WO2021000906A1 - 一种面向sar图像的小样本语义特征增强的方法及装置 - Google Patents

一种面向sar图像的小样本语义特征增强的方法及装置 Download PDF

Info

Publication number
WO2021000906A1
WO2021000906A1 PCT/CN2020/099880 CN2020099880W WO2021000906A1 WO 2021000906 A1 WO2021000906 A1 WO 2021000906A1 CN 2020099880 W CN2020099880 W CN 2020099880W WO 2021000906 A1 WO2021000906 A1 WO 2021000906A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
sar
neural network
semantic
deep neural
Prior art date
Application number
PCT/CN2020/099880
Other languages
English (en)
French (fr)
Inventor
翟懿奎
邓文博
柯琪锐
曹鹤
Original Assignee
五邑大学
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 五邑大学 filed Critical 五邑大学
Publication of WO2021000906A1 publication Critical patent/WO2021000906A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88—Radar or analogous systems specially adapted for specific applications
    • G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
    • G01S13/90—Radar or analogous systems specially adapted for specific applications for mapping or imaging using synthetic aperture techniques, e.g. synthetic aperture radar [SAR] techniques
    • G01S13/9021—SAR image post-processing techniques
    • G01S13/9027—Pattern recognition for feature extraction
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88—Radar or analogous systems specially adapted for specific applications
    • G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
    • G01S13/90—Radar or analogous systems specially adapted for specific applications for mapping or imaging using synthetic aperture techniques, e.g. synthetic aperture radar [SAR] techniques
    • G01S13/9004—SAR image acquisition techniques
    • G01S13/9005—SAR image acquisition techniques with optical processing of the SAR signals
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
    • G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
    • G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88—Radar or analogous systems specially adapted for specific applications
    • G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
    • G01S13/90—Radar or analogous systems specially adapted for specific applications for mapping or imaging using synthetic aperture techniques, e.g. synthetic aperture radar [SAR] techniques
    • G01S13/9021—SAR image post-processing techniques
    • G01S13/9029—SAR image post-processing techniques specially adapted for moving target detection within a single SAR image or within multiple SAR images taken at the same time

