WO2020037960A1 - 一种sar目标识别方法、装置、计算机设备及存储介质 - Google Patents

一种sar目标识别方法、装置、计算机设备及存储介质 Download PDF

Info

Publication number
WO2020037960A1
WO2020037960A1 PCT/CN2019/075175 CN2019075175W WO2020037960A1 WO 2020037960 A1 WO2020037960 A1 WO 2020037960A1 CN 2019075175 W CN2019075175 W CN 2019075175W WO 2020037960 A1 WO2020037960 A1 WO 2020037960A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
network model
pictures
sar
residual neural
Prior art date
Application number
PCT/CN2019/075175
Other languages
English (en)
French (fr)
Inventor
徐雪菲
廖斌
张安国
万环
肖鹏
魏通
Original Assignee
深圳大学
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 深圳大学 filed Critical 深圳大学
Publication of WO2020037960A1 publication Critical patent/WO2020037960A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/10Terrestrial scenes
    • G06V20/13Satellite images
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks

Definitions

  • the present application relates to the field of image processing technology, and in particular, to a SAR target recognition method, device, computer equipment, and storage medium.
  • the technical problem to be solved in the present application is to provide a method, device, computer equipment, and storage medium for SAR target recognition in view of the above-mentioned shortcomings of the prior art. And the model recognition accuracy is low.
  • a SAR target recognition method wherein the recognition method is implemented based on a deep learning network and includes:
  • the SAR image to be detected is input to the trained residual neural network model for detection and recognition, and the recognition result is output.
  • the SAR target recognition method wherein the acquiring a SAR original image sample specifically includes:
  • the SAR image original sample includes multiple recognition targets, and each recognition target corresponds to multiple cropped original grayscale pictures.
  • the SAR target recognition method wherein the performing data enhancement on the SAR original image samples to generate an expanded sample set specifically includes:
  • the denoised samples and the noised samples form an expanded sample set.
  • a value range of the residual control factor is -0.5 to 0.5.
  • the method randomly extracts a certain number of pictures from the augmented sample set, and inputs the extracted pictures to the optimized residual neural network model for training, after obtaining the training
  • the residual neural network model includes:
  • the accuracy requirements are not met, the number of samples in the expanded sample set needs to be further increased, and the steps of constructing, optimizing and training the residual neural network model are re-executed;
  • the pictures randomly selected from the expanded sample set include original grayscale pictures corresponding to each recognition target, and the number of original grayscale pictures for each recognition target is the same.
  • the method randomly extracts a certain number of pictures from the augmented sample set, and inputs the extracted pictures to the optimized residual neural network model for training, after obtaining the training
  • the residual neural network model also includes:
  • a second number of pictures are randomly extracted from the original sample of the SAR image, and the extracted pictures are input to the basic residual neural network model and the optimized residual neural network model for training to obtain the initial model accuracy.
  • the accuracy requirement is that the accuracy of the current model is higher than the accuracy of the initial model by 2%.
  • a SAR target recognition device wherein the device includes:
  • Original sample acquisition module used to obtain SAR original image samples
  • a sample data enhancement module configured to perform data enhancement on the SAR original image samples to generate an expanded sample set
  • a model construction and optimization module for constructing a basic residual neural network model, adding a residual control factor to the basic residual neural network model for optimization, and constructing an optimized residual neural network model;
  • a model training module configured to randomly extract a certain number of pictures from the augmented sample set, and input the extracted pictures into the optimized residual neural network model for training to obtain a trained residual neural network model;
  • An image recognition module is configured to input a SAR image to be detected into a trained residual neural network model for detection and recognition, and output a recognition result.
  • a computer device includes a memory and a processor, and the memory stores a computer program, wherein when the processor executes the computer program, the steps of any one of the foregoing methods are implemented.
  • a storage medium stores a computer program thereon, wherein when the computer program is executed by a processor, the steps of the method according to any one of the foregoing are implemented.
  • This application expands the training samples and constructs a residual neural network model with residual control factors.
  • the expanded training samples can meet the needs of the residual neural network model, effectively reducing the overfitting.
  • Situation, and the residual neural network model added with the residual control factor can improve the convergence speed during the training process, shorten the model training time, and then improve the efficiency and accuracy of target recognition.
  • FIG. 1 is a schematic flowchart of a SAR target recognition method according to an embodiment of the present application.
  • FIG. 2 is a schematic diagram of a unit structure of a basic residual neural network model constructed in an embodiment of the present application.
  • FIG. 3 is a schematic diagram of a unit structure of a residual neural network model in which a residual control factor is added in an embodiment of the present application.
  • FIG. 4 is a schematic structural diagram of an optimized residual neural network model in an embodiment of the present application.
  • FIG. 5 is a schematic structural diagram of a quick connection unit in a residual neural network model according to an embodiment of the present application.
  • FIG. 6 is a system block diagram of a SAR target recognition method according to an embodiment of the present application.
  • FIG. 7 is a structural block diagram of a SAR target recognition device in an embodiment of the present application.
  • FIG. 8 is an internal structure diagram of a computer device in an embodiment of the present application.
  • the application research of machine learning theory in radar target detection, classification and recognition has great potential.
  • the main object of detection and estimation theory is the radar echo signal matrix after pulse compression, such as the research on target detection and estimation theory based on Bayesian theory;
  • the main research object of target classification and recognition is synthetic aperture radar SAR (Synthetic Aperture Radar) complex matrix of imaging results, such as target classification and recognition based on convolutional neural networks.
  • Deep learning theory has powerful feature extraction capabilities. Its application in SAR image processing is mainly divided into the following two aspects according to the radar system and background: SAR / ISAR (Inverse Synthetic Aperture Radar) image processing Target classification and recognition based on Polarimetric Synthetic Aperture Radar (Polarimetric Synthetic Aperture Radar) image processing, research shows that, compared with traditional pre-designed, inelastic feature extraction systems, deep neural network systems are processing It has obvious advantages in the fields of optical images, acoustic signals, and machine translation.
  • the SAR target recognition method provided in this application can be applied to a terminal.
  • the terminal may be, but is not limited to, various personal computers, notebook computers, mobile phones, tablet computers, in-vehicle computers, and portable wearable devices.
  • the terminal of the present invention uses a multi-core processor.
  • the processor of the terminal may be at least one of a central processing unit (Central Processing Unit, CPU), a graphics processor (Graphics Processing Unit, GPU), and a video processing unit (Video Processing Unit (VPU)).
  • CPU Central Processing Unit
  • GPU Graphics Processing Unit
  • VPU Video Processing Unit
  • a SAR target recognition method is provided.
  • the method is applied to the above terminal as an example for description, and includes the following steps;
  • Step S100 Obtain a SAR raw image sample.
  • the present application selects multiple imaging results from MSTAR data (MSTAR mixed target data includes slice images of a group of military targets) and rotates 360 ° at a certain elevation angle, for example, selecting MSTAR data to synthesize Aperture radar imaging results at 360 ° rotation at 15 ° and 17 °. Then the format conversion software is used to convert the imaging results into JPG original grayscale images, and then all original grayscale images are centered, and 100 pixels are taken in both the horizontal and vertical directions for cropping. From this, a sample with a size of 100 * 100 pixels is obtained, which is the original sample of the SAR image.
  • MSTAR mixed target data includes slice images of a group of military targets
  • the SAR image original sample in the present application includes multiple recognition targets, and each recognition target corresponds to multiple cropped original grayscale pictures.
  • common ground military weapons can be selected from MSTAR data as targets, and the targets are different types of tanks and armored vehicles: T62, BRDM2, BTR-6, 2S1, D7, ZIL131, ZSU-234, and T72, respectively. -A04.
  • the number of original grayscale images corresponding to the eight types of recognition targets is: T62: 572, BRDM2: 572, BTR-60: 451, 2S1: 573, D7: 573, ZIL131: 573, ZSU -234: 573, T72-A04: 573, a total of 4,460 original grayscale pictures, which are cropped to form the original sample of SAR image.
  • Step 1 Place the original grayscale images of the eight types of recognition targets T62, BRDM2, BTR-60, 2S1, D7, ZIL131, ZSU-234, T72-A04 in different folders and name them.
  • Step 2 Set the MATLAB file path to these eight recognition targets, and set this operation as a large loop.
  • Step 3 Set up a small loop.
  • I imread (imgpath)
  • I an original grayscale picture of a recognition target
  • imgpath is a folder path.
  • step 4 a small loop is performed to crop the original gray image in each of the identified target folders.
  • rect [ ⁇ ] to set the reserved pixel range
  • execute rect [col, row, spacing-1, spacing-1]
  • execute newI imcrop (I, rect) to crop the picture
  • newI is the cropped picture result.
  • Step 6 Repeat steps 3, 4 and 5 until all the original grayscale pictures of the eight recognition targets are trimmed and saved.
  • Step S200 Perform data enhancement on the SAR original image samples to generate an expanded sample set.
  • the model is prone to overfitting.
  • This application mainly starts from the two aspects of increasing noise and reducing noise, and performs data enhancement on the original image samples to obtain noise-enhanced and de-noised samples. In order to achieve the expansion of the sample set.
  • all of the original grayscale pictures after being cropped in the SAR image are filtered in three different smoothing dimensions to obtain three sets of denoising samples with different parameters.
  • the filter smoothing dimensions are set to 3 * 3, 5 * 5, and 7 * 7, and 4,460 original grayscale pictures are filtered by using filters, three different sets of parameters can be obtained.
  • noisy samples For obtaining noise-enhanced samples, this embodiment generates three sets of speckle-noise image sets with different parameters of 0.5, 1.0, and 1.5 for all cropped original grayscale images in the original sample of the SAR image.
  • the speckle noise picture sets with different group parameters are respectively multiplied with the original samples of the SAR image to obtain three sets of noise samples.
  • Three sets of denoised samples and three sets of noisy samples obtained after data enhancement processing form an expanded sample set. It can be seen that the present invention can expand the original data by 6 times by performing data enhancement on the original samples of the SAR image, thereby greatly increasing the number of training samples.
  • the specific steps for obtaining the denoised samples are as follows:
  • Step 1 Read the cropped original grayscale images of the eight identified targets.
  • Step 4 Change the large loop parameter T to 5 and 7, and execute the large loop to obtain two sets of denoised samples.
  • Step 1 Read the cropped original reply pictures of the eight recognition targets.
  • the parameters MU of the large loop are changed to 1.0 and 1.5, and the large loop is executed to obtain another two sets of samples with increased speckle noise.
  • Step S300 Construct a basic residual neural network model, and add a residual control factor to the basic residual neural network model for optimization, and construct an optimized residual neural network model.
  • the recognition of SAR targets basically uses a deep convolutional network model. This network model has a slow training convergence rate. At the same time, due to the limit on the number of samples, it will also encounter overfitting problems, resulting in the network model layer. The number is limited, so the learning depth is limited.
  • a residual neural network model is used in this application, and a residual control factor is introduced to optimize the residual neural network model.
  • this embodiment first constructs a basic residual neural network model, as follows:
  • First build a basic basic unit including: the first layer is a 3 * 3 convolution kernel layer, followed by a BN (batch normalization) layer, followed by an activation function ReLU (rectified linear unit), and then the same 3 * 3
  • the structure of the basic basic unit is shown on the left in FIG. 2.
  • the first layer is a 3 * 3 convolution layer, followed by a BN (batch normalization) layer, followed by an activation function ReLU (rectified linear unit), and then the same 3 * 3
  • the convolution layer of the convolution kernel a BN (batch normalization) layer, and an activation function ReLU (rectified linear unit)
  • link the head and tail of this stacked neural unit to add a 1 * 1 convolution kernel convolution layer.
  • a BN layer merged with two convolutional layers, and then activate the function through a ReLU and output.
  • the structure of the Basic unit is shown on the right side of FIG. 2.
  • the network model includes a total of 20 convolutional layers, which are in turn: "input ⁇ convolutional layer + basic * 3 + basic * inc + basic *" 2 + basic inc + basic * 2 + fully connected layer ⁇ output ".
  • a residual control factor is added at the end of the quick connection to optimize the basic residual neural network model, as follows:
  • the first layer is a 3 * 3 convolution kernel layer, followed by a BN (batch normalization) layer, and then an activation function ReLU (rectified linear unit)
  • the convolution layer of the same 3 * 3 convolution kernel, a BN (batch normalization) layer, and an activation function ReLU (rectified linear unit) are implemented.
  • the residual control factor Cr is added. It is then linked at the beginning and end of this stacked neural unit, merged with the two convolutional layers, and then activated by a ReLU function and output.
  • the structure of the IB unit is shown on the left side of FIG. 3.
  • a propagation TB (transformational block) unit including: the first layer is a 3 * 3 convolution layer, followed by a BN (batch normalization) layer, followed by an activation function ReLU (rectified linear unit), Then follow the same convolution layer of 3 * 3 convolution kernel, a BN (batch normalization) layer, and an activation function ReLU (rectified linear unit).
  • the residual control factor Cr is added.
  • link the head and tail of this stacked neural unit add a 1 * 1 convolution kernel convolution layer, followed by a BN layer, then merge with two convolution layers, and then activate the function through a ReLU and output.
  • the structure of the TB unit is shown on the right side in FIG. 3.
  • the IB unit and the BN unit are stacked and combined to construct the optimized residual neural network in this application:
  • the network model contains a total of 20 convolutional layers, which are "input + convolutional layer + IB * 3 + TB + IB * 2" + TB + IB * 2 + pooling layer + dense layer + output ”.
  • the optimized residual neural network model is shown in Figure 4.
  • a quick connection unit is used in the residual neural network model in this application, compared with the traditional convolutional neural network, a deeper network structure can be realized.
  • a residual control factor is added to the tail of the quick connection unit.
  • the convergence speed can be increased during the training process, and the model training time is greatly shortened.
  • the value of the residual control factor ranges from -0.5 to 0.5, and the purpose of increasing the control factor is to improve the model training convergence speed.
  • the optimized residual neural network model constructed in this embodiment is further described here.
  • the illustrated optimized residual neural network model includes a 20-layer structure from top to bottom:
  • the first layer is a convolution layer, which is used to convolve the input picture of 100 * 100 * 1 * n, n is the number of sample inputs, and the size of the convolution kernel window is 3 * 3, the step size is 1, and the output 16 feature maps, the output image size is 100 * 100;
  • the second layer to the third layer are an IB structural unit.
  • the convolution layer has a convolution kernel of 3 * 3, the step size is 2, and 16 feature maps are also output, and the output image size is 100 * 100;
  • the fourth layer-the fifth layer and the sixth layer-the seventh layer are stacked by two identical IB units.
  • Each convolution layer has a convolution kernel of 3 * 3, the step length is 1, and 16 feature maps are output.
  • the image size is 100 * 100;
  • the eighth to ninth layers are a TB structural unit.
  • the convolution layer convolution kernel 3 * 3 outputs 32 feature maps with a step size of 2 and an output image size of 50 * 50;
  • the tenth layer-the eleventh layer and the twelfth layer-the thirteenth layer are two identical IB unit stacks, each convolution layer has a convolution kernel of 3 * 3, the step length is 2, and 32 feature maps are output.
