracted by a RAM ship detection network, an RPN is adopted to predict a plurality of candidate target areas which are possible to be foreground in an SAR image; and classifying and regressing the candidate target area based on the Fast RCNN network to obtain a target frame of a ship in the SAR image.
More specifically, a plurality of target frames are predicted in the feature map based on the RPN, and the offset of the target frames relative to the anchor frame and the probability that the target frames are foreground are calculated; determining a ship prediction target frame in the SAR image according to the probability that the target frame is a foreground; acquiring a corresponding region in the SAR image corresponding to the ship prediction target frame, and classifying the corresponding region based on a Fast RCNN network to obtain a detection frame and a regression value of the ship in the SAR image.
In the first stage, the region suggestion network RPN network extracts regions of interest (Propusals) according to the bottom layer features, and foreground and background classification and coarse regression are achieved.
And in the second stage, performing region-of-interest feature extraction (ROI Align) operation on the extracted Propusals by using a Fast RCNN network, performing fine regression and classification, and finally obtaining coordinate information and a category of a target in an image, thereby realizing SAR ship detection.
After the bottom-layer shared features are input into the RPN network, firstly, the offset (dx, dy, dw, dh) of the target frame relative to the Anchor and the probability that the target frame is foreground are predicted based on the Anchor frame (Anchor) designed in advance, and in the embodiment, 2000 candidate target frames with the highest probability are selected according to the target frame probability and the strategies such as non-maximum suppression (NMS).
And the Fast RCNN network cuts out the bottom layer characteristics corresponding to the region of interest and normalizes the bottom layer characteristics to a uniform size according to the target frames selected in the first stage through cutting operation in the bottom layer characteristics, then the characteristic diagrams of the target frames are sent into a convolution network in the second stage to carry out secondary regression and classification, finally the category and the coordinate of the target are obtained, and SAR target detection is finished.
After the network is built, the network needs to be optimized, and the optimization of the model branch and the optimization of the detection branch are introduced respectively.
1) Model branch optimization strategy
Since the SAR image in the experimental data set is a single view (SLC), the view n is set to 1. z is an implicit variable corresponding to x, wherein the x is a non-ship type pixel point represented by the x being 0, the x is a ship type pixel point represented by the x being 1, the alpha represents a shape parameter corresponding to x, and the gamma represents a scale parameter corresponding to x.
According to the maximum likelihood optimization criterion, it is desirable to maximize logP (x | z), and therefore the loss function of the model branch (i.e., the first loss function) is:
Figure BDA0002820918550000121
wherein the content of the first and second substances,
Figure BDA0002820918550000122
thus, the overall optimization function of the model branches is:
Figure BDA0002820918550000123
2) detection branch optimization strategy
The Loss functions of the discriminant branches include RPN network Loss and Fast RCNN network Loss. The RPN network comprises a regression Loss and a classification Loss, wherein the regression Loss adopts Smooth L1 Loss, and the classification Loss adopts cross entropy Loss (Crossentcopy Loss). The regression Loss and classification Loss functions of the Fast RCNN are the same as those of the RPN, namely the second Loss function:
Figure BDA0002820918550000124
wherein, Loss _2 represents a second Loss function, which comprises two parts of Loss functions of Loss _ RPN and Loss _ FastRCCNN of the RPN network, Loss _ RPN _ cla represents the classification Loss of the RPN network, Loss _ RPN _ loc represents the regression Loss of the RPN network, Loss _ Fast _ RCNN _ cla represents the classification Loss of the Fast RCNN network, and Loss _ Fast _ RCNN _ cla represents the regression Loss of the Fast RCNN network.
Therefore, the total optimized objective function in this embodiment is Loss _ total ═ Loss _1+ Loss _2, the network is optimized by using an Adam optimizer, and the parameters are set as defaults.
In conclusion, the SAR ship detection is realized by adopting a Group-G0-based target detection method, SAR images are modeled by introducing category hidden variables, and detection branches based on pixel parameter estimation of a neural network are constructed, so that the detection branches and the bottom layer characteristics of the model branches are shared, and the bottom layer characteristic subspace is continuously and alternately optimized under the combined action of the region-level detection branches and the pixel-level model branches, thereby realizing the extraction of interpretable bottom layer characteristics, and completing the SAR ship detection based on the characteristic subspace.