Definitions

  • the invention relates to the technical field of image recognition, in particular to a method and a device for enhancing the semantic features of small samples for SAR images.
  • Synthetic Aperture Radar is a microwave imaging device with extremely high resolution. It uses pulse compression technology and the principle of synthetic aperture to achieve imaging of ground scenes. It is used in environmental monitoring, resource exploration, surveying and mapping, and battlefield reconnaissance. Both play an important role. With the advancement of the Guangdong-Hong Kong-Macao Greater Bay Area, the demand for SAR images for safety monitoring in the Bay Area continues to increase, especially in the areas of remote sensing aviation aircraft monitoring, ship target dynamic monitoring, and oil spill early warning dynamic monitoring.
  • the purpose of the present invention is to provide a method and device for enhancing the semantic features of small samples of SAR images, which can solve the problem of insufficient sample size in the Bay Area, improve the network convergence speed, and enhance the semantic feature space of SAR images
  • the expressive ability of SAR images can improve the accuracy of target recognition in SAR images.
  • an embodiment of the present invention proposes a method for SAR image-oriented small sample semantic feature enhancement, which includes the following steps:
  • the sample set includes the SAR target image and the SAR target virtual image;
  • the autoencoder is used to map the feature space and the semantic space on the feature map to obtain a depth visual feature enhanced by semantic features.
  • obtaining a sample set of the SAR target image, and performing migration learning training on the sample set to obtain a deep neural network for initializing the SAR target image includes the following steps:
  • the parameters and weights of the sample set are used to initialize the deep neural network of the SAR target image.
  • deep neural networks include deep residual ResNet networks or DenseNet networks.
  • using the activation function to perform network optimization on the deep neural network and using the optimized deep neural network to perform feature extraction on the SAR target image to obtain a feature map includes the following steps:
  • An additional fully connected layer is connected behind the ReLU activation function layer, and the local optimal weight is input into the additional fully connected layer for classification processing to obtain a feature map.
  • the Softmax-MSE loss function is used to perform gradient loss calculation on the feature map, and the optimized deep neural network is reversely adjusted to minimize the gradient loss to obtain a fitted feature map.
  • the self-encoder includes an encoder and a decoder.
  • using an autoencoder to map the feature space and semantic space to obtain a depth visual feature enhanced by semantic features includes the following steps:
  • a decoder is used to map the semantic feature value from the semantic space to the feature space to obtain feature-enhanced depth visual features.
  • an embodiment of the present invention also proposes a small sample semantic feature enhancement device for SAR images, which includes at least one control processor and a memory for communicating with the at least one control processor; the memory stores There are instructions that can be executed by the at least one control processor, and the instructions are executed by the at least one control processor, so that the at least one control processor can execute the SAR-oriented A method for enhancing the semantic features of small samples of images.
  • an embodiment of the present invention also proposes a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to make a computer execute the above Any one of the methods for enhancing the semantic features of small samples of SAR images.
  • the technical solution provided in the embodiment of the present invention has at least the following beneficial effects: by performing transfer learning training on the sample set, the deep neural network of the SAR target image is initialized, which improves the accuracy of the training due to insufficient sample size of the SAR target image.
  • the shortcomings of eigenvalues, and effectively improve the convergence speed of the deep neural network use the activation function to optimize the deep neural network to make the eigenvalues more accurate; use the autoencoder to map the feature space and semantic space of the SAR target graph , Can enhance the expression ability of the semantic space and feature space of SAR images, thereby improving the accuracy of SAR image target recognition.
  • FIG. 1 is an overall flowchart of an embodiment of the method for enhancing the semantic features of small samples for SAR images according to the present invention
  • FIG. 2 is a flow chart of obtaining a sample set of a SAR target image and performing migration learning training on the sample set in an embodiment of the method for enhancing semantic features of small samples of SAR images according to the present invention
  • FIG. 3 is a flowchart of an embodiment of the method for enhancing the semantic features of small samples for SAR images of the present invention using an activation function to perform network optimization on the deep neural network;
  • FIG. 4 is a flowchart of an embodiment of the method for enhancing the semantic features of small samples for SAR images of the present invention using an autoencoder to map the feature space and semantic space;
  • FIG. 5 is a schematic diagram of an embodiment of the method for enhancing semantic features of small samples for SAR images of the present invention, using an autoencoder to map the feature space and semantic space of the feature map.
  • the present invention provides a method and device for SAR image-oriented semantic feature enhancement of small samples, which can solve the problem of insufficient sample size in the Bay Area, improve the convergence speed of the network, and enhance the expression of the semantic feature space of SAR images Ability to improve the accuracy of SAR image target recognition.
  • an embodiment of the present invention provides a method for enhancing the semantic features of small samples of SAR images, including the following steps:
  • Step S100 Obtain a sample set of a SAR target image, perform transfer learning training on the sample set to obtain a deep neural network for initializing the SAR target image, the sample set includes the SAR target image and the SAR target virtual image;
  • Step S200 Use the activation function to perform network optimization on the deep neural network, and use the optimized deep neural network to perform feature extraction on the SAR target image to obtain a feature map;
  • Step S300 Use an auto-encoder to map the feature space and semantic space to obtain a depth visual feature enhanced by semantic features.
  • step S100 collects sample sets of virtual images in the same category as the SAR target image in different angles, different shapes, etc., and trains the sample sets of the virtual images to obtain each layer after the sample set is trained. Corresponding eigenvalues; at the same time, the eigenvalues are transferred and learned to initialize the deep neural network of the SAR target image, so that each layer of the deep neural network has a fixed initial value instead of a randomly set initial value.
  • the migration learning training of the sample set of the virtual image the problem of insufficient sample size of the SAR target image is solved, the convergence speed of the deep neural network is effectively improved, and the accuracy of SAR image recognition is improved.
  • the sample set of virtual images can be the front and side of SAR target images such as airplanes and ships collected on the Internet, as well as images of various shapes, etc.;
  • the network layer of the initialization deep neural network can be set as the last layer. Any network layer before the connection layer, that is, the network layer before the fully connected layer of the last layer, can be initialized with the feature value of the sample set of the virtual image.
  • Step S200 uses the activation function to optimize the deep neural network, so that the network has the learning ability of hierarchical nonlinear mapping, and makes the model training more efficient; the activation function can use sigmoid function, tanh function, relu function, maxout function, etc.
  • the optimized deep neural network is used to extract the features of the SAR target image, and the extracted features are more accurate and the fitting effect is better.
  • Step S300 Use the autoencoder to map the feature space and semantic space of the feature map of the SAR target image, which can enhance the expression ability of the semantic space and feature space of the SAR image, effectively explore the difference between the multi-layer features, and realize the information Complementarity and enhancement to improve the accuracy of SAR image target recognition.
  • another embodiment of the present invention also provides a method for SAR image-oriented small-sample semantic feature enhancement, in which a sample set of SAR target images is obtained, and migration learning training is performed on the sample set to obtain
  • Initializing the deep neural network of the SAR target image includes the following steps:
  • Step S110 Obtain a SAR target image, and establish a deep neural network of the SAR target image;
  • Step S120 Collect a sample set of virtual images of the same category as the SAR target image in different states, and extract the parameters and weights of the sample set using the feature extraction layer of the convolutional network;
  • Step S130 Use the parameters and weights of the sample set to initialize the deep neural network of the SAR target image.
  • step S110 acquires a SAR target image and uses the SAR target image to build a deep neural network
  • step S120 acquires a sample set of virtual images in the same category as the SAR target image with different angles and shapes, and uses convolution
  • the feature extraction layer of the network extracts the optimal parameters and weights of each layer of the sample set
  • step S130 uses the optimal parameters and weights of each layer of the sample set to initialize the deep neural network of the SAR target image, and the initialized network layer is set as the last of the deep neural network Any network layer before a fully connected layer, that is, the number of migrated layers does not exceed the number of layers where the last fully connected layer is located, so that the deep neural network obtains a fixed initial value for each layer, rather than a randomly set initial value, increasing The learning ability of the deep neural network is solved, and the problem of insufficient sample size of the SAR target image is solved, thereby effectively improving the convergence speed of the deep neural network and improving the accuracy of SAR target image recognition.
  • the SAR target image can be the target image of the aircraft, ship, etc. obtained by radar detection
  • the sample set of the virtual image can be the front, side, and various shapes of the image of the aircraft, ship, tank and other targets collected on the Internet. .
  • another embodiment of the present invention also provides a method for SAR image-oriented small sample semantic feature enhancement, wherein the deep neural network includes a deep residual ResNet network or a DenseNet network.
  • the deep neural network can be set as a deep residual ResNet network
  • the deep residual ResNet network preferably uses the residual network structure of ResNet-50, that is, the deep residual network has a network structure of 50 layers, which greatly increases The depth of the network is improved, and the accuracy of feature extraction is improved.