  • the output image size is 50 * 50;
  • the fourteenth to fifteenth layers are a TB structural unit.
  • the convolution layer convolution kernel 3 * 3 outputs 64 feature maps with a step size of 2 and an output image size of 25 * 25;
  • the sixteenth layer-the seventeenth layer and the eighteenth layer-the nineteenth layer are two identical IB unit stacks, each convolution layer has a convolution kernel of 3 * 3, the step size is 2, and 64 feature maps are output. , The output image size is 25 * 25;
  • the twentieth layer is a fully connected layer composed of 8 neurons.
  • the above layers are connected together to form an optimized residual neural network model for SAR target recognition.
  • Step S400 Randomly extract a certain number of pictures from the extended sample set, and input the extracted pictures into the optimized residual neural network model for training to obtain a trained residual neural network model.
  • the optimized residual neural network needs to be trained so that the network model meets the accuracy requirements of SAR recognition.
  • a first number of pictures are randomly selected from the extended sample, and the extracted pictures are input to the basic residual neural network model and the optimized residual neural network model for training to obtain the current model accuracy. Then compare the accuracy of the current model with the accuracy of the initial model to determine whether the accuracy of the current model meets the accuracy requirements. If the accuracy requirements are met, a residual neural network model trained with the first number of pictures is output. The residual neural network The model can then be used to identify SAR images. If the accuracy requirements are not met, the number of samples in the expanded sample set needs to be further increased, and the steps of constructing, optimizing, and training the residual neural network model are re-executed so that the trained residual neural network mechanism can meet the requirements.
  • the pictures randomly selected from the extended sample set include original grayscale pictures corresponding to each recognition target, and the number of original grayscale pictures for each recognition target is the same.
  • the accuracy requirement in this embodiment is that the accuracy of the current model is higher than the accuracy of the initial model by 2%.
  • the accuracy of the initial model is obtained by randomly extracting a second number of pictures from the original sample of the SAR image, and inputting the extracted pictures to the basic residual neural network model and the optimized residual neural network model after training. owned.
  • the network model is trained to obtain the initial model accuracy, which is mainly used to compare with the current model accuracy.
  • Step S500 Input the SAR image to be detected into the trained residual neural network model for detection and recognition, and output the recognition result.
  • the SAR image to be detected is input to the trained residual neural network model for detection and recognition, and the residual neural network model will classify and identify the SAR image. To output the recognition result.
  • the accuracy of the trained residual neural network model in this application has been improved from 94.58% to 99.65%, and the model operation speed, compared with the original residual structure without increasing the "residual control factor"
  • the convergence time of the model of this application is reduced from 1200 minutes to 300 minutes, which effectively improves the computing efficiency.
  • steps in the flowchart of FIG. 1 are sequentially displayed in accordance with the directions of the arrows, these steps are not necessarily performed sequentially in the order indicated by the arrows. Unless explicitly stated in this document, the execution of these steps is not strictly limited, and these steps can be performed in other orders. Moreover, at least a part of the steps in FIG. 1 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but may be executed at different times. The execution of these sub-steps or stages The sequence is not necessarily performed sequentially, but may be performed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.
  • FIG. 6 a system block diagram of the SAR target recognition method of the present application is provided.
  • the sample to be tested that is, the SAR original image sample
  • the training samples need to undergo two types of data enhancement processing, that is, data enhancement processing 1 (denoising processing) to obtain denoised samples and data enhancement processing 2 (noising processing) to obtain noise-enhanced samples, thereby forming extended samples.
  • data enhancement processing 1 denoising processing
  • data enhancement processing 2 noisesing processing
  • a SAR target recognition device which includes: an original sample acquisition module 710, a sample data enhancement module 720, a model construction and optimization module 730, a model training module 740, and an image recognition module 750. among them,
  • An original sample acquisition module 710 configured to acquire a SAR original image sample
  • a sample data enhancement module 720 configured to perform data enhancement on the SAR original image samples to generate an expanded sample set
  • a model construction and optimization module 730 is configured to construct a basic residual neural network model, and add a residual control factor to the basic residual neural network model for optimization to construct an optimized residual neural network model;
  • a model training module 740 is configured to randomly extract a certain number of pictures from the expanded sample set, and input the extracted pictures to the optimized residual neural network model for training to obtain a trained residual neural network model;
  • An image recognition module 750 is configured to input a SAR image to be detected into a trained residual neural network model for detection and recognition, and output a recognition result.
  • the original sample acquisition module 7107 specifically expands and selects multiple imaging results from the MSTAR data at a certain elevation angle and rotates 360 °; the formatted software is used to convert the imaging results into JPG Format the original grayscale pictures; take all the original grayscale pictures with the center position as the reference, and take 100 pixels in both the horizontal and vertical directions to crop to form the original sample of the SAR image; It includes multiple recognition targets, and each recognition target corresponds to multiple cropped original grayscale pictures.
  • the sample data enhancement module 720 specifically includes performing filtering processing on all original grayscale pictures in the original samples of the SAR image in three different smoothing dimensions to obtain three sets of denoised samples with different parameters; All original grayscale pictures in the original sample of the SAR image are respectively generated into three sets of speckle noise picture sets with different parameters of 0.5, 1.0, and 1.5; and the three sets of speckle noise picture sets with different parameters are respectively associated with the SAR image.
  • the original samples are multiplied to obtain three sets of noise samples; the denoised samples and the noise samples form an expanded sample set.
  • the value of the residual control factor ranges from -0.5 to 0.5.
  • the model training module 740 specifically includes randomly extracting a first number of pictures from the extended sample, and inputting the extracted pictures to a basic residual neural network model and an optimized residual neural network, respectively.
  • the model is trained to obtain the current model accuracy; the current model accuracy is compared with the initial model accuracy to determine whether the current model accuracy meets the accuracy requirements; if the accuracy requirements are met, the first number of pictures after training are output Residual neural network model; if it does not meet the accuracy requirements, the number of samples in the expanded sample set needs to be further increased, and the steps of constructing, optimizing and training the residual neural network model are re-executed; pictures randomly drawn from the expanded sample set The original grayscale pictures corresponding to each recognition target are included, and the number of original grayscale pictures for each recognition target is the same.
  • the model training module 740 further includes: randomly extracting a second number of pictures from the SAR image original sample, and inputting the extracted pictures to the basic residual neural network model and the optimized residual neural network, respectively.
  • the model is trained to obtain the initial model accuracy.
  • the accuracy requirement is that the accuracy of the current model is higher than the accuracy of the initial model by 2%.
  • Each module of the above-mentioned SAR target recognition device can be realized in whole or in part by software, hardware, and a combination thereof.
  • the above-mentioned modules may be embedded in the hardware form or independent of the processor in the computer device, or may be stored in the memory of the computer device in the form of software, so that the processor calls and performs the operations corresponding to the above modules.
  • a computer device is provided.
  • the computer device may be a terminal, and its internal structure diagram may be as shown in FIG. 8.
  • the computer equipment includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus.
  • the processor of the computer device is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system and a computer program.
  • the internal memory provides an environment for running an operating system and computer programs in a non-volatile storage medium.
  • the network interface of the computer equipment is used to communicate with external terminals through a network connection.
  • the computer program is executed by a processor to implement a method for detecting the color of an image light source.
  • the display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen.
  • the input device of the computer device may be a touch layer covered on the display screen, or a button, a trackball, or a touchpad provided on the computer device casing. , Or an external keyboard, trackpad, or mouse.
  • FIG. 8 is only a block diagram of a part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer equipment to which the scheme of the present application is applied.
  • the specific computer equipment may be Include more or fewer parts than shown in the figure, or combine certain parts, or have a different arrangement of parts.
  • a computer device including a memory and a processor.
  • the memory stores a computer program
  • the processor implements the following steps when the computer program is executed:
  • the SAR image to be detected is input to the trained residual neural network model for detection and recognition, and the recognition result is output.
  • obtaining the SAR raw image sample specifically includes: selecting a plurality of imaging results from the MSTAR data at a certain elevation angle and rotating 360 °; using format conversion software to convert the imaging results into a JPG format Original grayscale pictures; take all the original grayscale pictures to the center position, and take 100 pixels in both horizontal and vertical directions to crop to form an original sample of SAR image; the original sample of SAR image includes There are multiple recognition targets, and each recognition target corresponds to multiple cropped original grayscale pictures.
  • performing data enhancement on the SAR original image samples and generating the extended sample set specifically includes: performing filtering processing on all original grayscale pictures in the SAR image original samples in three different smoothing dimensions, Three sets of denoising samples with different parameters are obtained; three sets of speckle noise pictures with different average values of 0.5, 1.0, and 1.5 are generated for all original grayscale pictures in the original sample of the SAR image; The different speckle noise picture sets are respectively multiplied with the original samples of the SAR image to obtain three sets of noise samples; the denoised samples and the noise samples form an expanded sample set.
  • the value of the residual control factor ranges from -0.5 to 0.5.
  • a certain number of pictures are randomly selected from the above-mentioned expanded sample set, and the extracted pictures are input to the optimized residual neural network model for training.
  • the trained residual neural network model includes: A first number of pictures are randomly selected from the expanded sample, and the extracted pictures are input to the basic residual neural network model and the optimized residual neural network model for training to obtain the current model accuracy; the current model accuracy is obtained; Compare with the accuracy of the initial model to determine whether the current model accuracy meets the accuracy requirements; if it meets the accuracy requirements, output the residual neural network model trained with the first number of pictures; if it does not meet the accuracy requirements, you need to further Increase the number of samples in the expanded sample set, and re-execute the steps of constructing, optimizing, and training the residual neural network model; the pictures randomly extracted from the expanded sample set include the original grayscale pictures corresponding to each recognition target, and each The number of original gray-scale pictures of each recognition target is the same.
  • a certain number of pictures are randomly selected from the above-mentioned expanded sample set, and the extracted pictures are input to the optimized residual neural network model for training.
  • the trained residual neural network model further includes: random Extract a second number of pictures from the original sample of the SAR image, and input the extracted pictures to the basic residual neural network model and the optimized residual neural network model for training to obtain the initial model accuracy.
  • the accuracy requirement is that the accuracy of the current model is higher than the accuracy of the initial model by 2%.
  • the present application further provides a computer-readable storage medium having a computer program stored thereon.
  • the computer program When the computer program is executed by a processor, the following steps are implemented:
  • the SAR image to be detected is input to the trained residual neural network model for detection and recognition, and the recognition result is output.
  • obtaining the SAR raw image sample specifically includes: selecting a plurality of imaging results from the MSTAR data at a certain elevation angle and rotating 360 °; using format conversion software to convert the imaging results into a JPG format Original grayscale pictures; take all the original grayscale pictures with the center position as the reference, and take 100 pixels in both the horizontal and vertical directions to crop to form an original sample of the SAR image; the original sample of the SAR image includes There are multiple recognition targets, and each recognition target corresponds to multiple cropped original grayscale pictures.
  • performing data enhancement on the SAR original image samples and generating the extended sample set specifically includes: performing filtering processing on all original grayscale pictures in the SAR image original samples in three different smoothing dimensions, Three sets of denoising samples with different parameters are obtained; three sets of speckle noise pictures with different average values of 0.5, 1.0, and 1.5 are generated for all original grayscale pictures in the SAR image original sample; The different speckle noise picture sets are respectively multiplied with the original samples of the SAR image to obtain three sets of noise samples; the denoised samples and the noise samples form an expanded sample set.
  • the value of the residual control factor ranges from -0.5 to 0.5.
  • a certain number of pictures are randomly selected from the above-mentioned expanded sample set, and the extracted pictures are input to the optimized residual neural network model for training.
  • the trained residual neural network model includes: A first number of pictures are randomly selected from the expanded sample, and the extracted pictures are input to the basic residual neural network model and the optimized residual neural network model for training to obtain the current model accuracy; the current model accuracy is obtained; Compare with the accuracy of the initial model to determine whether the current model accuracy meets the accuracy requirements; if it meets the accuracy requirements, output the residual neural network model trained with the first number of pictures; if it does not meet the accuracy requirements, you need to further Increase the number of samples in the expanded sample set, and re-execute the steps of constructing, optimizing, and training the residual neural network model; the pictures randomly extracted from the expanded sample set include the original grayscale pictures corresponding to each recognition target, and each The number of original gray-scale pictures of each recognition target is the same.
  • a certain number of pictures are randomly selected from the above-mentioned expanded sample set, and the extracted pictures are input to the optimized residual neural network model for training.
  • the trained residual neural network model further includes: random Extract a second number of pictures from the original sample of the SAR image, and input the extracted pictures to the basic residual neural network model and the optimized residual neural network model for training to obtain the initial model accuracy.
  • the accuracy requirement is that the accuracy of the current model is higher than the accuracy of the initial model by 2%.
  • Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory can include random access memory (RAM) or external cache memory.
  • RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Synchlink DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
  • SRAM static RAM
  • DRAM dynamic RAM
  • SDRAM synchronous DRAM
  • DDRSDRAM dual data rate SDRAM
  • ESDRAM enhanced SDRAM
  • SLDRAM synchronous chain Synchlink DRAM
  • Rambus direct RAM
  • DRAM direct memory bus dynamic RAM
  • RDRAM memory bus dynamic RAM