In addition, the method introduces a model parameter estimation branch, so that the neural network focuses more on the statistical model characteristic of the target in the process of learning the optimal solution, and the overfitting phenomenon easily caused by the fact that the original discriminant model only depends on a label to search a target feature subspace is effectively avoided, so that the generalization capability and the robustness of the model are greatly improved.
The invention (double-current network architecture) is different from a mainstream target detection algorithm which relies on a manual label, and only carries out region level feature extraction on a target by using CNN. The invention provides a Group-G0 SAR image statistical distribution model, and a Group-G0 statistical distribution model is constructed by introducing category hidden variables into the SAR image pixel point statistical distribution model. In addition, the design model branch utilizes a neural network to realize the adaptive estimation of the Group-G0 distribution parameters.
The invention adopts a feature learning framework with model data dual drive, so that the bottom layer shared features are continuously optimized under the combined action of the pixel level feature information of the model branches and the region level pixel information of the detection branches, thereby effectively improving the recall rate of the detection algorithm and greatly reducing the omission factor.
In addition, due to the addition of the model branch of the parameter self-adaptive estimation, the target feature subspace extracted by the network does not only depend on the manually labeled label, and the statistical model of the target obedience is concerned more, so that the interpretability of the feature extracted by the network is enhanced. The present invention has unique advantages for solving the problem of small samples. The small sample problem is that when the number of labels in the training set is extremely limited, the mainstream deep learning target detection algorithm can be over-fitted by trapping. In the invention, the SAR image data characteristics are modeled, and SAR image data statistical model branches are introduced, so that the bottom layer characteristics are constrained around the statistical model distribution, and the overfitting phenomenon in the training process is avoided, thereby greatly enhancing the generalization capability of the model and effectively solving the problem of small samples.
Verification of the examples:
in order to illustrate the effectiveness of the method for solving the SAR image thumbnail problem, 5 scene large graphs shot by a high-resolution three-number camera over 5 cities are selected as a data set of the experiment, and 5 pairs of large graphs are segmented into 29399 small graphs with the size of 1024 × 768. 4644 images are randomly selected from the 29399 training set as a small sample training set, so as to verify the advantages of the method in solving the problem of the small sample. The test set selects 1 scene from 5 scenes as the test set, and is divided into 8889 test charts, and the experimental result comparison indexes of the method provided by the invention and other target detection methods are shown in table 1.
Figure BDA0002820918550000141
Figure BDA0002820918550000151
In order to verify the efficiency of the method, the invention is compared with the main target detection algorithms of fast RCNN, YOLOV3, FCOS, FPN of the target mainstream. Table 1 shows the experimental results, and it can be seen from the table that even for a small sample dataset, under the condition that the number of manually labeled labels is very limited, the detection accuracy of the method is still higher than that of other algorithms, which is sufficient to show that the method has a good advantage in solving the problem of the small sample and shows the generalization of the algorithm.
In addition, the method and the FPN method are used for detection and comparison in a scene with strong speckle noise, the comparison result is shown in fig. 5, the ship target can not be easily detected by the FPN algorithm in the scene with strong speckle noise, and only partial targets at the edge can not be detected. Compared with the original FPN, the method provided by the invention has good performance on ship recall performance in both near shore and ocean scenes. Specific quantitative performance indexes are shown in table 1, and compared with the FPN algorithm, the recall rate of the detection algorithm provided by the invention is improved by 4.21 percentage points, which reflects that the detection algorithm provided by the invention is good in reducing target detection false alarm. The accuracy index Precision is observed to be reduced, because after the model branch is added, compared with the previous extraction of only the focus region features, the bottom-layer feature extraction network needs to be responsible for the mining of pixel-level information at the same time, and some false alarms can occur in the ship detection result. The overall performance of the detector is measured by the MAP index, and the overall MAP index of the method provided by the invention is improved by 6 percent.