  • the residual block structure is applied in the deep residual ResNet network, which makes the network layer jump between every two or three layers, which solves the problem that the network is too deep and difficult to train The problem of simplifies the training steps of the deep network and effectively improves the recognition ability and training speed of SAR target images.
  • the deep neural network can also be set as a DenseNet network.
  • the DenseNet network is a convolutional neural network with dense connections. There are direct connections between any two layers of networks, and the feature maps of each layer are merged in the dimension of the channel. It effectively reduces the number of feature maps and alleviates the problem of gradient disappearance; inputting the SAR target image into the DenseNet network for feature extraction can strengthen the propagation of features, the deep network training effect is good, and the recognition rate of the SAR target image is improved.
  • another embodiment of the present invention also provides a method for SAR image-oriented small sample semantic feature enhancement, wherein the activation function is used to perform network optimization on the deep neural network, and the optimized The deep neural network performs feature extraction on the SAR target image to obtain a feature map, including the following steps:
  • Step S210 Connect the Maxout activation function layer behind the fully connected layer of the last layer of the deep neural network, and input the SAR target image into the deep neural network containing the Maxout activation function layer for feature extraction to obtain the local maximum weight of the feature ;
  • Step S220 Connect the ReLU activation function layer behind the Maxout activation function layer, and input the local maximum weight into the ReLU activation function layer for processing to obtain the local optimal weight;
  • Step S230 Connect an additional fully connected layer behind the ReLU activation function layer, and input the local optimal weights into the additional fully connected layer for classification processing to obtain a feature map.
  • step S210 the SAR target image is input to the deep neural network containing the Maxout activation function layer for feature extraction.
  • Each input feature of the Maxout activation function layer is composed of m selectable hidden layer nodes.
  • the following formula extracts the local maximum weight of the feature of the SAR target image:
  • h i is the local maximum weight of the i th feature weight
  • x T is the right input feature weight set
  • W ij is d * m * k three-dimensional weight matrix
  • b is a two-dimensional offset vector m * k's
  • d Represents the number of input weights
  • m represents the number of hidden layer nodes, where each hidden layer node is composed of k "hidden layer” nodes, and the structure of the "hidden layer” node is similar to the hidden layer node
  • k “hidden layer” nodes are all linear output.
  • the Maxout activation function layer is a continuous learning activation function layer, the change state is determined by W, and the Maxout activation function is a piecewise linear function, which can approximate any convex function, and has strong fitting ability and nonlinear characterization ability. Make the network obtain better local solutions.
  • Step S220 connects the ReLU activation function layer behind the Maxout activation function layer.
  • the ReLU activation function layer has unilateral inhibition. The negative value in the local maximum weight output by the Maxout activation function layer is changed to 0, while the positive value remains unchanged.
  • the local maximum weight is processed with one-sided inhibition to obtain the local optimal solution.
  • the one-sided inhibition of the ReLU activation function layer makes the features of the SAR target image have sparse activation, alleviating the network depth.
  • the gradient divergence problem speeds up the convergence speed of the network, enhances the generalization ability of the characteristics of the SAR target image, and makes the characteristics of the SAR target image more representative.
  • the Maxout activation function is a piecewise linear function, which can process the data and reduce the output of the data in the dimension;
  • the ReLU activation function layer has sparse connectivity, which makes the deep neural network training process easier;
  • the Maxout activation function layer and ReLU The continuous connection setting of the activation function layer constitutes the dual activation function layer.
  • the dual activation function layer has the properties of the Maxout activation function and the ReLU activation function, and has a certain data reduction for the data of the SAR target image characteristics processed by the deep neural network. It also makes the network training process easier and improves the feature recognition ability of SAR target images.
  • an additional fully connected layer is connected behind the ReLU activation function layer.
  • the additional fully connected layer performs fusion classification on the features of the SAR target image and its local optimal weights, and calculates the probability distribution of each feature to obtain the feature Figure.
  • another embodiment of the present invention also provides a method for SAR image-oriented small sample semantic feature enhancement, wherein the activation function is used to perform network optimization on the deep neural network, and the optimized The deep neural network performs feature extraction on the SAR target image to obtain a feature map, which also includes the following steps:
  • Step S240 Perform gradient loss calculation on the feature map using the Softmax-MSE loss function, and reversely adjust the optimized deep neural network to minimize the gradient loss to obtain a fitted feature map.
  • the Softmax-MSE loss function layer is connected after the additional fully connected layer, and the gradient loss calculation is performed on the feature map.
  • the loss value L of the Softmax-MSE loss function layer is:
  • another embodiment of the present invention also provides a method for SAR image-oriented small sample semantic feature enhancement, wherein the self-encoder includes an encoder and a decoder.
  • the self-encoder includes an encoder and a decoder.
  • the encoder can convert the original input set into an intermediate output set.
  • the dimension of the intermediate output set is lower than that of the original input set, that is, the encoder can reduce the original input.
  • the dimensionality of the set merge and combine the data of the original input set, realize the complementation of information, transform the relationship between the data, and improve the data expression ability;
  • the decoder can receive the intermediate output set of the encoder, and The output set reconstructs the data and transforms the data into the spatial structure of the original input set's dimensions.
  • the original input set can be images, vectors, video or audio data, etc.
  • another embodiment of the present invention also provides a method for SAR image-oriented small-sample semantic feature enhancement, wherein an autoencoder is used to perform feature space and semantics on the feature map.
  • Spatial mapping to obtain depth visual features enhanced by semantic features includes the following steps:
  • Step S310 Input the feature map into a neural network, and extract the feature value of each layer of the neural network by using an encoder;
  • Step S320 Use Gaussian white noise to perform simulation processing on the eigenvalues to obtain simulated eigenvalues
  • Step S330 Use the encoder to map the simulated feature values from feature space to semantic space to obtain semantically enhanced semantic feature values
  • Step S340 Use a decoder to map the semantic feature value from the semantic space to the feature space to obtain a feature-enhanced depth visual feature.
  • step S310 inputs the feature map into a neural network, and uses the encoder to extract the feature value of each layer of the neural network;
  • the Gaussian white noise of step S320 is the noise whose probability density obeys the normal distribution.
  • noise is often not caused by a single source, but a complex of noise from different sources, that is, real noise is composed of random variables with different probability distributions, and each random variable is independent.
  • the normalization of real noise will always approach a Gaussian distribution as the number of noise sources increases. Therefore, using Gaussian white noise to simulate the eigenvalues can better simulate the unknown Real noise, get simulated characteristic value.
  • step S330 the encoder is used to process the simulated feature value, so that the feature value of each layer has different encoding information, where each encoded information represents different levels of abstract semantic information, and the encoder effectively utilizes the encoded information of the simulated feature value
  • the difference between the coded information enables the merging and combination of information to achieve complementary information, so that the coded information can match the output feature space of each layer of the neural network, so that the simulated feature value can be mapped from the feature space to the semantic space, and the semantics can be obtained.
  • Feature-enhanced semantic feature value; and the semantic feature value contains the semantic feature label of the category, for example, the category includes aircraft, ship, tank, etc.
  • the decoder in step S340 and the encoder have exactly the same architecture.
  • the decoder is used to process semantic feature values, and the enhanced semantic space is remapped and projected to the feature space of the neural network through deformation, size transformation and splicing methods , To obtain deep visual features with enhanced expressive power in semantic feature space, enrich the expressive power of features, and improve the performance of SAR target recognition.
  • another embodiment of the present invention also provides a method for enhancing the semantic features of small samples of SAR images.
  • the identification method includes the following steps:
  • Step S110 Obtain a SAR target image, and establish a deep neural network of the SAR target image;
  • Step S120 Collect a sample set of virtual images of the same category as the SAR target image in different states, and extract the parameters and weights of the sample set using the feature extraction layer of the convolutional network;
  • Step S130 Initialize the deep neural network of the SAR target image by using the parameters and weights of the sample set;
  • Step S210 Connect the Maxout activation function layer behind the fully connected layer of the last layer of the deep neural network, and input the SAR target image into the deep neural network containing the Maxout activation function layer for feature extraction to obtain the local maximum weight of the feature ;
  • Step S220 Connect the ReLU activation function layer behind the Maxout activation function layer, and input the local maximum weight into the ReLU activation function layer for processing to obtain the local optimal weight;
  • Step S230 Connect an additional fully connected layer behind the ReLU activation function layer, and input the local optimal weights into the additional fully connected layer for classification processing to obtain a feature map;
  • Step S240 Perform gradient loss calculation on the feature map by using the Softmax-MSE loss function, and reversely adjust the optimized deep neural network to minimize the gradient loss to obtain a fitted feature map;
  • Step S310 Input the feature map into a neural network, and extract the feature value of each layer of the neural network by using an encoder;