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

一种SAR目标识别方法、装置、计算机设备及存储介质,方法包括:获取SAR原始图像样本(S100);对SAR原始图像样本进行数据增强,生成扩充样本集(S200);构建基础残差神经网络模型,并在基础残差神经网络模型中加入残差控制因子进行优化,构建出优化后的残差神经网络模型(S300);从扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型(S400);将待检测的SAR图像输入至训练后的网络模型中进行识别,输出识别结果(S500)。该方法有效降低了过拟合的情况,并且加入残差控制因子的残差神经网络模型可以在训练的过程中提高收敛速度,使得模型训练时间缩短,提高了目标识别的效率及精度。

Description

一种SAR目标识别方法、装置、计算机设备及存储介质 技术领域
本申请涉及图像处理技术领域,尤其涉及的是一种SAR目标识别方法、装置、计算机设备及存储介质。
背景技术
目前,合成孔径雷达(Synthetic Aperture Radar,SAR)目标识别的需求越来越明显,专家和学者也都提出了众多的算法来提高目标识别精度。
但是在传统的雷达目标识别方法中,其主要的问题是缺乏训练样本而导致模型过拟合。虽然现有技术中也存在对训练样本进行扩充的技术,但是现有技术中也仅仅只考虑了部分因素,并不能从多方面进行样本扩充,因此训练样本数量仍然受限。此外,传统的识别方法中不管是基于模板还是基于模型,对先验知识和模型精度要求普遍较高,灵活性和适应性较差,识别准确性受限于模型的可靠性与特征提取的准确度,因此识别精度不高。
因此,现有技术还有待于改进和发展。
发明内容
本申请要解决的技术问题在于,针对现有技术的上述缺陷,提供一种SAR目标识别方法、装置、计算机设备及存储介质,旨在解决现有技术中的SAR目标识别方法中训练样本数量受限,且模型识别精度低的问题。
本申请解决技术问题所采用的技术方案如下:
一种SAR目标识别方法,其中,所述识别方法是基于深度学习网络所实现的,包括:
获取SAR原始图像样本;
对所述SAR原始图像样本进行数据增强,生成扩充样本集;
构建基础残差神经网络模型,并在所述基础残差神经网络模型中加入残差控制因子进行优化,构建出优化后的残差神经网络模型;
从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型;
将待检测的SAR图像输入至训练后的残差神经网络模型中进行检测识别,输出识别结果。
优选地,所述的SAR目标识别方法,其中,所述获取SAR原始图像样本具体包括:
从MSTAR数据中选取多个在一定角度的俯仰角下并旋转360°的成像结果;
利用格式转换软件将所述成像结果转换成JPG格式的原始灰度图片;
将所有的原始灰度图片以中心位置为基准,并在横向和纵向两个方向上取100个像素点进行裁剪,形成SAR图像原始样本;
所述SAR图像原始样本中包括有多个识别目标,且每个识别目标对应有多张经过裁剪的原始灰度图片。
优选地,所述的SAR目标识别方法,其中,所述对所述SAR原始图像样本进行数据增强,生成扩充样本集具体包括:
对所述SAR图像原始样本中的所有经过裁剪之后原始灰度图片分别进行三种不同平滑维度的滤波处理,得到三组参数不同的去噪样本;
对所述SAR图像原始样本中的所有经过裁剪之后原始灰度图片分别生成均值为0.5、1.0以及1.5的三组参数不同的斑点噪声图片集;
将所述三组参数不同的斑点噪声图片集分别与所述SAR图像原始样本相乘,得到三组加噪样本;
所述去噪样本与所述加噪样本组成扩充样本集。
优选地,所述的SAR目标识别方法,其中,所述残差控制因子的取值范围为-0.5~0.5。
优选地,所述的SAR目标识别方法,其中,所述从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型具体包括:
从所述扩充样本中随机抽取第一数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到当前模型精度;
将所述当前模型精度与初始模型精度进行比较,判断所述当前模型精度是否符合精度要求;
若符合精度要求,则输出以第一数量的图片进行训练后的残差神经网络模型;
若不符合精度要求,则需进一步增加扩充样本集中的样本数量,并重新执行残差神经网络模型的构建、优化以及训练的步骤;
从所述扩充样本集中随机抽取的图片中包括有各个识别目标对应的原始灰度图片,且每个识别目标的原始灰度图片的数量相同。
优选地,所述的SAR目标识别方法,其中,所述从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型还包括:
随机从SAR图像原始样本抽取第二数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到初始模型精度。
优选地,所述的SAR目标识别方法,其中,所述精度要求为当前模型精度比初始模型精度高于2%。
一种SAR目标识别装置,其中,所述装置包括:
原始样本获取模块,用于获取SAR原始图像样本;
样本数据增强模块,用于对所述SAR原始图像样本进行数据增强,生成扩充样本集;
模型构建且优化模块,用于构建基础残差神经网络模型,并在所述基础残差神经网络模型中加入残差控制因子进行优化,构建出优化后的残差神经网络模型;
模型训练模块,用于从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型;
图像识别模块,用于将待检测的SAR图像输入至训练后的残差神经网络模型中进行检测识别,输出识别结果。
一种计算机设备,包括存储器和处理器,所述存储器存储有计算机程序,其中,所述处理器执行所述计算机程序时实现上述任一项所述方法的步骤。
一种存储介质,其上存储有计算机程序,其中,所述计算机程序被处理器执行时实现上述中任一项所述的方法的步骤。
本申请的有益效果:本申请通过对训练样本进行扩充,并构建具有残差控制因子的残差神经网络模型,扩充的训练样本可以满足残差神经网络模型的需求,有效降低了过拟合的情况,而加入残差控制因子的残差神经网络模型可以在训练 的过程中提高收敛速度,使得模型训练的时间缩短,进而提高目标识别的效率及精度。
附图说明
图1是本申请一个实施例中的SAR目标识别方法的流程示意图。
图2是本申请一个实施例中构建的基础残差神经网络模型的单元结构示意图。
图3是本申请一个实施例中增加残差控制因子的残差神经网络模型的单元结构示意图。
图4是本申请一个实施例中优化后的残差神经网络模型的结构示意图。
图5是本申请一个实施例中残差神经网络模型中的快捷连接单元结构示意图。
图6是本申请一个实施例中的SAR目标识别方法的***框图。
图7本申请一个实施例中的SAR目标识别装置的结构框图。
图8是本申请一个实施例中计算机设备的内部结构图。
具体实施方式
为使本申请的目的、技术方案及优点更加清楚、明确,以下参照附图并举实施例对本申请进一步详细说明。应当理解,此处所描述的具体实施例仅仅用以解释本申请,并不用于限定本申请。
目前,机器学习理论在雷达目标的检测、分类与识别方面的应用研究有很大潜力。其中,检测与估计理论主要处理对象是经过脉冲压缩的雷达回波信号矩阵,例如基于贝叶斯理论的目标检测估计理论研究;目标的分类与识别主要研究对象为是合成孔径雷达SAR(Synthetic Aperture Radar)成像结果的复矩阵,例如基于卷积神经网络的目标分类与识别。
深度学习理论具有强大的特征提取能力,其在SAR图像处理方面的应用,根据雷达体制与背景不同,主要分为以下两个方面:SAR/ISAR(Inverse Synthetic Aperture Radar,逆合成孔径雷达)图像处理的目标分类与识别以及基于极化合成孔径雷达PolSAR(Polarimetric Synthetic Aperture Radar)图像处理的目标分类与识别,研究表明,相对于传统预设计、非弹性化特征提取的***,深度神经网络***在处理光学图像方面、声学信号方面以及机器翻译等领域具有明显优势。
因此,为了解决现有技术中SAR目标识别方法中训练样本数量受限,且模型识别精度低的问题,本申请提供一种SAR目标识别方法,该方法是基于深度学***板电脑、车载电脑和便携式可穿戴设备。本发明的终端采用多核处理器。其中,终端的处理器可以为中央处理器(Central Processing Unit,CPU),图形处理器(Graphics Processing Unit,GPU)、视频处理单元(Video Processing Unit,VPU)等中的至少一种。
具体地,在其中一个实施例中,如图1所示,提供一种SAR目标识别方法,该方法应用于上述终端为例进行说明,包括以下步骤;
步骤S100、获取SAR原始图像样本。
在其中一个实施例中,本申请从MSTAR数据(MSTAR混合目标数据包含一组军事目标的切片图像)中选取多个在一定角度的俯仰角下并旋转360°的成像结果,例如选取MSTAR数据合成孔径雷达俯仰角分别为15°和17°下,旋转360°的成像结果。然后利用格式转换软件将所述成像结果转换成JPG格式的原始灰度图片,再将所有的原始灰度图片以中心位置为基准,并在横向和纵向两个方向上取100个像素点进行裁剪,由此得到数据集为100*100像素大小的样本,即为SAR图像原始样本。
进一步地,本申请中的SAR图像原始样本中包括有多个识别目标,且每个识别目标对应有多张经过裁剪的原始灰度图片。例如,本实施例中可以从MSTAR数据中选取常见地面军事武器为识别目标,目标为不同型号的坦克和装甲车:分别为T62,BRDM2,BTR-6,2S1,D7,ZIL131,ZSU-234,T72-A04。那这8种识别目标所对应的原始灰度图片数量即为:T62:572张,BRDM2:572张,BTR-60:451张,2S1:573张,D7:573张,ZIL131:573张,ZSU-234:573张,T72-A04:573张,共计4,460张原始灰度图片,经裁剪后形成SAR图像原始样本。
进一步地,本实施例中对于原始灰度图片的裁剪的具体步骤如下:
第1步,将8种识别目标T62,BRDM2,BTR-60,2S1,D7,ZIL131,ZSU-234,T72-A04的原始灰度图片放置在不同的文件夹并命名。
第2步,将MATLAB获取文件路径分别设置为这八种识别目标,并将此操作设置为一个大循环。
第3步,设置小循环。调用MATTLAB读取图片函数imread(˙)逐次读取每种识别目标文件夹下的样本:执行I=imread(imgpath),其中I为某个识别目标的一张原始灰度图片,imgpath为文件夹路径。利用函数size(˙)和floor(˙)找到图片剪裁的中心点像素的位置:执行row=floor((size(I,1)-spacing)/2),col=floor((size(I,2)-spacing)/2),其中(row,col)代表中心点像素位置坐标,spacing=100。
第4步,小循环执行每个识别目标文件夹中的原始灰度图片剪裁。调用rect[˙]设置保留像素范围,执行rect=[col,row,spacing-1,spacing-1],执行newI=imcrop(I,rect)剪裁图片,newI为剪裁后的图片结果。
第5步,保存剪裁图片到新路径:将剪裁之后的图片重新命名,执行newname=[frames(jj).name],保存到新路径,执行newpath=fullfile(dstpath,newname),imwrite(newI,newpath),其中jj为小循环中间变量,newname为裁剪图片名字,framework(˙).name为重新命名格式,destpath为要保存剪裁图片的目标路径。