Claims (6)

1. A synthetic aperture radar image ship detection method based on a Group-G0 model is characterized by comprising the following steps:
acquiring a plurality of SAR images containing ship targets for training; wherein each ship in the SAR image is marked with a marking frame;
training a SAR ship detection network through the SAR image containing the ship target;
acquiring an SAR image to be detected, and extracting a characteristic map of the SAR image to be detected by using the SAR ship detection network;
and sequentially using an RPN (resilient packet network) and a Fast RCNN (Fast-forward neural network) to detect the characteristic graph to obtain an SAR image with a ship detection frame.
2. The Group-G0 model-based synthetic aperture radar image ship detection method of claim 1, wherein training a SAR ship detection network through the SAR image containing a ship target comprises:
carrying out multi-scale feature extraction on the SAR image containing the ship target through an SAR ship detection network to obtain a feature map;
classifying and regressing the characteristic diagram by sequentially using an RPN network and a Fast RCNN network;
and optimizing the parameters of the SAR ship detection network by combining a second loss function according to the classification and regression results to obtain the trained SAR ship detection network.
3. The Group-G0 model-based synthetic aperture radar image ship detection method according to claim 2, wherein the obtaining of the feature map further comprises:
generating a hidden variable characteristic diagram with the same size as the SAR image by adopting an encoder consisting of a convolutional neural network;
carrying out binarization on the hidden variable feature map to generate a ship type mask map and a non-ship type mask map;
multiplying the ship occultation map and the non-ship occultation map with the SAR image respectively, and inputting the multiplied products to a decoder consisting of a convolutional neural network;
decoding the multiplied image through the decoder to obtain a G0 distribution parameter value of each pixel point;
and alternately optimizing the SAR ship detection network based on the classification and regression result, the second loss function, the G0 distribution parameter value and the first loss function to obtain the trained SAR ship detection network.
4. The Group-G0 model-based synthetic aperture radar image ship detection method according to claim 3, wherein the classifying and regressing the feature map sequentially using RPN network and Fast RCNN network comprises:
predicting a plurality of candidate target areas which are possibly foreground in the SAR image by adopting an RPN (resilient packet network) based on the characteristic map;
classifying and regressing the candidate target area based on a Fast RCNN network to obtain a target frame of a ship in the SAR image.
5. The Group-G0 model-based synthetic aperture radar image ship detection method of claim 4, wherein the first loss function is:
Figure FDA0002820918540000021
wherein x isiA certain pixel point in the SAR image; z is a radical ofkIs a pixel point xiThe corresponding category hidden variable has the value range of {0, 1}, z11 and z2When equal to 0, it represents pixel point xiIs a non-ship type pixel point, z10 and z1When 1, the pixel point x is representediThe pixel points are ship-like pixel points; n is a radical ofkThe sample number N of two types of pixel points in the SAR image1The number of ship-like pixel points N in the SAR image2The number of non-ship type pixel points alpha in the SAR imageiDenotes a shape parameter, gammaiScale parameters are represented, and Gamma is a Gamma function.
6. The Group-G0 model-based synthetic aperture radar image ship detection method of claim 5, wherein the second loss function is:
Figure FDA0002820918540000031
wherein, Loss _2 represents a second Loss function, which comprises two parts of Loss functions of Loss _ RPN and Loss _ FastRCCNN of the RPN network, Loss _ RPN _ cla represents the classification Loss of the RPN network, Loss _ RPN _ loc represents the regression Loss of the RPN network, Loss _ Fast _ RCNN _ cla represents the classification Loss of the Fast RCNN network, and Loss _ Fast _ RCNN _ cla represents the regression Loss of the Fast RCNN network.
CN202011418137.0A 2020-12-07 2020-12-07 Synthetic aperture radar image ship detection method based on Group-G0 model Pending CN112597815A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011418137.0A CN112597815A (en) 2020-12-07 2020-12-07 Synthetic aperture radar image ship detection method based on Group-G0 model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011418137.0A CN112597815A (en) 2020-12-07 2020-12-07 Synthetic aperture radar image ship detection method based on Group-G0 model

Publications (1)

Publication Number Publication Date
CN112597815A true CN112597815A (en) 2021-04-02

Family

ID=75188899

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011418137.0A Pending CN112597815A (en) 2020-12-07 2020-12-07 Synthetic aperture radar image ship detection method based on Group-G0 model

Country Status (1)

Country Link
CN (1) CN112597815A (en)

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113111758A (en) * 2021-04-06 2021-07-13 中山大学 SAR image ship target identification method based on pulse neural network
CN113627356A (en) * 2021-08-13 2021-11-09 上海无线电设备研究所 Intelligent sea surface ship identification method
CN113627310A (en) * 2021-08-04 2021-11-09 中国电子科技集团公司第十四研究所 Background and scale perception SAR ship target detection method
CN113657252A (en) * 2021-08-16 2021-11-16 福州大学 Efficient SAR image ship target detection method based on codec
CN113657181A (en) * 2021-07-23 2021-11-16 西北工业大学 SAR image rotating target detection method based on smooth label coding and feature enhancement
CN113723182A (en) * 2021-07-21 2021-11-30 西安电子科技大学 SAR image ship detection method under limited training sample condition
CN113837377A (en) * 2021-08-30 2021-12-24 厦门大学 Neural network pruning method based on class mask
CN113962883A (en) * 2021-09-30 2022-01-21 西南电子技术研究所（中国电子科技集团公司第十研究所） Method for enhancing ship target detection SAR image data
CN114241307A (en) * 2021-12-09 2022-03-25 中国电子科技集团公司第五十四研究所 Synthetic aperture radar aircraft target identification method based on self-attention network
CN115661500A (en) * 2022-12-27 2023-01-31 南京邮电大学 Target detection method based on second-order distribution and uncertainty perception cluster fusion