  • Step S320 Use Gaussian white noise to perform simulation processing on the eigenvalues to obtain simulated eigenvalues
  • Step S330 Use the encoder to map the simulated feature values from feature space to semantic space to obtain semantically enhanced semantic feature values
  • Step S340 Use a decoder to map the semantic feature value from the semantic space to the feature space to obtain a feature-enhanced depth visual feature.
  • step S110 obtains a SAR target image, and uses the SAR target image to establish a deep residual ResNet network.
  • the deep residual ResNet network preferably uses the structure of the ResNet-50 residual network, which greatly increases the depth of the network and improves the characteristics Accuracy of extraction;
  • the residual block structure is applied in the deep residual ResNet network, which makes the network layer jump between every two or three layers, which solves the problem of the network being too deep and difficult to train, and simplifies the deep network
  • the training steps effectively improve the recognition ability and training speed of SAR target images.
  • the filters of the deep residual ResNet network use specifications of 5*5, 3*3, and 1*1, the learning rate is 0.001, the training algebra is 100, and the training and test batches are 64 and 32, respectively.
  • Step S120 collects a sample set of virtual images in the same category as the SAR target image with different angles, shapes, etc., and uses the feature extraction layer of the convolutional network to extract the optimal parameters and weights of each layer of the sample set;
  • Step S130 uses the sample set
  • the optimal parameters and weights of each layer initialize the deep residual ResNet network of the SAR target image, and the initialized network layer is set to any network layer before the last fully connected layer of the deep residual ResNet network, that is, the number of migrated layers does not exceed
  • the number of layers in the last fully connected layer allows the deep residual ResNet network to obtain a fixed initial value for each layer, rather than a randomly set initial value, which improves the learning ability of the deep residual ResNet network and solves the SAR target image sample The problem of insufficient quantity, thus effectively improving the convergence speed of the deep residual ResNet network, and improving the accuracy of SAR target image recognition.
  • the SAR target image can be the target image of the aircraft, ship, etc. obtained by radar detection
  • the sample set of the virtual image can be the front, side, and various shapes of the image of the aircraft, ship, tank and other targets collected on the Internet. .
  • Step S210 Input the SAR target image into the deep residual ResNet network containing the Maxout activation function layer for feature extraction.
  • Each input feature of the Maxout activation function layer is composed of m selectable hidden layer nodes.
  • the SAR is extracted by the following formula The local maximum weight of the feature of the target image:
  • h i is the local maximum weight of the i th feature weight
  • x T is the right input feature weight set
  • W ij is d * m * k three-dimensional weight matrix
  • b is a two-dimensional offset vector m * k's
  • d Represents the number of input weights
  • m represents the number of hidden layer nodes, where each hidden layer node is composed of k "hidden layer” nodes, and the structure of the "hidden layer” node is similar to the hidden layer node
  • k “hidden layer” nodes are all linear output.
  • the Maxout activation function layer is a continuous learning activation function layer, the change state is determined by W, and the Maxout activation function is a piecewise linear function, which can approximate any convex function, and has strong fitting ability and nonlinear characterization ability. Make the network obtain better local solutions.
  • Step S220 connects the ReLU activation function layer behind the Maxout activation function layer.
  • the ReLU activation function layer has unilateral inhibition. The negative value in the local maximum weight output by the Maxout activation function layer is changed to 0, while the positive value remains unchanged.
  • the local maximum weight is processed with one-sided inhibition to obtain the local optimal solution.
  • the one-sided inhibition of the ReLU activation function layer makes the features of the SAR target image have sparse activation, alleviating the network depth.
  • the gradient divergence problem speeds up the convergence speed of the network, enhances the generalization ability of the characteristics of the SAR target image, and makes the characteristics of the SAR target image more representative.
  • the Maxout activation function is a piecewise linear function, which can process the data and reduce the output of the data in dimensionality;
  • the ReLU activation function layer has sparse connectivity, which makes the training process of the deep residual ResNet network easier;
  • the Maxout activation function layer The setting of continuous connection with the ReLU activation function layer constitutes a dual activation function layer.
  • the dual activation function layer has the properties of the Maxout activation function and the ReLU activation function. It has the characteristics of the SAR target image data processed by the deep residual ResNet network Certain data dimensionality reduction capabilities, and make the network training process easier, improve the SAR target image feature recognition capabilities.
  • an additional fully connected layer is connected behind the ReLU activation function layer.
  • the additional fully connected layer performs fusion classification on the features of the SAR target image and its local optimal weights, and calculates the probability distribution of each feature to obtain the feature Figure.
  • Step S240 After the additional fully connected layer, connect the Softmax-MSE loss function layer, and perform gradient loss calculation on the feature map.
  • the number of input parameters of the Softmax-MSE loss function layer is m
  • the loss value L of the Softmax-MSE loss function layer is:
  • step S310 inputs the feature map into a neural network, and uses the encoder to extract the feature value of each layer of the neural network;
  • the Gaussian white noise of step S320 is the noise whose probability density obeys the normal distribution.
  • Noise is often not caused by a single source, but a complex of noise from different sources, that is, real noise is composed of random variables with different probability distributions, and each random variable is independent.
  • the normalization of real noise will always approach a Gaussian distribution as the number of noise sources increases. Therefore, using Gaussian white noise to simulate the eigenvalues can better simulate the unknown Real noise, get simulated characteristic value.
  • step S330 the encoder is used to process the simulated feature value, so that the feature value of each layer has different encoding information, where each encoded information represents different levels of abstract semantic information, and the encoder effectively utilizes the encoded information of the simulated feature value
  • the difference between the coded information enables the merging and combination of information to achieve complementary information, so that the coded information can match the output feature space of each layer of the neural network, so that the simulated feature value can be mapped from the feature space to the semantic space, and the semantics can be obtained.
  • Feature-enhanced semantic feature value; and the semantic feature value contains the semantic feature label of the category, for example, the category includes aircraft, ship, tank, etc.
  • the decoder in step S340 and the encoder have exactly the same architecture.
  • the decoder is used to process semantic feature values, and the enhanced semantic space is remapped and projected to the feature space of the neural network through deformation, size transformation and splicing methods , To obtain deep visual features with enhanced expressive power in semantic feature space, enrich the expressive power of features, and improve the performance of SAR target recognition.
  • another embodiment of the present invention also provides a small-sample semantic feature enhancement device for SAR images, which includes at least one control processor and a memory for communicating with the at least one control processor; the memory Stored are instructions that can be executed by the at least one control processor, and the instructions are executed by the at least one control processor, so that the at least one control processor can execute the one of the orientations described in any one of the above A method to enhance the semantic features of small samples of SAR images.
  • the feature enhancement device includes: one or more control processors and memories, and the control processors and memories may be connected by a bus or in other ways.
  • the memory can be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as program instructions/modules corresponding to the feature enhancement method in the embodiment of the present invention.
  • the control processor executes various functional applications and data processing of the feature enhancement device by running non-transitory software programs, instructions, and modules stored in the memory, that is, implementing the feature enhancement method of the above method embodiment.
  • the memory may include a storage program area and a storage data area.
  • the storage program area may store an operating system and an application program required by at least one function; the storage data area may store data created according to the use of the feature enhancement device and the like.
  • the memory may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid state storage devices.
  • the storage may optionally include storage remotely provided with respect to the control processor, and these remote storages may be connected to the feature enhancement device via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.
  • the one or more modules are stored in the memory, and when executed by the one or more control processors, the feature enhancement method in the above method embodiment is executed, for example, steps S100 to S100 to the feature enhancement method described above are executed.
  • the embodiment of the present invention further provides a computer-readable storage medium, the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are executed by one or more control processors, for example, a control processor Execution may cause the above-mentioned one or more control processors to execute the feature enhancement method in the above-mentioned method embodiment, for example, execute the above-described method steps S100 to S300, S110 to S130, S210 to S240, and S310 to S340.
  • the device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.
  • each implementation manner can be implemented by means of software plus a general hardware platform.
  • All or part of the processes in the methods of the above-mentioned embodiments can be implemented by computer programs instructing relevant hardware.
  • the programs can be stored in a computer-readable storage medium, and the program is executed At this time, it may include the process of the embodiment of the above method.
  • the storage medium may be a magnetic disk, an optical disc, a read-only memory (Read Only Memory, ROM), or a random access memory (Random Access Memory, RAM), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Remote Sensing (AREA)
  • Physics & Mathematics (AREA)
  • Radar, Positioning & Navigation (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Electromagnetism (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Signal Processing (AREA)
  • Astronomy & Astrophysics (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