第6步,重复步骤3、4和5,直到完成八种识别目标的全部原始灰度图片完成剪裁和保存。
步骤S200、对所述SAR原始图像样本进行数据增强,生成扩充样本集。
由于现有技术中的SAR目标识别方法中主要问题在于缺乏训练样本,从而导致模型容易发生过拟合的情况。为了解决过拟合的现象,因此需要对训练样本进行扩充,本申请中主要是从增加噪声以及减少噪声两个方面入手,对原始图像样本进行数据增强,以得到加噪样本与去噪样本,从而实现对样本集的扩充。
在其中一个实施例中,本实施例中对SAR图像原始样本中的所有经过裁剪之后的原始灰度图片分别进行三种不同平滑维度的滤波处理,得到三组参数不同的去噪样本。例如设置滤波器平滑维度分别为3*3、5*5和7*7,利用滤波器对经过裁剪之后的4,460张原始灰度图片分别进行滤波处理,即可得到三组参数不同的去噪样本。而对于获得加噪样本,本实施例对SAR图像原始样本中的所有经过裁剪的原始灰度图片分别生成均值为0.5、1.0以及1.5的三组参数不同的斑点噪声图片集,再将所述三组参数不同的斑点噪声图片集分别与所述SAR图像原始样本相乘,即可得到三组加噪样本。经过数据增强处理之后获得的三组去噪样本以及三组加噪样本形成一个扩充样本集。由此可见,本发明通过对SAR图像原始样本进行数据增强,可以将原始的数据扩充6倍,从而大大地增加了训 练样本的数量。
进一步地,在本实施例中,对于获得于去噪样本的具体步骤如下:
第1步,分别读取八种识别目标的剪裁后的原始灰度图片。
第2步,在大循环中设置中值滤波参数T为3,即对图片进行3*3的平滑,执行Ibw=medfilt2(I,T),其中Ibw为中值滤波结果,medfilt2(˙)为中值滤波函数。
第3步,设置小循环,依次对八种识别目标分别进行参数T=3的中值滤波处理,并保存图片到新路径,得到一组去噪样本。
第4步,改变大循环参数T为5和7,执行大循环,得到两组去噪样本。
而对于获得于去噪样本的具体步骤如下:
第1步,分别读取八种识别目标的剪裁后的原始回复图片。
第2步,在大循环中设置乘性噪声参数MU=0.5,执行S=exprnd(MU,spacing,spacing)生成均值为0.5的斑点噪声,S为一个噪声矩阵,exprnd(˙)函数可生成随机指数分布矩阵。将噪声与原始图片相乘,执行newI=I.*uint8(S),其中uint8(˙)为数据格式转变函数,newI为加噪图片结果。
第3步,设置小循环,依次对八种识别目标分别进行参数MU=0.5的加噪处理,并保存图片到新路径,得到一组加噪样本。
第4步,改变大循环参数MU为1.0和1.5,执行大循环,得到另外两组增加斑点噪声样本。
步骤S300、构建基础残差神经网络模型,并在所述基础残差神经网络模型中加入残差控制因子进行优化,构建出优化后的残差神经网络模型。
由于传统的识别方法中不管是基于模板还是基于模型,对先验知识和模型精度要求普遍较高,灵活性和适应性较差,识别准确性受限于模型的可靠性与特征提取的准确度,因此识别精度不高。并且,现有技术中对于SAR目标的识别基本都是采用深度卷积网络模型,该网络模型训练收敛速度较慢,同时由于样本数量限制,还会遇到过拟合的问题,导致网络模型层数受限,因此学习深度受限。为了解决上述问题,本申请中所采用的是残差神经网络模型,并且引入残差控制因子对所述残差神经网络模型进行优化。
在其中一个实施例中,本实施例首先构建一个基础的残差神经网络模型,具体如下:
首先构建一个基本basic单元,包括:第一层为一个3*3卷积核的卷积层,后面接一个BN(batch normalization)层,接着一个激活函数ReLU(rectified linear  unit),再接着同样3*3卷积核的卷积层、一个BN(batch normalization)层和一个激活函数ReLU(rectified linear unit),之后链接在这个堆叠神经单元的首尾,与两个卷积层合并,之后再通过一个ReLU激活函数并输出。所述基本basic单元的结构如图2中的左侧所示。
然后构建一个basic inc单元,包括:第一层为一个3*3卷积核的卷积层,后面接一个BN(batch normalization)层,接着一个激活函数ReLU(rectified linear unit),再接着同样3*3卷积核的卷积层、一个BN(batch normalization)层和一个激活函数ReLU(rectified linear unit),之后链接这个堆叠神经单元的首尾,增加一个1*1卷积核的卷积层,后接一个BN层,之后与两个卷积层合并,之后再通过一个ReLU激活函数并输出。所述Basic inc单元的结构在图2右侧所示。
最后堆叠组合basic单元和basic inc单元,构建本申请中的基本残差神经网路模型:网络模型一共包括20卷积层,依次为“输入→卷积层+basic*3+basic inc+basic*2+basic inc+basic*2+全连接层→输出”。
进一步地,当构建好基本残差网络模型之后,在快捷连接末端增加残差控制因子,对基本残差神经网络模型进行优化,具体如下:
首先构建一个一致性IB(identity block)单元,包括:第一层为一个3*3卷积核的卷积层,后面接一个BN(batch normalization)层,接着一个激活函数ReLU(rectified linear unit),再接着同样3*3卷积核的卷积层、一个BN(batch normalization)层和一个激活函数ReLU(rectified linear unit),这里实施增加残差控制因子Cr。之后链接在这个堆叠神经单元的首尾,与两个卷积层合并,之后再通过一个ReLU激活函数并输出。所述IB单元的结构如图3的左侧所示。
然后构建一个传播TB(transformational block)单元,包括:第一层为一个3*3卷积核的卷积层,后面接一个BN(batch normalization)层,接着一个激活函数ReLU(rectified linear unit),再接着同样3*3卷积核的卷积层、一个BN(batch normalization)层和一个激活函数ReLU(rectified linear unit),这里实施增加残差控制因子Cr。之后链接这个堆叠神经单元的首尾,增加一个1*1卷积核的卷积层,后接一个BN层,之后与两个卷积层合并,之后再通过一个ReLU激活函数并输出。所述TB单元的结构如图3中右侧所示。
最后堆叠组合IB单元和BN单元,构建本申请中的优化后的残差神经网路:网络模型一共包含20卷积层,依次为“输入+卷积层+IB*3+TB+IB*2+TB+IB*2+池化层+密集层+输出”。优化后的残差神经网路模型如图4所示。
较佳地,由于本申请中的残差神经网络模型中采用了快捷连接单元,相比于传统的卷积神经网络,可以实现更深的网络结构。并且申请中是在快捷连接单元的尾端中增加残差控制因子,如图5中所示,可以在训练过程中提高收敛速度,使得模型训练时间大大缩短。优选地,所述残差控制因子的取值范围为-0.5~0.5,增加该控制因子的目的是提高模型训练收敛速度。
更为具体地,此处对本实施例中所构建的优化后的残差神经网络模型作进一步地说明,所示优化后的残差神经网络模型自上而下包括20层结构:
第一层为卷积层,用于对100*100*1*n的输入图片进行卷积,n表示样本输入个数,该层卷积核窗口大小为3*3,步长为1,输出16个特征图,输出图像大小为100*100;
第二层-第三层为一个IB结构单元,卷积层卷积核3*3,步长为2,同样输出16个特征图,输出图像大小为100*100;
第四层-第五层以及第六层-第七层是两个相同的IB单元堆叠而成,每个卷积层卷积核3*3,步长为1,输出16个特征图,输出图像大小为100*100;
第八层-第九层为一个TB结构单元,卷积层卷积核3*3输出32个特征图,步长为2,输出图像大小为50*50;
第十层-第十一层以及第十二层-第十三层是两个相同的IB单元堆叠,每个卷积层卷积核3*3,步长为2,输出32个特征图,输出图像大小为50*50;
第十四层-第十五层为一个TB结构单元,卷积层卷积核3*3输出64个特征图,步长为2,输出图像大小为25*25;
第十六层-第十七层以及第十八层-第十九层是两个相同的IB单元堆叠,每个卷积层卷积核3*3,步长为2,输出64个特征图,输出图像大小为25*25;
第二十层为全连接层,由8个神经元组成,将以上各层连接在一起,构成一个用于SAR目标识别的优化后的残差神经网络模型。
步骤S400、从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型。
在其中一个实施例中,本实施例中需要对优化后的残差神经网络进行训练,以使网络模型满足SAR识别的精度要求。具体的,本实施例中从扩充样本中随机抽取第一数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到当前模型精度,然后将当前模型精度 与初始模型精度进行比较,判断所述当前模型精度是否符合精度要求若符合精度要求,则输出以第一数量的图片进行训练后的残差神经网络模型,该残差神经网络模型即可用来对SAR图像进行识别。若不符合精度要求,则需进一步增加扩充样本集中的样本数量,并重新执行残差神经网络模型的构建、优化以及训练的步骤,以使训练后的残差神经网络机构可以满足要求。
优选地,本实施例中从扩充样本集中随机抽取的图片中包括有各个识别目标对应的原始灰度图片,且每个识别目标的原始灰度图片的数量相同。并且,本实施例中的精度要求为当前模型精度比初始模型精度高于2%。更进一步地,所述初始模型精度是通过随机从SAR图像原始样本抽取第二数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练后得到的。
例如,首先随机抽取SAR图像原始样本中的3200张原始灰度图片(从8种识别目标中分别取400张原始灰度图片),分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到初始模型精度,该初始模型精度主要用于与当前模型精度进行对比。
然后从扩充样本集中随机抽取19200张原始灰度图片(从8种识别目标中分别取400*6张原始灰度图片),同样地分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到当前模型精度。
然后将当前模型精度与初始模型精度对比,如当前模型精度相比于初始模型精度不能有效提高2个百分点,则需增加数据增强操作,进一步扩充样本数量,并重新执行残差神经网络模型的构建、优化以及训练的步骤,即需要重复执行上述的步骤200-400。
步骤S500、将待检测的SAR图像输入至训练后的残差神经网络模型中进行检测识别,输出识别结果。
当经过训练后的残差神经网络模型满足精度要求,则将待检测的SAR图像输入至训练后的残差神经网络模型中进行检测识别,残差神经网络模型就会对SAR图像进行分类与识别,输出识别结果。通过本申请中的数据增强,本申请中的训练后的残差神经网络模型精度从94.58%提高到了99.65%,并且模型运算速度方面,相比没有增加“残差控制因数”的原始残差结构,本申请的模型收敛时间从1200分钟缩短至300分钟,有效提高了运算效率。
应该理解的是,虽然图1的流程图中的各个步骤按照箭头的指示依次显示, 但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明,这些步骤的执行并没有严格的顺序限制,这些步骤可以以其它的顺序执行。而且,图1中的至少一部分步骤可以包括多个子步骤或者多个阶段,这些子步骤或者阶段并不必然是在同一时刻执行完成,而是可以在不同的时刻执行,这些子步骤或者阶段的执行顺序也不必然是依次进行,而是可以与其它步骤或者其它步骤的子步骤或者阶段的至少一部分轮流或者交替地执行。