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113111758B (en) * 2021-04-06 2024-01-12 中山大学 SAR image ship target recognition method based on impulse neural network
CN113111758A (en) * 2021-04-06 2021-07-13 中山大学 SAR image ship target identification method based on pulse neural network
CN113723182A (en) * 2021-07-21 2021-11-30 西安电子科技大学 SAR image ship detection method under limited training sample condition
CN113657181B (en) * 2021-07-23 2024-01-23 西北工业大学 SAR image rotation target detection method based on smooth tag coding and feature enhancement
CN113657181A (en) * 2021-07-23 2021-11-16 西北工业大学 SAR image rotating target detection method based on smooth label coding and feature enhancement
CN113627310B (en) * 2021-08-04 2023-11-24 中国电子科技集团公司第十四研究所 SAR ship target detection method based on background and scale sensing
CN113627310A (en) * 2021-08-04 2021-11-09 中国电子科技集团公司第十四研究所 Background and scale perception SAR ship target detection method
CN113627356A (en) * 2021-08-13 2021-11-09 上海无线电设备研究所 Intelligent sea surface ship identification method
CN113627356B (en) * 2021-08-13 2024-02-23 上海无线电设备研究所 Sea surface ship intelligent identification method
CN113657252B (en) * 2021-08-16 2023-07-28 福州大学 Efficient SAR image ship target detection method based on encoding and decoding device
CN113657252A (en) * 2021-08-16 2021-11-16 福州大学 Efficient SAR image ship target detection method based on codec
CN113837377A (en) * 2021-08-30 2021-12-24 厦门大学 Neural network pruning method based on class mask
CN113962883A (en) * 2021-09-30 2022-01-21 西南电子技术研究所（中国电子科技集团公司第十研究所） Method for enhancing ship target detection SAR image data
CN113962883B (en) * 2021-09-30 2023-02-03 西南电子技术研究所（中国电子科技集团公司第十研究所） Method for enhancing ship target detection SAR image data
CN114241307A (en) * 2021-12-09 2022-03-25 中国电子科技集团公司第五十四研究所 Synthetic aperture radar aircraft target identification method based on self-attention network
CN115661500A (en) * 2022-12-27 2023-01-31 南京邮电大学 Target detection method based on second-order distribution and uncertainty perception cluster fusion

Similar Documents

Publication Publication Date Title
CN112597815A (en) Synthetic aperture radar image ship detection method based on Group-G0 model
CN109934200B (en) RGB color remote sensing image cloud detection method and system based on improved M-Net
US10198657B2 (en) All-weather thermal-image pedestrian detection method
CN111914686B (en) SAR remote sensing image water area extraction method, device and system based on surrounding area association and pattern recognition
CN113591968A (en) Infrared weak and small target detection method based on asymmetric attention feature fusion
CN113449660B (en) Abnormal event detection method of space-time variation self-coding network based on self-attention enhancement
CN110991257B (en) Polarized SAR oil spill detection method based on feature fusion and SVM
CN107424175B (en) Target tracking method combined with space-time context information
CN110705412A (en) Video target detection method based on motion history image
CN110084181B (en) Remote sensing image ship target detection method based on sparse MobileNet V2 network
CN114255403A (en) Optical remote sensing image data processing method and system based on deep learning
CN115984714B (en) Cloud detection method based on dual-branch network model
CN111666801A (en) Large-scene SAR image ship target detection method
CN113743505A (en) Improved SSD target detection method based on self-attention and feature fusion
Zhang et al. YoloXT: A object detection algorithm for marine benthos
CN114926826A (en) Scene text detection system
CN117218545A (en) LBP feature and improved Yolov 5-based radar image detection method
Yang et al. SAR image target detection and recognition based on deep network
CN115953312A (en) Joint defogging detection method and device based on single image and storage medium
Yao et al. A Forecast-Refinement Neural Network Based on DyConvGRU and U-Net for Radar Echo Extrapolation
Zhang et al. Research on rainy day traffic sign recognition algorithm based on PMRNet
Li et al. A fast detection method for polynomial fitting lane with self-attention module added
CN113449611B (en) Helmet recognition intelligent monitoring system based on YOLO network compression algorithm
CN114973390B (en) Complex background infrared weak target detection method combined with eye movement attention mechanism
Li et al. A-YOLO: small target vehicle detection based on improved YOLOv5

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