一种面向SAR图像的小样本语义特征增强的方法，包括如下步骤：获取SAR目标图像的样本集，对样本集进行迁移学习训练，得到初始化SAR目标图像的深度神经网络（S100），改善了由于SAR目标图像样本量不足，而无法训练得到精确特征值的缺点，并且有效地提升深度神经网络的收敛速度；利用激活函数对所述深度神经网络进行网络优化，采用优化后的深度神经网络对SAR目标图像进行特征提取，得到特征图（S200），使特征值更加准确；利用自编码器对所述特征图进行特征空间与语义空间的映射，能够增强SAR图像的语义空间和特征空间的表达能力，得到语义特征增强的深度视觉特征（S300），从而提高SAR图像目标识别的准确率。

Description

一种面向SAR图像的小样本语义特征增强的方法及装置 技术领域
本发明涉及图像识别技术领域，具体涉及一种面向SAR图像的小样本语义特征增强的方法及装置。
背景技术
合成孔径雷达(Synthetic Aperture Radar,SAR)是一种分辨率极高的微波成像设备，采用脉冲压缩技术和合成孔径原理实现对地面场景的成像，在环境监测、资源勘查、测绘以及战场侦查等领域都发挥着重要作用。随着粤港澳大湾区建设的推进，湾区的安全监测对SAR图像的应用需求不断增强，特别是在遥感航空飞机监测、船舰目标动态监视、溢油预警动态监测等方面。但是由于湾区的标注样本量不足，在传统的特征提取模型训练过程中容易出现过拟合、陷入局部最优解、收敛速度慢等问题，造成模型泛化能力退化，并且SAR目标检测对复杂场景多目标的检测还存在一定的难度，影响了湾区目标特征的准确提取，因此亟需建立起一个可以解决湾区样本量少，同时使SAR图像特征增强的机制。
发明内容
为解决上述问题，本发明的目的在于提供一种面向SAR图像的小样本语义特征增强的方法及装置，能够解决湾区的样本量不足的问题，提升网络收敛速度，增强SAR图像的语义特征空间的表达能力，从而提高SAR图像目标识别的准确率。
本发明解决其问题所采用的技术方案是：第一方面，本发明实施例提出了一种面向SAR图像的小样本语义特征增强的方法，包括如下步骤：
获取SAR目标图像的样本集，对样本集进行迁移学习训练，得到初始化SAR目标图像的深度神经网络，所述样本集包括SAR目标图像与SAR目标虚拟图像；
利用激活函数对所述深度神经网络进行网络优化，采用优化后的深度神经网络对SAR目标图像进行特征提取，得到特征图；
利用自编码器对所述特征图进行特征空间与语义空间的映射，得到语义特征增强的深度视觉特征。
进一步，获取SAR目标图像的样本集，对样本集进行迁移学习训练，得到初始化SAR目标图像的深度神经网络，包括如下步骤：
获取SAR目标图像，建立SAR目标图像的深度神经网络；
采集与SAR目标图像相同类别的不同状态的虚拟图像的样本集，利用卷积网络的特征提取层提取样本集的参数和权重；
利用所述样本集的参数和权重初始化SAR目标图像的深度神经网络。
进一步，深度神经网络包括深度残差ResNet网络或DenseNet网络。
进一步，利用激活函数对所述深度神经网络进行网络优化，采用优化后的深度神经网络对SAR目标图像进行特征提取，得到特征图，包括如下步骤：
在深度神经网络的最后一层的全连接层后面连接Maxout激活函数层，将SAR目标图像输入到所述含有Maxout激活函数层的深度神经网络中进行特征提取，得到特征的局部最大权重；
在Maxout激活函数层后面连接ReLU激活函数层，将所述局部最大权重输入到ReLU激活函数层进行处理，得到局部最优权重；
在ReLU激活函数层后面连接一层额外的全连接层，将所述局部最优权重输入到额外的全连接层中进行分类处理，得到特征图。
进一步，利用激活函数对所述深度神经网络进行网络优化，采用优化后的深度神经网络对SAR目标图像进行特征提取，得到特征图，还包括如下步骤：
利用Softmax-MSE损失函数对所述特征图进行梯度损失计算，并通过反向调整所述优化后的深度神经网络使梯度损失最小，得到拟合的特征图。
进一步，自编码器包括编码器和译码器。
进一步，利用自编码器对所述特征图进行特征空间与语义空间的映射，得到语义特征增强的深度视觉特征，包括如下步骤：
将所述特征图输入到一个神经网络中，利用编码器提取所述神经网络每层的特征值；
利用高斯白噪声对所述特征值进行仿真处理，得到仿真特征值；
利用所述编码器将所述仿真特征值进行特征空间向语义空间的映射，得到语义增强的语义特征值；
采用译码器对所述语义特征值进行语义空间向特征空间的映射，得到特征增强的深度视觉特征。
第二方面，本发明实施例还提出了一种面向SAR图像的小样本语义特征增强装置，包括至少一个控制处理器和用于与所述至少一个控制处理器通信连接的存储器；所述存储器存储有可被所述至少一个控制处理器执行的指令，所述指令被所述至少一个控制处理器执行，以使所述至少一个控制处理器能够执行如以上任一项所述的一种面向SAR图像的小样本语义特征增强的方法。
第三方面，本发明实施例还提出了一种计算机可读存储介质，其特征在于，所述计算机可读存储介质存储有计算机可执行指令，所述计算机可执行指令用于使计算机执行如以上任一项所述的一种面向SAR图像的小样本语义特征增强的方法。
本发明实施例中提供的技术方案，至少具有如下有益效果：通过对样本集进行迁移学习训练，实现初始化SAR目标图像的深度神经网络，改善了由于SAR目标图像样本量不足，而无法训练得到精确特征值的缺点，并且有效地提升深度神经网络的收敛速度；利用激活函数对深度神经网络进行网络优化，使特征值更加准确；利用自编码器对SAR目标图形的进行特征空间与语义空间的映射，能够增强SAR图像的语义空间和特征空间的表达能力，从而提高SAR图像目标识别的准确率。
附图说明
下面结合附图和实例对本发明作进一步说明：
图1是本发明的面向SAR图像的小样本语义特征增强的方法的一个实施例的整体流程图；
图2是本发明的面向SAR图像的小样本语义特征增强的方法的一个实施例的获取SAR目标图像的样本集，对样本集进行迁移学习训练的流程图；
图3是本发明的面向SAR图像的小样本语义特征增强的方法的一个实施例的利用激活函数对所述深度神经网络进行网络优化的流程图；
图4是本发明的面向SAR图像的小样本语义特征增强的方法的一个实施例的利用自编码器对所述特征图进行特征空间与语义空间的映射的流程图；
图5是本发明的面向SAR图像的小样本语义特征增强的方法的一个实施例的利用自编码器对所述特征图进行特征空间与语义空间的映射的示意图。
具体实施方式
随着粤港澳大湾区建设的推进，湾区的安全监测对SAR图像的应用需求不断增强，但是由于湾区的标注样本量不足，在传统的特征提取模型训练过程中容易出现过拟合、陷入局部最优解、收敛速度慢等问题，造成模型泛化能力退化，并且SAR目标检测对复杂场景多目标的检测还存在一定的难度，影响了湾区目标特征的准确提取。
基于此，本发明提供了一种面向SAR图像的小样本语义特征增强的方法及装置，能够解决湾区的样本量不足的问题，提升网络的收敛速度，以及增强SAR图像的语义特征空间的表达能力，从而提高SAR图像目标识别的准确率。
下面结合附图，对本发明实施例作进一步阐述。
参照图1，本发明的一个实施例提供了一种面向SAR图像的小样本语义特征增强的方法，包括如下步骤：
步骤S100：获取SAR目标图像的样本集，对样本集进行迁移学习训练，得到初始化SAR目标图像的深度神经网络，所述样本集包括SAR目标图像与SAR目标虚拟图像；
步骤S200：利用激活函数对所述深度神经网络进行网络优化，采用优化后的深度神经网络对SAR目标图像进行特征提取，得到特征图；
步骤S300：利用自编码器对所述特征图进行特征空间与语义空间的映射，得到语义特征增强的深度视觉特征。
在本实施例中，步骤S100通过搜集与SAR目标图像的相同类别的不同角度、不同形状等状态的虚拟图像的样本集，通过对虚拟图像的样本集进行训练，得到样本集训练完成后每层对应的特征值；同时，对特征值进行迁移学习，初始化SAR目标图像的深度神经网络，使得深度神经网络的每层网络具有固定的初始化值，而不是随机设置的初始值。通过对虚拟图像的样本集进行迁移学习训练，解决SAR目标图像样本量不足的问题，有效提升深度神经网络收敛速度，提高SAR图像识别的准确率。其中，虚拟图像的样本集可以是互联网上搜集的飞机、船舰等SAR目标图像的正面、侧面，以及各种形状的图像等；初始化深度神经网络的网络层，可以设置为最后一层的全连接层之前的任意网络层，即最后一层的全连接层之前的网络层都可以使用虚拟图像的样本集的特征值进行初始化。
步骤S200利用激活函数对深度神经网络进行网络优化，使网络具备分层的非线性映射的学习能力，使模型训练更加高效；其中激活函数可以采用sigmoid函数、tanh函数、relu函数、maxout函数等。采用优化后的深度神经网络对SAR目标图像进行特征提取，提取的特征更加准确、拟合效果好。
步骤S300利用自编码器对SAR目标图形的特征图进行特征空间与语义空间的映射，能够增强SAR图像的语义 空间和特征空间的表达能力，有效地探索多层特征间的差异性，实现信息的互补与增强，从而提高SAR图像目标识别的准确率。
进一步地，参照图2，本发明的另一个实施例还提供了一种面向SAR图像的小样本语义特征增强的方法，其中，获取SAR目标图像的样本集，对样本集进行迁移学习训练，得到初始化SAR目标图像的深度神经网络，包括如下步骤：
步骤S110：获取SAR目标图像，建立SAR目标图像的深度神经网络；
步骤S120：采集与SAR目标图像相同类别的不同状态的虚拟图像的样本集，利用卷积网络的特征提取层提取样本集的参数和权重；
步骤S130：利用所述样本集的参数和权重初始化SAR目标图像的深度神经网络。
在本实施例中，步骤S110获取SAR目标图像，利用SAR目标图像建立深度神经网络；步骤S120采集与SAR目标图像的相同类别的不同角度、形状等状态的虚拟图像的样本集，并利用卷积网络的特征提取层提取样本集每层的最优参数和权重；步骤S130利用样本集每层的最优参数和权重初始化SAR目标图像的深度神经网络，初始化的网络层设置为深度神经网络的最后一层全连接层之前的任意网络层，即迁移的层数不超过最后一层全连接层所在的层数，使得深度神经网络得到每层固定的初始化值，而不是随机设置的初始值，提高了深度神经网络的学习能力，解决了SAR目标图像样本量不足的问题，从而有效提升深度神经网络的收敛速度，提高了SAR目标图像识别的准确度。其中，SAR目标图像可以是雷达检测得到的飞机、船舰等目标图像，那么虚拟图像的样本集就可以是互联网上搜集飞机、船舰、坦克等目标的正面、侧面、各种形状的图像等。