进一步地,如图6所示,在一个实施例中,提供了本申请的SAR目标识别方法的***框图,从图6中可以看出,待测试样本(即SAR原始图像样本)需要经过数据增强,并将数据增强后的扩充样本输入至经过网络训练的残差神经网络中进行分类与识别,输出识别结果。而在本实施例中训练样本需要经过两种数据增强处理即数据增强处理1(去噪处理)获得去噪样本以及数据增强处理2(加噪处理)获得加噪样本,从而形成扩充样本。
在一个实施例中,如图7所示,提供了一种SAR目标识别装置,包括:原始样本获取模块710、样本数据增强模块720、模型构建且优化模块730、模型训练模块740、图像识别模块750。其中,
原始样本获取模块710,用于获取SAR原始图像样本;
样本数据增强模块720,用于对所述SAR原始图像样本进行数据增强,生成扩充样本集;
模型构建且优化模块730,用于构建基础残差神经网络模型,并在所述基础残差神经网络模型中加入残差控制因子进行优化,构建出优化后的残差神经网络模型;
模型训练模块740,用于从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型;
图像识别模块750,用于将待检测的SAR图像输入至训练后的残差神经网络模型中进行检测识别,输出识别结果。
在一个实施例中,所述原始样本获取模块7107具体包扩从MSTAR数据中选取多个在一定角度的俯仰角下并旋转360°的成像结果;利用格式转换软件将所述成像结果转换成JPG格式的原始灰度图片;将所有的原始灰度图片以中心位置为基准,并在横向和纵向两个方向上取100个像素点进行裁剪,形成SAR图像原始样本;所述SAR图像原始样本中包括有多个识别目标,且每个识别目 标对应有多张经过裁剪的原始灰度图片。
在一个实施例中,样本数据增强模块720具体包括对所述SAR图像原始样本中的所有原始灰度图片分别进行三种不同平滑维度的滤波处理,得到三组参数不同的去噪样本;对所述SAR图像原始样本中的所有原始灰度图片分别生成均值为0.5、1.0以及1.5的三组参数不同的斑点噪声图片集;将所述三组参数不同的斑点噪声图片集分别与所述SAR图像原始样本相乘,得到三组加噪样本;所述去噪样本与所述加噪样本组成扩充样本集。
在一个实施例中,所述残差控制因子的取值范围为-0.5~0.5。
在一个实施例中,所述模型训练模块740具体包括从所述扩充样本中随机抽取第一数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到当前模型精度;将所述当前模型精度与初始模型精度进行比较,判断所述当前模型精度是否符合精度要求;若符合精度要求,则输出以第一数量的图片进行训练后的残差神经网络模型;若不符合精度要求,则需进一步增加扩充样本集中的样本数量,并重新执行残差神经网络模型的构建、优化以及训练的步骤;从所述扩充样本集中随机抽取的图片中包括有各个识别目标对应的原始灰度图片,且每个识别目标的原始灰度图片的数量相同。
在一个实施例中,所述模型训练模块740还包括:随机从SAR图像原始样本抽取第二数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到初始模型精度。
在一个实施例中,所述精度要求为当前模型精度比初始模型精度高于2%。
关于SAR目标识别装置具体限定可以参见上文中对于SAR目标识别方法的限定,在此不再赘述。上述SAR目标识别装置各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中,也可以以软件形式存储于计算机设备中的存储器中,以便于处理器调用执行以上各个模块对应的操作。
在一个实施例中,提供了一种计算机设备,该计算机设备可以是终端,其内部结构图可以如图8所示。该计算机设备包括通过***总线连接的处理器、存储器、网络接口、显示屏和输入装置。其中,该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作***和计算机程序。该内存储器为非易失性存储介质中的操作***和计算机程序的运行提供环境。该计算机设备的网络接口用于与外部 的终端通过网络连接通信。该计算机程序被处理器执行时以实现一种图像光源颜色的检测方法。该计算机设备的显示屏可以是液晶显示屏或者电子墨水显示屏,该计算机设备的输入装置可以是显示屏上覆盖的触摸层,也可以是计算机设备外壳上设置的按键、轨迹球或触控板,还可以是外接的键盘、触控板或鼠标等。
本领域技术人员可以理解,图8中示出的结构,仅仅是与本申请方案相关的部分结构的框图,并不构成对本申请方案所应用于其上的计算机设备的限定,具体的计算机设备可以包括比图中所示更多或更少的部件,或者组合某些部件,或者具有不同的部件布置。
在一个实施例中,提供了一种计算机设备,包括存储器和处理器,存储器中存储有计算机程序,该处理器执行计算机程序时实现以下步骤:
获取SAR原始图像样本;
对所述SAR原始图像样本进行数据增强,生成扩充样本集;
构建基础残差神经网络模型,并在所述基础残差神经网络模型中加入残差控制因子进行优化,构建出优化后的残差神经网络模型;
从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型;
将待检测的SAR图像输入至训练后的残差神经网络模型中进行检测识别,输出识别结果。
在其中一个实施例中,获取SAR原始图像样本具体包括:从MSTAR数据中选取多个在一定角度的俯仰角下并旋转360°的成像结果;利用格式转换软件将所述成像结果转换成JPG格式的原始灰度图片;将所有的原始灰度图片以中心位置为基准,并在横向和纵向两个方向上取100个像素点进行裁剪,形成SAR图像原始样本;所述SAR图像原始样本中包括有多个识别目标,且每个识别目标对应有多张经过裁剪的原始灰度图片。
在其中一个实施例中,对所述SAR原始图像样本进行数据增强,生成扩充样本集具体包括:对所述SAR图像原始样本中的所有原始灰度图片分别进行三种不同平滑维度的滤波处理,得到三组参数不同的去噪样本;对所述SAR图像原始样本中的所有原始灰度图片分别生成均值为0.5、1.0以及1.5的三组参数不同的斑点噪声图片集;将所述三组参数不同的斑点噪声图片集分别与所述SAR图像原始样本相乘,得到三组加噪样本;所述去噪样本与所述加噪样本组成扩充样本集。
在其中一个实施例中,残差控制因子的取值范围为-0.5~0.5。
在其中一个实施例中,上述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型具体包括:从所述扩充样本中随机抽取第一数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到当前模型精度;将所述当前模型精度与初始模型精度进行比较,判断所述当前模型精度是否符合精度要求;若符合精度要求,则输出以第一数量的图片进行训练后的残差神经网络模型;若不符合精度要求,则需进一步增加扩充样本集中的样本数量,并重新执行残差神经网络模型的构建、优化以及训练的步骤;从所述扩充样本集中随机抽取的图片中包括有各个识别目标对应的原始灰度图片,且每个识别目标的原始灰度图片的数量相同。
在其中一个实施例中,上述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型还包括:随机从SAR图像原始样本抽取第二数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到初始模型精度。
在其中一个实施例中,精度要求为当前模型精度比初始模型精度高于2%。
在一个实施例中,本申请还提供了一种计算机可读存储介质,其上存储有计算机程序,计算机程序被处理器执行时实现以下步骤:
获取SAR原始图像样本;
对所述SAR原始图像样本进行数据增强,生成扩充样本集;
构建基础残差神经网络模型,并在所述基础残差神经网络模型中加入残差控制因子进行优化,构建出优化后的残差神经网络模型;
从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型;
将待检测的SAR图像输入至训练后的残差神经网络模型中进行检测识别,输出识别结果。
在其中一个实施例中,获取SAR原始图像样本具体包括:从MSTAR数据中选取多个在一定角度的俯仰角下并旋转360°的成像结果;利用格式转换软件将所述成像结果转换成JPG格式的原始灰度图片;将所有的原始灰度图片以中心位置为基准,并在横向和纵向两个方向上取100个像素点进行裁剪,形成SAR 图像原始样本;所述SAR图像原始样本中包括有多个识别目标,且每个识别目标对应有多张经过裁剪的原始灰度图片。
在其中一个实施例中,对所述SAR原始图像样本进行数据增强,生成扩充样本集具体包括:对所述SAR图像原始样本中的所有原始灰度图片分别进行三种不同平滑维度的滤波处理,得到三组参数不同的去噪样本;对所述SAR图像原始样本中的所有原始灰度图片分别生成均值为0.5、1.0以及1.5的三组参数不同的斑点噪声图片集;将所述三组参数不同的斑点噪声图片集分别与所述SAR图像原始样本相乘,得到三组加噪样本;所述去噪样本与所述加噪样本组成扩充样本集。
在其中一个实施例中,残差控制因子的取值范围为-0.5~0.5。
在其中一个实施例中,上述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型具体包括:从所述扩充样本中随机抽取第一数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到当前模型精度;将所述当前模型精度与初始模型精度进行比较,判断所述当前模型精度是否符合精度要求;若符合精度要求,则输出以第一数量的图片进行训练后的残差神经网络模型;若不符合精度要求,则需进一步增加扩充样本集中的样本数量,并重新执行残差神经网络模型的构建、优化以及训练的步骤;从所述扩充样本集中随机抽取的图片中包括有各个识别目标对应的原始灰度图片,且每个识别目标的原始灰度图片的数量相同。
在其中一个实施例中,上述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型还包括:随机从SAR图像原始样本抽取第二数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到初始模型精度。
在其中一个实施例中,精度要求为当前模型精度比初始模型精度高于2%。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机程序来指令相关的硬件来完成,所述的计算机程序可存储于一非易失性计算机可读取存储介质中,该计算机程序在执行时,可包括如上述各方法的实施例的流程。其中,本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用,均可包括非易失性和/或易失性存储器。非易失 性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限,RAM以多种形式可得,诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
以上实施例的各技术特征可以进行任意的组合,为使描述简洁,未对上述实施例中的各个技术特征所有可能的组合都进行描述,然而,只要这些技术特征的组合不存在矛盾,都应当认为是本说明书记载的范围。
以上所述实施例仅表达了本申请的几种实施方式,其描述较为具体和详细,但并不能因此而理解为对发明专利范围的限制。应当指出的是,对于本领域的普通技术人员来说,在不脱离本申请构思的前提下,还可以做出若干变形和改进,这些都属于本申请的保护范围。因此,本申请专利的保护范围应以所附权利要求为准。

Claims (10)

  1. 一种SAR目标识别方法,其特征在于,所述识别方法是基于深度学习网络所实现的,包括:
    获取SAR原始图像样本;
    对所述SAR原始图像样本进行数据增强,生成扩充样本集;
    构建基础残差神经网络模型,并在所述基础残差神经网络模型中加入残差控制因子进行优化,构建出优化后的残差神经网络模型;
    从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型;
    将待检测的SAR图像输入至训练后的残差神经网络模型中进行检测识别,输出识别结果。
  2. 根据权利要求1所述的SAR目标识别方法,其特征在于,所述获取SAR原始图像样本具体包括:
    从MSTAR数据中选取多个在一定角度的俯仰角下并旋转360°的成像结果;
    利用格式转换软件将所述成像结果转换成JPG格式的原始灰度图片;
    将所有的原始灰度图片以中心位置为基准,并在横向和纵向两个方向上取100个像素点进行裁剪,形成SAR图像原始样本;
    所述SAR图像原始样本中包括有多个识别目标,且每个识别目标对应有多张经过裁剪的原始灰度图片。
  3. 根据权利要求1所述的SAR目标识别方法,其特征在于,所述对所述SAR原始图像样本进行数据增强,生成扩充样本集具体包括:
    对所述SAR图像原始样本中的所有经过裁剪之后原始灰度图片分别进行三种不同平滑维度的滤波处理,得到三组参数不同的去噪样本;
    对所述SAR图像原始样本中的所有经过裁剪之后原始灰度图片分别生成均值为0.5、1.0以及1.5的三组参数不同的斑点噪声图片集;
    将所述三组参数不同的斑点噪声图片集分别与所述SAR图像原始样本相乘,得到三组加噪样本;
    所述去噪样本与所述加噪样本组成扩充样本集。
  4. 根据权利要求1所述的SAR目标识别方法,其特征在于,所述残差控制因子的取值范围为-0.5~0.5。
  5. 根据权利要求1所述的SAR目标识别方法,其特征在于,所述从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型具体包括:
    从所述扩充样本中随机抽取第一数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到当前模型精度;将所述当前模型精度与初始模型精度进行比较,判断所述当前模型精度是否符合精度要求;
    若符合精度要求,则输出以第一数量的图片进行训练后的残差神经网络模型;若不符合精度要求,则需进一步增加扩充样本集中的样本数量,并重新执行残差神经网络模型的构建、优化以及训练的步骤;
    从所述扩充样本集中随机抽取的图片中包括有各个识别目标对应的原始灰度图片,且每个识别目标的原始灰度图片的数量相同。
  6. 根据权利要求5所述的SAR目标识别方法,其特征在于,所述从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型还包括:
    随机从SAR图像原始样本抽取第二数量的图片,并将抽取的图片分别输入至基础残差神经网络模型以及优化后的残差神经网络模型进行训练,得到初始模型精度。
  7. 根据权利要求5所述的SAR目标识别方法,其特征在于,所述精度要求为当前模型精度比初始模型精度高于2%。
  8. 一种SAR目标识别装置,其特征在于,所述装置包括:
    原始样本获取模块,用于获取SAR原始图像样本;
    样本数据增强模块,用于对所述SAR原始图像样本进行数据增强,生成扩充样本集;
    模型构建且优化模块,用于构建基础残差神经网络模型,并在所述基础残差神经网络模型中加入残差控制因子进行优化,构建出优化后的残差神经网络模型;模型训练模块,用于从所述扩充样本集中随机抽取一定数量的图片,并将抽取的图片输入至优化后的残差神经网络模型中进行训练,得到训练后的残差神经网络模型;
    图像识别模块,用于将待检测的SAR图像输入至训练后的残差神经网络模型中进行检测识别,输出识别结果。
  9. 一种计算机设备,包括存储器和处理器,所述存储器存储有计算机程序,其特征在于,所述处理器执行所述计算机程序时实现权利要求1至7中任一项所述方法的步骤。
  10. 一种存储介质,其上存储有计算机程序,其特征在于,所述计算机程序被处理器执行时实现权利要求1至7中任一项所述的方法的步骤。
PCT/CN2019/075175 2018-08-21 2019-02-15 一种sar目标识别方法、装置、计算机设备及存储介质 WO2020037960A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810952342.1A CN109145815B (zh) 2018-08-21 2018-08-21 一种sar目标识别方法、装置、计算机设备及存储介质
CN201810952342.1 2018-08-21

Publications (1)

Publication Number Publication Date
WO2020037960A1 true WO2020037960A1 (zh) 2020-02-27

Family

ID=64790372

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/075175 WO2020037960A1 (zh) 2018-08-21 2019-02-15 一种sar目标识别方法、装置、计算机设备及存储介质

Country Status (2)

Country Link
CN (1) CN109145815B (zh)
WO (1) WO2020037960A1 (zh)