进一步地，本发明的另一个实施例还提供了一种面向SAR图像的小样本语义特征增强的方法，其中，深度神经网络包括深度残差ResNet网络或DenseNet网络。
在本实施例中，深度神经网络可以设置为深度残差ResNet网络，并且深度残差ResNet网络优选使用了ResNet-50的残差网络结构，即深度残差网络共有50层的网络结构，大大增加了网络的深度，提高特征提取的准确度；同时深度残差ResNet网络中应用了残差块的结构，使网络层在每两层或三层之间发生跳跃连接，解决了网络太深难训练的问题，简化了深度网络的训练步骤，有效提高了对SAR目标图像的识别能力和训练速度。
另外，深度神经网络还可以设置为DenseNet网络，DenseNet网络是一种具有密集连接的卷积神经网络，任意两层网络间都有直接的连接，并且每层的特征图以通道的维度进行合并，有效地减少了特征图的数量，缓解了梯度消失的问题；将SAR目标图像输入到DenseNet网络中进行特征提取，能够加强特征的传播，深度网络训练效果好，提高了SAR目标图像的识别率。
进一步地，参照图3，本发明的另一个实施例还提供了一种面向SAR图像的小样本语义特征增强的方法，其中，利用激活函数对所述深度神经网络进行网络优化，采用优化后的深度神经网络对SAR目标图像进行特征提取，得到特征图，包括如下步骤：
步骤S210：在深度神经网络的最后一层的全连接层后面连接Maxout激活函数层，将SAR目标图像输入到所述含有Maxout激活函数层的深度神经网络中进行特征提取，得到特征的局部最大权重；
步骤S220：在Maxout激活函数层后面连接ReLU激活函数层，将所述局部最大权重输入到ReLU激活函数层进行处理，得到局部最优权重；
步骤S230：在ReLU激活函数层后面连接一层额外的全连接层，将所述局部最优权重输入到额外的全连接层中进行分类处理，得到特征图。
在本实施例中，步骤S210将SAR目标图像输入到含有Maxout激活函数层的深度神经网络中进行特征提取，Maxout激活函数层的每个输入特征由m个可选择的隐含层节点组成，通过以下公式提取SAR目标图像的特征的局部最大权重：
Figure PCTCN2020099880-appb-000001
其中，h i是第i个特征的局部最大权重；x T是输入的特征的权重集合，W ij为d*m*k的三维权重矩阵，b为m*k的二维偏置向量，d表示输入的权重的个数，m表示隐含层节点的个数，其中每个隐含层节点由k个“隐隐含层”节点组成，“隐隐含层”节点的结构类似于隐含层节点，并且k个“隐隐含层”节点都是线性输出的。
因此，Maxout激活函数层是一个可持续学习的激活函数层，变化状态由W确定，并且Maxout激活函数为分段线性函数，可以逼近任意凸函数，拟合能力和非线性表征力较强，能够使网络获得更好的局部解。
步骤S220在Maxout激活函数层后面连接ReLU激活函数层，ReLU激活函数层具有单侧抑制性，把Maxout激活函数层输出的局部最大权重中的负值变为0，而正值不变，具体公式为：
Figure PCTCN2020099880-appb-000002
通过ReLU激活函数层对局部最大权重进行单侧抑制性的处理，得到局部最优解，并且ReLU激活函数层的单侧抑制性使得SAR目标图像的特征具有稀疏激活性，缓解了网络深度带来的梯度发散问题，加快了网络的收敛速度，增强SAR目标图像的特征的泛化能力，使SAR目标图像的特征更具代表性。
因此，Maxout激活函数是分段线性函数，能够对数据进行处理，在维度上减少数据的输出；ReLU激活函数层具有稀疏连接性，使深度神经网络训练过程更加容易；而Maxout激活函数层和ReLU激活函数层的连续连接的设置，构成了双激活函数层，双激活函数层具备了Maxout激活函数和ReLU激活函数的性质，对深度神经网络处理后的SAR目标图像特征的数据具有一定的数据降维能力，并且使网络训练过程更加容易，提高SAR目标图像特征识别能力。
步骤S230在ReLU激活函数层后面连接一层额外的全连接层，额外的全连接层对SAR目标图像的特征，及其局部最优权重重新进行融合分类处理，计算各个特征的概率分布，得到特征图。
进一步地，参照图3，本发明的另一个实施例还提供了一种面向SAR图像的小样本语义特征增强的方法，其中，利用激活函数对所述深度神经网络进行网络优化，采用优化后的深度神经网络对SAR目标图像进行特征提取，得到特征图，还包括如下步骤：
步骤S240：利用Softmax-MSE损失函数对所述特征图进行梯度损失计算，并通过反向调整所述优化后的深度神 经网络使梯度损失最小，得到拟合的特征图。
在本实施例中，在额外的全连接层后连接Softmax-MSE损失函数层，对特征图进行梯度损失计算，假设Softmax-MSE损失函数层的输入参数的个数为m,输入参数集为X＝{x 0,x 1,...,x m-1}，其中，参数集中每个元素表示参数的权重，则Softmax-MSE损失函数层的第k个参数权重经Softmax函数变换得到：
Figure PCTCN2020099880-appb-000003
其中，p k为识别预测值，k∈[0,m-1]，p k＝max([p 0,p 1,...,p m-1])，则参数输出的最终类别预测值为：
Figure PCTCN2020099880-appb-000004
假设SAR目标图像的输入训练样本数量为n，则Softmax-MSE损失函数层的损失值L为：
Figure PCTCN2020099880-appb-000005
其中，y j是
Figure PCTCN2020099880-appb-000006
的期望值，也是第j个图像的标签值，则Softmax-MSE损失函数层的梯度为：
Figure PCTCN2020099880-appb-000007
在Softmax-MSE损失函数计算梯度时，对于网络模型在额外的全连接层中的第i个参数的权重输出，若权重与样本的期望值y i相等，则采用p i-1计入梯度值，若不相等，则采用p i计入梯度值。
计算Softmax-MSE损失函数层的梯度值总和，若梯度值总和较大，则通过反向调整优化后的深度神经网络的初始化参数的权重大小或学习速率等因素，使得Softmax-MSE损失函数层的梯度值总和较小，增强特征图的拟合程度。
进一步地，本发明的另一个实施例还提供了一种面向SAR图像的小样本语义特征增强的方法，其中，自编码器包括编码器和译码器。
在本实施例中，自编码器包括编码器和译码器，编码器能够把原始输入集转换为中间输出集，一般中间输出集的维度比原始输入集低，即编码器能够降低了原始输入集的维度，并将原始输入集的数据进行合并和组合，实现信息的互补，转化了数据之间的关系，提高了数据表达能力；译码器能够接收编码器的中间输出集，并对中间输出集进行数据的重构，将数据转化为原始输入集的维度的空间架构。其中，原始输入集可以是图像、向量、视频或音频数据等。
进一步地，参照图4和图5，本发明的另一个实施例还提供了一种面向SAR图像的小样本语义特征增强的方法，其中，利用自编码器对所述特征图进行特征空间与语义空间的映射，得到语义特征增强的深度视觉特征，包括如下步骤：
步骤S310：将所述特征图输入到一个神经网络中，利用编码器提取所述神经网络每层的特征值；
步骤S320：利用高斯白噪声对所述特征值进行仿真处理，得到仿真特征值；
步骤S330：利用所述编码器将所述仿真特征值进行特征空间向语义空间的映射，得到语义增强的语义特征值；
步骤S340：采用译码器对所述语义特征值进行语义空间向特征空间的映射，得到特征增强的深度视觉特征。
在本实施例中，步骤S310将特征图输入到一个神经网络中，并利用编码器提取神经网络每层的特征值；步骤S320的高斯白噪声是概率密度服从正态分布的噪声，在真实环境中，噪音往往不是由单一源头造成的，而是不同来源的噪音复合体，即真实噪音是由不同概率分布的随机变量组成的，并且每一个随机变量都是独立的。根据中心极限定理可知，真实噪音的归一化总会随着噪音源数量的上升，而趋近于一个高斯分布，因此，利用高斯白噪声对特征值进行仿真处理，能够更好地模拟未知的真实噪音，得到仿真特征值。
步骤S330利用编码器对仿真特征值进行处理，使每层的特征值具有不同的编码信息，其中每个编码信息代表不同级别的抽象语义信息，并且编码器有效地利用了仿真特征值的编码信息的差异性，使编码信息之间进行合并和组合，实现信息的互补，使编码信息能够匹配神经网络每层的输出特征空间，从而使仿真特征值实现从特征空间向语义空间的映射，得到语义特征增强的语义特征值；并且所述语义特征值含有所属类别的语义特征标签，例如，所属类别有飞机、船舰、坦克等。
步骤S340的译码器与编码器具有完全相同的架构，采用译码器对语义特征值进行处理，通过变形、尺寸变换和拼接的方法，将增强的语义空间重新映射投射到神经网络的特征空间，得到语义特征空间的表达能力增强的深度视觉特征，丰富了特征的表达力，提升了SAR目标识别性能。
另外，参照图1-图4，本发明的另一个实施例还提供了一种面向SAR图像的小样本语义特征增强的方法。该识别方法包括如下步骤：
步骤S110：获取SAR目标图像，建立SAR目标图像的深度神经网络；
步骤S120：采集与SAR目标图像相同类别的不同状态的虚拟图像的样本集，利用卷积网络的特征提取层提取样本集的参数和权重；
步骤S130：利用所述样本集的参数和权重初始化SAR目标图像的深度神经网络；
步骤S210：在深度神经网络的最后一层的全连接层后面连接Maxout激活函数层，将SAR目标图像输入到所述含有Maxout激活函数层的深度神经网络中进行特征提取，得到特征的局部最大权重；
步骤S220：在Maxout激活函数层后面连接ReLU激活函数层，将所述局部最大权重输入到ReLU激活函数层进行处理，得到局部最优权重；
步骤S230：在ReLU激活函数层后面连接一层额外的全连接层，将所述局部最优权重输入到额外的全连接层中进行分类处理，得到特征图；
步骤S240：利用Softmax-MSE损失函数对所述特征图进行梯度损失计算，并通过反向调整所述优化后的深度神经网络使梯度损失最小，得到拟合的特征图；
步骤S310：将所述特征图输入到一个神经网络中，利用编码器提取所述神经网络每层的特征值；
步骤S320：利用高斯白噪声对所述特征值进行仿真处理，得到仿真特征值；
步骤S330：利用所述编码器将所述仿真特征值进行特征空间向语义空间的映射，得到语义增强的语义特征值；
步骤S340：采用译码器对所述语义特征值进行语义空间向特征空间的映射，得到特征增强的深度视觉特征。
在本实施例中，步骤S110获取SAR目标图像，利用SAR目标图像建立深度残差ResNet网络，深度残差ResNet网络优选使用了ResNet-50残差网络的结构，大大增加了网络的深度，提高特征提取的准确度；同时深度残差ResNet网络中应用了残差块的结构，使网络层在每两层或三层之间发生跳跃连接，解决了网络太深难训练的问题，简化了深度网络的训练步骤，有效提高了对SAR目标图像的识别能力和训练速度。特别地，深度残差ResNet网络的滤波器选用5*5、3*3、1*1的规格，学习率选取0.001，训练代数为100，训练测试批次分别为64和32。
步骤S120采集与SAR目标图像的相同类别的不同角度、形状等状态的虚拟图像的样本集，并利用卷积网络的特征提取层提取样本集每层的最优参数和权重；步骤S130利用样本集每层的最优参数和权重初始化SAR目标图像的深度残差ResNet网络，初始化的网络层设置为深度残差ResNet网络的最后一层全连接层之前的任意网络层，即迁移的层数不超过最后一层全连接层所在的层数，使得深度残差ResNet网络得到每层固定的初始化值，而不是随机设置的初始值，提高了深度残差ResNet网络的学习能力，解决了SAR目标图像样本量不足的问题，从而有效提升深度残差ResNet网络的收敛速度，提高了SAR目标图像识别的准确度。其中，SAR目标图像可以是雷达检测得到的飞机、船舰等目标图像，那么虚拟图像的样本集就可以是互联网上搜集飞机、船舰、坦克等目标的正面、侧面、各种形状的图像等。
步骤S210将SAR目标图像输入到含有Maxout激活函数层的深度残差ResNet网络中进行特征提取，Maxout激活函数层的每个输入特征由m个可选择的隐含层节点组成，通过以下公式提取SAR目标图像的特征的局部最大权重：
Figure PCTCN2020099880-appb-000008
其中，h i是第i个特征的局部最大权重；x T是输入的特征的权重集合，W ij为d*m*k的三维权重矩阵，b为m*k的二维偏置向量，d表示输入的权重的个数，m表示隐含层节点的个数，其中每个隐含层节点由k个“隐隐含层”节点组成，“隐隐含层”节点的结构类似于隐含层节点，并且k个“隐隐含层”节点都是线性输出的。
因此，Maxout激活函数层是一个可持续学习的激活函数层，变化状态由W确定，并且Maxout激活函数为分段线性函数，可以逼近任意凸函数，拟合能力和非线性表征力较强，能够使网络获得更好的局部解。
步骤S220在Maxout激活函数层后面连接ReLU激活函数层，ReLU激活函数层具有单侧抑制性，把Maxout激活函数层输出的局部最大权重中的负值变为0，而正值不变，具体公式为：
Figure PCTCN2020099880-appb-000009
通过ReLU激活函数层对局部最大权重进行单侧抑制性的处理，得到局部最优解，并且ReLU激活函数层的单侧抑制性使得SAR目标图像的特征具有稀疏激活性，缓解了网络深度带来的梯度发散问题，加快了网络的收敛速度，增强SAR目标图像的特征的泛化能力，使SAR目标图像的特征更具代表性。
因此，Maxout激活函数是分段线性函数，能够对数据进行处理，在维度上减少数据的输出；ReLU激活函数层具有稀疏连接性，使深度残差ResNet网络训练过程更加容易；而Maxout激活函数层和ReLU激活函数层的连续连接的设置，构成了双激活函数层，双激活函数层具备了Maxout激活函数和ReLU激活函数的性质，对深度残差ResNet网络处理后的SAR目标图像特征的数据具有一定的数据降维能力，并且使网络训练过程更加容易，提高SAR目标图像特征识别能力。
步骤S230在ReLU激活函数层后面连接一层额外的全连接层，额外的全连接层对SAR目标图像的特征，及其局部最优权重重新进行融合分类处理，计算各个特征的概率分布，得到特征图。
步骤S240在额外的全连接层后连接Softmax-MSE损失函数层，对特征图进行梯度损失计算，假设Softmax-MSE损失函数层的输入参数的个数为m,输入参数集为X＝{x 0,x 1,...,x m-1}，其中，参数集中每个元素表示参数的权重，则Softmax-MSE损失函数层的第k个参数权重经Softmax函数变换得到：
Figure PCTCN2020099880-appb-000010
其中，p k为识别预测值，k∈[0,m-1]，p k＝max([p 0,p 1,...,p m-1])，则参数输出的最终类别预测值为：
Figure PCTCN2020099880-appb-000011
假设SAR目标图像的输入训练样本数量为n，则Softmax-MSE损失函数层的损失值L为：
Figure PCTCN2020099880-appb-000012
其中，y j是
Figure PCTCN2020099880-appb-000013
的期望值，也是第j个图像的标签值，则Softmax-MSE损失函数层的梯度为：
Figure PCTCN2020099880-appb-000014
在Softmax-MSE损失函数计算梯度时，对于网络模型在额外的全连接层中的第i个参数的权重输出，若权重与样本的期望值y i相等，则采用p i-1计入梯度值，若不相等，则采用p i计入梯度值。
计算Softmax-MSE损失函数层的梯度值总和，若梯度值总和较大，则通过反向调整优化后的深度残差ResNet网络的初始化参数的权重大小或学习速率等因素，使得Softmax-MSE损失函数层的梯度值总和较小，增强特征图的拟合程度。
参照图5，步骤S310将特征图输入到一个神经网络中，并利用编码器提取神经网络每层的特征值；步骤S320的高斯白噪声是概率密度服从正态分布的噪声，在真实环境中，噪音往往不是由单一源头造成的，而是不同来源的噪音复合体，即真实噪音是由不同概率分布的随机变量组成的，并且每一个随机变量都是独立的。根据中心极限定 理可知，真实噪音的归一化总会随着噪音源数量的上升，而趋近于一个高斯分布，因此，利用高斯白噪声对特征值进行仿真处理，能够更好地模拟未知的真实噪音，得到仿真特征值。
步骤S330利用编码器对仿真特征值进行处理，使每层的特征值具有不同的编码信息，其中每个编码信息代表不同级别的抽象语义信息，并且编码器有效地利用了仿真特征值的编码信息的差异性，使编码信息之间进行合并和组合，实现信息的互补，使编码信息能够匹配神经网络每层的输出特征空间，从而使仿真特征值实现从特征空间向语义空间的映射，得到语义特征增强的语义特征值；并且所述语义特征值含有所属类别的语义特征标签，例如，所属类别有飞机、船舰、坦克等。
步骤S340的译码器与编码器具有完全相同的架构，采用译码器对语义特征值进行处理，通过变形、尺寸变换和拼接的方法，将增强的语义空间重新映射投射到神经网络的特征空间，得到语义特征空间的表达能力增强的深度视觉特征，丰富了特征的表达力，提升了SAR目标识别性能。
此外，本发明的另一个实施例还提供了一种面向SAR图像的小样本语义特征增强装置，包括至少一个控制处理器和用于与所述至少一个控制处理器通信连接的存储器；所述存储器存储有可被所述至少一个控制处理器执行的指令，所述指令被所述至少一个控制处理器执行，以使所述至少一个控制处理器能够执行如上的任一项所述的一种面向SAR图像的小样本语义特征增强的方法。
在本实施例中，特征增强装置包括：一个或多个控制处理器和存储器，控制处理器和存储器可以通过总线或者其他方式连接。
存储器作为一种非暂态计算机可读存储介质，可用于存储非暂态软件程序、非暂态性计算机可执行程序以及模块，如本发明实施例中的特征增强方法对应的程序指令/模块。控制处理器通过运行存储在存储器中的非暂态软件程序、指令以及模块，从而执行特征增强装置的各种功能应用以及数据处理，即实现上述方法实施例的特征增强的方法。
存储器可以包括存储程序区和存储数据区，其中，存储程序区可存储操作系统、至少一个功能所需要的应用程序；存储数据区可存储根据特征增强装置的使用所创建的数据等。此外，存储器可以包括高速随机存取存储器，还可以包括非暂态存储器，例如至少一个磁盘存储器件、闪存器件、或其他非暂态固态存储器件。在一些实施方式中，存储器可选包括相对于控制处理器远程设置的存储器，这些远程存储器可以通过网络连接至该特征增强装置。上述网络的实例包括但不限于互联网、企业内部网、局域网、移动通信网及其组合。
所述一个或者多个模块存储在所述存储器中，当被所述一个或者多个控制处理器执行时，执行上述方法实施例中的特征增强方法，例如，执行以上描述特征增强方法步骤S100至S300、S110至S130、S210至S240，以及S310至S340的功能。
本发明实施例还提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可执行指令，该计算机可执行指令被一个或多个控制处理器执行，例如，一个控制处理器执行，可使得上述一个或多个控制处理器执行上述方法实施例中的特征增强方法，例如，执行以上描述的方法步骤S100至S300、S110至S130、S210至S240，以及S310至S340的功能。
以上所描述的装置实施例仅仅是示意性的，其中所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。
通过以上的实施方式的描述，本领域技术人员可以清楚地了解到各实施方式可借助软件加通用硬件平台的方式来实现。本领域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，所述的程序可存储于一计算机可读取存储介质中，该程序在执行时，可包括如上述方法的实施例的流程。其中，所述的存储介质可为磁碟、光盘、只读存储记忆体(ReadOnly Memory,ROM)或随机存储记忆体(Random Access Memory,RAM)等。
以上是对本发明的较佳实施进行了具体说明，但本发明并不局限于上述实施方式，熟悉本领域的技术人员在不违背本发明精神的前提下还可作出种种的等同变形或替换，这些等同的变形或替换均包含在本申请权利要求所限定的范围内。