Cited By (38)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111401387A (zh) * 2020-03-10 2020-07-10 腾讯科技(深圳)有限公司 异常样本构建方法、装置、计算机设备和存储介质
CN111476269A (zh) * 2020-03-04 2020-07-31 中国平安人寿保险股份有限公司 均衡样本集构建、翻拍图像识别方法、装置、设备及介质
CN111523564A (zh) * 2020-03-24 2020-08-11 北京航天自动控制研究所 一种用于深度学习训练的sar时敏目标样本增广方法
CN111583271A (zh) * 2020-04-13 2020-08-25 华东师范大学 一种基于癌症ct图像自动预测基因表达类别的方法
CN111582463A (zh) * 2020-06-08 2020-08-25 佛山金华信智能科技有限公司 伺服电机故障识别及模型训练方法、装置、介质及终端
CN111652059A (zh) * 2020-04-27 2020-09-11 西北大学 基于计算鬼成像的目标识别模型构建、识别方法及装置
CN111753887A (zh) * 2020-06-09 2020-10-09 军事科学院***工程研究院后勤科学与技术研究所 一种点源靶标像控点检测模型训练方法及装置
CN111784564A (zh) * 2020-06-30 2020-10-16 稿定(厦门)科技有限公司 自动抠图方法及***
CN111914815A (zh) * 2020-09-05 2020-11-10 广东鲲鹏智能机器设备有限公司 一种垃圾目标的机器视觉智能识别***及方法
CN112016675A (zh) * 2020-08-13 2020-12-01 北京首钢自动化信息技术有限公司 用于钢板表面缺陷检测的cnn模型训练方法和装置
CN112052762A (zh) * 2020-08-27 2020-12-08 西安电子科技大学 基于高斯原型的小样本isar图像目标识别方法
CN112464718A (zh) * 2020-10-23 2021-03-09 西安电子科技大学 一种基于YOLO-Terse网络的目标检测方法及存储介质
CN112580689A (zh) * 2020-11-23 2021-03-30 北京迈格威科技有限公司 神经网络模型的训练方法、应用方法、装置和电子设备
CN112614125A (zh) * 2020-12-30 2021-04-06 湖南科技大学 手机玻璃缺陷检测方法、装置、计算机设备及存储介质
CN112733425A (zh) * 2020-12-10 2021-04-30 东华理工大学 神经网络训练方法、装置、计算机设备及存储介质
CN112906716A (zh) * 2021-02-25 2021-06-04 北京理工大学 基于小波去噪阈值自学习的带噪sar图像目标识别方法
CN113159300A (zh) * 2021-05-15 2021-07-23 南京逸智网络空间技术创新研究院有限公司 图像检测神经网络模型及其训练方法、图像检测方法
CN113344046A (zh) * 2021-05-24 2021-09-03 电子科技大学 一种提高sar图像船只分类精度的方法
CN113361588A (zh) * 2021-06-03 2021-09-07 北京文安智能技术股份有限公司 基于图像数据增强的图像训练集生成方法和模型训练方法
CN113420812A (zh) * 2021-06-23 2021-09-21 西安电子科技大学 基于进化卷积神经网络的极化sar图像分类方法
CN113468906A (zh) * 2021-07-12 2021-10-01 深圳思谋信息科技有限公司 图形码提取模型构建方法、识别方法、装置、设备和介质
CN113469114A (zh) * 2021-07-19 2021-10-01 国网陕西省电力公司电力科学研究院 一种输电线路环水保扰动土范围识别方法、装置和设备
CN113537399A (zh) * 2021-08-11 2021-10-22 西安电子科技大学 多目标进化图卷积神经网络的极化sar图像分类方法及***
CN113553916A (zh) * 2021-06-30 2021-10-26 广西大学 一种基于卷积神经网络的轨道危险区域障碍物检测方法
CN113673554A (zh) * 2021-07-07 2021-11-19 西安电子科技大学 一种基于宽度学习的雷达高分辨距离像目标识别方法
CN113705001A (zh) * 2021-08-27 2021-11-26 西安交通大学 输电线路故障检测方法、***及嵌入式设备
CN113723182A (zh) * 2021-07-21 2021-11-30 西安电子科技大学 一种训练样本受限条件下的sar图像舰船检测方法
CN113902975A (zh) * 2021-10-08 2022-01-07 电子科技大学 一种用于sar舰船检测的场景感知数据增强方法
CN113935417A (zh) * 2021-10-15 2022-01-14 安徽理工大学 一种基于全过程分析的矿下钻杆计数方法
CN114022671A (zh) * 2021-11-11 2022-02-08 中国矿业大学 一种基于残差神经网络的索道缆绳绳位智能识别方法
CN114462646A (zh) * 2022-03-15 2022-05-10 成都中轨轨道设备有限公司 一种基于接触网安全巡检的杆号牌识别方法及***
CN114998956A (zh) * 2022-05-07 2022-09-02 北京科技大学 一种基于类内差异的小样本图像数据扩充方法及装置
CN115482248A (zh) * 2022-09-22 2022-12-16 推想医疗科技股份有限公司 图像分割方法、装置、电子设备及存储介质
CN115967609A (zh) * 2021-10-11 2023-04-14 ***通信集团山东有限公司 内容分发网络故障检测方法和设备
CN112488038B (zh) * 2020-12-15 2023-07-07 中国人民解放军国防科技大学 一种基于图网络学习的目标识别方法
CN116452696A (zh) * 2023-06-16 2023-07-18 山东省计算中心(国家超级计算济南中心) 一种基于双域特征采样的图像压缩感知重构方法及***
CN116959477A (zh) * 2023-09-19 2023-10-27 杭州爱华仪器有限公司 一种基于卷积神经网络的噪声源分类的方法及装置
CN116975611A (zh) * 2023-08-07 2023-10-31 天津大学 一种基于扩散模型ode形式的高频负荷数据生成方法和***

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109145815B (zh) * 2018-08-21 2022-05-03 深圳大学 一种sar目标识别方法、装置、计算机设备及存储介质
CN110009097B (zh) * 2019-04-17 2023-04-07 电子科技大学 胶囊残差神经网络、胶囊残差神经网络的图像分类方法
CN110110625A (zh) * 2019-04-24 2019-08-09 北方工业大学 Sar图像目标识别方法及装置
CN110751061B (zh) * 2019-09-29 2023-04-07 五邑大学 基于sar网络的sar图像识别方法、装置、设备和存储介质
CN111681218A (zh) * 2020-06-02 2020-09-18 浙江同花顺智能科技有限公司 一种骨折识别模型的训练方法、装置、设备及存储介质
CN112381140B (zh) * 2020-11-13 2024-02-06 国家能源集团泰州发电有限公司 一种基于新特征参数的磨粒图像机器学习识别方法
CN112672052A (zh) * 2020-12-24 2021-04-16 华中光电技术研究所(中国船舶重工集团公司第七一七研究所) 一种图像数据增强方法、***、电子设备及存储介质
CN112580554B (zh) * 2020-12-25 2024-01-05 北京环境特性研究所 基于cnn的mstar数据噪声强度控制的分类识别方法
CN112613575B (zh) * 2020-12-30 2024-02-09 清华大学 一种图像分类模型的数据集扩充方法、训练方法及装置
CN113177497B (zh) * 2021-05-10 2024-04-12 百度在线网络技术(北京)有限公司 视觉模型的训练方法、车辆识别方法及装置
CN113537020B (zh) * 2021-07-07 2022-12-09 中国人民解放军国防科技大学 基于改进神经网络的复数sar图像目标识别方法
CN113780364A (zh) * 2021-08-18 2021-12-10 西安电子科技大学 一种模型与数据联合驱动的sar图像目标识别方法
CN115082288B (zh) * 2022-05-16 2023-04-07 西安电子科技大学 基于偏微分方程启发的sar图像到光学图像的转换方法
CN117392545B (zh) * 2023-10-26 2024-02-09 南昌航空大学 一种基于深度学习的sar图像目标检测方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107145846A (zh) * 2017-04-26 2017-09-08 贵州电网有限责任公司输电运行检修分公司 一种基于深度学习的绝缘子识别方法
CN107492095A (zh) * 2017-08-02 2017-12-19 西安电子科技大学 基于深度学习的医学图像肺结节检测方法
CN108038445A (zh) * 2017-12-11 2018-05-15 电子科技大学 一种基于多视角深度学习框架的sar自动目标识别方法
CN109145815A (zh) * 2018-08-21 2019-01-04 深圳大学 一种sar目标识别方法、装置、计算机设备及存储介质

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9086484B2 (en) * 2011-06-30 2015-07-21 The Boeing Company Context-based target recognition
CN104732243B (zh) * 2015-04-09 2018-04-17 西安电子科技大学 基于cnn的sar目标识别方法
CN106156744B (zh) * 2016-07-11 2019-01-29 西安电子科技大学 基于cfar检测与深度学习的sar目标检测方法
CN106683666B (zh) * 2016-12-23 2019-11-08 云知声(上海)智能科技有限公司 一种基于深度神经网络的领域自适应方法
CN107689034B (zh) * 2017-08-16 2020-12-01 清华-伯克利深圳学院筹备办公室 一种去噪方法及装置
CN108052884A (zh) * 2017-12-01 2018-05-18 华南理工大学 一种基于改进残差神经网络的手势识别方法
CN108280460B (zh) * 2017-12-04 2021-07-27 西安电子科技大学 基于改进卷积神经网络的sar车辆目标识别方法
CN107944496A (zh) * 2017-12-06 2018-04-20 电子科技大学 基于改进后的残差网络的骨龄自动化识别***
CN108416318A (zh) * 2018-03-22 2018-08-17 电子科技大学 基于数据增强的合成孔径雷达图像目标深度模型识别方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107145846A (zh) * 2017-04-26 2017-09-08 贵州电网有限责任公司输电运行检修分公司 一种基于深度学习的绝缘子识别方法
CN107492095A (zh) * 2017-08-02 2017-12-19 西安电子科技大学 基于深度学习的医学图像肺结节检测方法
CN108038445A (zh) * 2017-12-11 2018-05-15 电子科技大学 一种基于多视角深度学习框架的sar自动目标识别方法
CN109145815A (zh) * 2018-08-21 2019-01-04 深圳大学 一种sar目标识别方法、装置、计算机设备及存储介质

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
DING, JUN: "Convolutional Neural Network With Data Augmentation for SAR Target Recognition", IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, vol. 13, no. 3, 31 March 2016 (2016-03-31), pages 31, XP011608921, DOI: 10.1109/LGRS.2015.2513754 *
GAO, FEI: "A New Algorithm for SAR Image Target Recognition Based on an Improved Deep Convolutional Neural Network", COGNITIVE COMPUTATION, 26 June 2018 (2018-06-26), XP036946730 *

Cited By (62)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111476269A (zh) * 2020-03-04 2020-07-31 中国平安人寿保险股份有限公司 均衡样本集构建、翻拍图像识别方法、装置、设备及介质
CN111476269B (zh) * 2020-03-04 2024-04-16 中国平安人寿保险股份有限公司 均衡样本集构建、翻拍图像识别方法、装置、设备及介质
CN111401387B (zh) * 2020-03-10 2023-09-19 腾讯科技(深圳)有限公司 异常样本构建方法、装置、计算机设备和存储介质
CN111401387A (zh) * 2020-03-10 2020-07-10 腾讯科技(深圳)有限公司 异常样本构建方法、装置、计算机设备和存储介质
CN111523564A (zh) * 2020-03-24 2020-08-11 北京航天自动控制研究所 一种用于深度学习训练的sar时敏目标样本增广方法
CN111523564B (zh) * 2020-03-24 2023-05-12 北京航天自动控制研究所 一种用于深度学习训练的sar时敏目标样本增广方法
CN111583271A (zh) * 2020-04-13 2020-08-25 华东师范大学 一种基于癌症ct图像自动预测基因表达类别的方法
CN111652059A (zh) * 2020-04-27 2020-09-11 西北大学 基于计算鬼成像的目标识别模型构建、识别方法及装置
CN111652059B (zh) * 2020-04-27 2023-03-24 西北大学 基于计算鬼成像的目标识别模型构建、识别方法及装置
CN111582463B (zh) * 2020-06-08 2024-02-09 佛山金华信智能科技有限公司 伺服电机故障识别及模型训练方法、装置、介质及终端
CN111582463A (zh) * 2020-06-08 2020-08-25 佛山金华信智能科技有限公司 伺服电机故障识别及模型训练方法、装置、介质及终端
CN111753887B (zh) * 2020-06-09 2024-05-28 军事科学院***工程研究院后勤科学与技术研究所 一种点源靶标像控点检测模型训练方法及装置
CN111753887A (zh) * 2020-06-09 2020-10-09 军事科学院***工程研究院后勤科学与技术研究所 一种点源靶标像控点检测模型训练方法及装置
CN111784564B (zh) * 2020-06-30 2022-05-10 稿定(厦门)科技有限公司 自动抠图方法及***
CN111784564A (zh) * 2020-06-30 2020-10-16 稿定(厦门)科技有限公司 自动抠图方法及***
CN112016675A (zh) * 2020-08-13 2020-12-01 北京首钢自动化信息技术有限公司 用于钢板表面缺陷检测的cnn模型训练方法和装置
CN112016675B (zh) * 2020-08-13 2023-12-19 北京首钢自动化信息技术有限公司 用于钢板表面缺陷检测的cnn模型训练方法和装置
CN112052762A (zh) * 2020-08-27 2020-12-08 西安电子科技大学 基于高斯原型的小样本isar图像目标识别方法
CN111914815A (zh) * 2020-09-05 2020-11-10 广东鲲鹏智能机器设备有限公司 一种垃圾目标的机器视觉智能识别***及方法
CN112464718A (zh) * 2020-10-23 2021-03-09 西安电子科技大学 一种基于YOLO-Terse网络的目标检测方法及存储介质
CN112464718B (zh) * 2020-10-23 2024-02-20 西安电子科技大学 一种基于YOLO-Terse网络的目标检测方法及存储介质
CN112580689A (zh) * 2020-11-23 2021-03-30 北京迈格威科技有限公司 神经网络模型的训练方法、应用方法、装置和电子设备
CN112733425A (zh) * 2020-12-10 2021-04-30 东华理工大学 神经网络训练方法、装置、计算机设备及存储介质
CN112733425B (zh) * 2020-12-10 2023-01-03 东华理工大学 神经网络训练方法、装置、计算机设备及存储介质
CN112488038B (zh) * 2020-12-15 2023-07-07 中国人民解放军国防科技大学 一种基于图网络学习的目标识别方法
CN112614125A (zh) * 2020-12-30 2021-04-06 湖南科技大学 手机玻璃缺陷检测方法、装置、计算机设备及存储介质
CN112614125B (zh) * 2020-12-30 2023-12-01 湖南科技大学 手机玻璃缺陷检测方法、装置、计算机设备及存储介质
CN112906716A (zh) * 2021-02-25 2021-06-04 北京理工大学 基于小波去噪阈值自学习的带噪sar图像目标识别方法
CN113159300A (zh) * 2021-05-15 2021-07-23 南京逸智网络空间技术创新研究院有限公司 图像检测神经网络模型及其训练方法、图像检测方法
CN113159300B (zh) * 2021-05-15 2024-02-27 南京逸智网络空间技术创新研究院有限公司 图像检测神经网络模型及其训练方法、图像检测方法
CN113344046A (zh) * 2021-05-24 2021-09-03 电子科技大学 一种提高sar图像船只分类精度的方法
CN113361588A (zh) * 2021-06-03 2021-09-07 北京文安智能技术股份有限公司 基于图像数据增强的图像训练集生成方法和模型训练方法
CN113420812A (zh) * 2021-06-23 2021-09-21 西安电子科技大学 基于进化卷积神经网络的极化sar图像分类方法
CN113420812B (zh) * 2021-06-23 2024-04-19 西安电子科技大学 基于进化卷积神经网络的极化sar图像分类方法
CN113553916A (zh) * 2021-06-30 2021-10-26 广西大学 一种基于卷积神经网络的轨道危险区域障碍物检测方法
CN113673554A (zh) * 2021-07-07 2021-11-19 西安电子科技大学 一种基于宽度学习的雷达高分辨距离像目标识别方法
CN113468906A (zh) * 2021-07-12 2021-10-01 深圳思谋信息科技有限公司 图形码提取模型构建方法、识别方法、装置、设备和介质
CN113468906B (zh) * 2021-07-12 2024-03-26 深圳思谋信息科技有限公司 图形码提取模型构建方法、识别方法、装置、设备和介质
CN113469114A (zh) * 2021-07-19 2021-10-01 国网陕西省电力公司电力科学研究院 一种输电线路环水保扰动土范围识别方法、装置和设备
CN113723182A (zh) * 2021-07-21 2021-11-30 西安电子科技大学 一种训练样本受限条件下的sar图像舰船检测方法
CN113537399B (zh) * 2021-08-11 2024-07-05 西安电子科技大学 多目标进化图卷积神经网络的极化sar图像分类方法及***
CN113537399A (zh) * 2021-08-11 2021-10-22 西安电子科技大学 多目标进化图卷积神经网络的极化sar图像分类方法及***
CN113705001B (zh) * 2021-08-27 2024-05-07 西安交通大学 输电线路故障检测方法、***及嵌入式设备
CN113705001A (zh) * 2021-08-27 2021-11-26 西安交通大学 输电线路故障检测方法、***及嵌入式设备
CN113902975B (zh) * 2021-10-08 2023-05-05 电子科技大学 一种用于sar舰船检测的场景感知数据增强方法
CN113902975A (zh) * 2021-10-08 2022-01-07 电子科技大学 一种用于sar舰船检测的场景感知数据增强方法
CN115967609A (zh) * 2021-10-11 2023-04-14 ***通信集团山东有限公司 内容分发网络故障检测方法和设备
CN115967609B (zh) * 2021-10-11 2024-05-24 ***通信集团山东有限公司 内容分发网络故障检测方法和设备
CN113935417A (zh) * 2021-10-15 2022-01-14 安徽理工大学 一种基于全过程分析的矿下钻杆计数方法
CN114022671A (zh) * 2021-11-11 2022-02-08 中国矿业大学 一种基于残差神经网络的索道缆绳绳位智能识别方法
CN114462646B (zh) * 2022-03-15 2022-11-15 成都中轨轨道设备有限公司 一种基于接触网安全巡检的杆号牌识别方法及***
CN114462646A (zh) * 2022-03-15 2022-05-10 成都中轨轨道设备有限公司 一种基于接触网安全巡检的杆号牌识别方法及***
CN114998956B (zh) * 2022-05-07 2023-03-28 北京科技大学 一种基于类内差异的小样本图像数据扩充方法及装置
CN114998956A (zh) * 2022-05-07 2022-09-02 北京科技大学 一种基于类内差异的小样本图像数据扩充方法及装置
CN115482248B (zh) * 2022-09-22 2023-12-08 推想医疗科技股份有限公司 图像分割方法、装置、电子设备及存储介质
CN115482248A (zh) * 2022-09-22 2022-12-16 推想医疗科技股份有限公司 图像分割方法、装置、电子设备及存储介质
CN116452696B (zh) * 2023-06-16 2023-08-29 山东省计算中心(国家超级计算济南中心) 一种基于双域特征采样的图像压缩感知重构方法及***
CN116452696A (zh) * 2023-06-16 2023-07-18 山东省计算中心(国家超级计算济南中心) 一种基于双域特征采样的图像压缩感知重构方法及***
CN116975611A (zh) * 2023-08-07 2023-10-31 天津大学 一种基于扩散模型ode形式的高频负荷数据生成方法和***
CN116975611B (zh) * 2023-08-07 2024-04-16 天津大学 一种基于扩散模型ode形式的高频负荷数据生成方法和***
CN116959477B (zh) * 2023-09-19 2023-12-19 杭州爱华仪器有限公司 一种基于卷积神经网络的噪声源分类的方法及装置
CN116959477A (zh) * 2023-09-19 2023-10-27 杭州爱华仪器有限公司 一种基于卷积神经网络的噪声源分类的方法及装置

Also Published As

Publication number Publication date
CN109145815A (zh) 2019-01-04
CN109145815B (zh) 2022-05-03

Similar Documents

Publication Publication Date Title
WO2020037960A1 (zh) 一种sar目标识别方法、装置、计算机设备及存储介质
WO2020038205A1 (zh) 目标检测方法、装置、计算机可读存储介质及计算机设备
US20200250497A1 (en) Image classification method, server, and computer-readable storage medium
CN109791625B (zh) 使用人工神经网络进行面部识别
CN110047069B (zh) 一种图像检测装置
CN108765425B (zh) 图像分割方法、装置、计算机设备和存储介质
WO2019218136A1 (zh) 图像分割方法、计算机设备和存储介质
CN110648334A (zh) 一种基于注意力机制的多特征循环卷积显著性目标检测方法
CN113034358B (zh) 一种超分辨率图像处理方法以及相关装置
US11538244B2 (en) Extraction of spatial-temporal feature representation
WO2016054779A1 (en) Spatial pyramid pooling networks for image processing
WO2021135773A1 (zh) 图像重建方法、装置、设备、***及计算机可读存储介质
WO2021022543A1 (en) Computer-implemented method, computer-implemented diagnosis method, apparatus for classifying image, and computer-program product
US10438088B2 (en) Visual-saliency driven scene description
Wu et al. FW-GAN: Underwater image enhancement using generative adversarial network with multi-scale fusion
CN110766127A (zh) 神经网络计算专用电路及其相关计算平台与实现方法
WO2021043023A1 (zh) 图像处理方法及装置、分类器训练方法以及可读存储介质
Cheng et al. Exploring more diverse network architectures for single image super-resolution
Azaza et al. Context proposals for saliency detection
Hua et al. Dynamic scene deblurring with continuous cross-layer attention transmission
Wang et al. Joint depth map super-resolution method via deep hybrid-cross guidance filter
US20230343137A1 (en) Method and apparatus for detecting key point of image, computer device and storage medium
Zhang et al. Thermal image super-resolution via multi-path residual attention network
Yan et al. Joint image-to-image translation with denoising using enhanced generative adversarial networks
Wu et al. Hybrid sparsity learning for image restoration: An iterative and trainable approach

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19852072

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSNOTING OF LOSS OF RIGHTS (EPO FORM 1205A DATED 01.06.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19852072

Country of ref document: EP

Kind code of ref document: A1