Claims (9)

  1. 一种面向SAR图像的小样本语义特征增强的方法，其特征在于：包括如下步骤：
    获取SAR目标图像的样本集，对样本集进行迁移学习训练，得到初始化SAR目标图像的深度神经网络，所述样本集包括SAR目标图像与SAR目标虚拟图像；
    利用激活函数对所述深度神经网络进行网络优化，采用优化后的深度神经网络对SAR目标图像进行特征提取，得到特征图；
    利用自编码器对所述特征图进行特征空间与语义空间的映射，得到语义特征增强的深度视觉特征。
  2. 根据权利要求1所述的一种面向SAR图像的小样本语义特征增强的方法，其特征在于：所述获取SAR目标图像的样本集，对样本集进行迁移学习训练，得到初始化SAR目标图像的深度神经网络，包括如下步骤：
    获取SAR目标图像，建立SAR目标图像的深度神经网络；
    采集与SAR目标图像相同类别的不同状态的虚拟图像的样本集，利用卷积网络的特征提取层提取样本集的参数和权重；
    利用所述样本集的参数和权重初始化SAR目标图像的深度神经网络。
  3. 根据权利要求2所述的一种面向SAR图像的小样本语义特征增强的方法，其特征在于：所述深度神经网络包括深度残差ResNet网络或DenseNet网络。
  4. 根据权利要求1所述的一种面向SAR图像的小样本语义特征增强的方法，其特征在于：所述利用激活函数对所述深度神经网络进行网络优化，采用优化后的深度神经网络对SAR目标图像进行特征提取，得到特征图，包括如下步骤：
    在深度神经网络的最后一层的全连接层后面连接Maxout激活函数层，将SAR目标图像输入到所述含有Maxout激活函数层的深度神经网络中进行特征提取，得到特征的局部最大权重；
    在Maxout激活函数层后面连接ReLU激活函数层，将所述局部最大权重输入到ReLU激活函数层进行处理，得到局部最优权重；
    在ReLU激活函数层后面连接一层额外的全连接层，将所述局部最优权重输入到额外的全连接层中进行分类处理，得到特征图。
  5. 根据权利要求4所述的一种面向SAR图像的小样本语义特征增强的方法，其特征在于：所述利用激活函数对所述深度神经网络进行网络优化，采用优化后的深度神经网络对SAR目标图像进行特征提取，得到特征图，还包括如下步骤：
    利用Softmax-MSE损失函数对所述特征图进行梯度损失计算，并通过反向调整所述优化后的深度神经网络使梯度损失最小，得到拟合的特征图。
  6. 根据权利要求1所述的一种面向SAR图像的小样本语义特征增强的方法，其特征在于：所述自编码器包括编码器和译码器。
  7. 根据权利要求6所述的一种面向SAR图像的小样本语义特征增强的方法，其特征在于：所述利用自编码器对所述特征图进行特征空间与语义空间的映射，得到语义特征增强的深度视觉特征，包括如下步骤：
    将所述特征图输入到一个神经网络中，利用编码器提取所述神经网络每层的特征值；
    利用高斯白噪声对所述特征值进行仿真处理，得到仿真特征值；
    利用所述编码器将所述仿真特征值进行特征空间向语义空间的映射，得到语义增强的语义特征值；
    采用译码器对所述语义特征值进行语义空间向特征空间的映射，得到特征增强的深度视觉特征。
  8. 一种面向SAR图像的小样本语义特征增强装置，其特征在于，包括至少一个控制处理器和用于与所述至少一个控制处理器通信连接的存储器；所述存储器存储有可被所述至少一个控制处理器执行的指令，所述指令被所述至少一个控制处理器执行，以使所述至少一个控制处理器能够执行如权利要求1-7任一项所述的一种面向SAR图像的小样本语义特征增强的方法。
  9. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质存储有计算机可执行指令，所述计算机可执行指令用于使计算机执行如权利要求1-7任一项所述的一种面向SAR图像的小样本语义特征增强的方法。
PCT/CN2020/099880 2019-07-02 2020-07-02 一种面向sar图像的小样本语义特征增强的方法及装置 WO2021000906A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910588329.7 2019-07-02
CN201910588329.7A CN110472483B (zh) 2019-07-02 2019-07-02 一种面向sar图像的小样本语义特征增强的方法及装置

Publications (1)

Publication Number Publication Date
WO2021000906A1 true WO2021000906A1 (zh) 2021-01-07

Family

ID=68507043

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/099880 WO2021000906A1 (zh) 2019-07-02 2020-07-02 一种面向sar图像的小样本语义特征增强的方法及装置

Country Status (3)

Country Link
US (1) US11402496B2 (zh)
CN (1) CN110472483B (zh)
WO (1) WO2021000906A1 (zh)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112801166A (zh) * 2021-01-25 2021-05-14 上海海事大学 一种基于特征数据及维度缩放的黑色素瘤分类方法
CN112990334A (zh) * 2021-03-29 2021-06-18 西安电子科技大学 基于改进原型网络的小样本sar图像目标识别方法
CN113095416A (zh) * 2021-04-16 2021-07-09 西安电子科技大学 基于混合损失与图注意力的小样本sar目标分类方法
CN113111909A (zh) * 2021-03-04 2021-07-13 西北工业大学 一种面向训练目标视角不完备的sar目标识别的自学习方法
CN115169499A (zh) * 2022-08-03 2022-10-11 中国电子科技集团公司信息科学研究院 资产数据降维方法、装置、电子设备和计算机存储介质
CN116012364A (zh) * 2023-01-28 2023-04-25 北京建筑大学 Sar图像变化检测方法和装置
CN116071609A (zh) * 2023-03-29 2023-05-05 中国科学技术大学 基于目标特征动态自适应提取的小样本图像分类方法
CN117422855A (zh) * 2023-12-19 2024-01-19 浙江省北大信息技术高等研究院 面向机器视觉的图像预处理方法、装置、设备及存储介质

Families Citing this family (34)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110472483B (zh) * 2019-07-02 2022-11-15 五邑大学 一种面向sar图像的小样本语义特征增强的方法及装置
CN111126226B (zh) * 2019-12-17 2023-07-28 杭州电子科技大学 一种基于小样本学习和特征增强的辐射源个体识别方法
CN111050174A (zh) * 2019-12-27 2020-04-21 清华大学 图像压缩方法、装置及系统
CN111432211B (zh) * 2020-04-01 2021-11-12 山东浪潮科学研究院有限公司 一种用于视频编码的残差信息压缩方法
CN111507388A (zh) * 2020-04-10 2020-08-07 上海眼控科技股份有限公司 一种天气图像分类方法及设备
CN112070123B (zh) * 2020-08-14 2023-11-24 五邑大学 小样本sar图像识别方法、装置及存储介质
US11300652B1 (en) * 2020-10-30 2022-04-12 Rebellion Defense, Inc. Systems and methods for generating images from synthetic aperture radar data using neural networks
CN112926381B (zh) * 2021-01-08 2022-06-21 浙江大学 一种基于激光检测技术的水下目标识别系统
CN112731330B (zh) * 2021-01-12 2022-04-19 西安电子科技大学 基于迁移学习的雷达载频参数变化稳健目标识别方法
CN113030001B (zh) * 2021-03-19 2022-08-23 北京农业智能装备技术研究中心 一种水果糖度检测方法及系统
CN113111758B (zh) * 2021-04-06 2024-01-12 中山大学 一种基于脉冲神经网络的sar图像舰船目标识别方法
CN112990365B (zh) * 2021-04-22 2021-08-17 宝略科技(浙江)有限公司 一种面向遥感图像语义分割的深度学习模型的训练方法
CN113763296A (zh) * 2021-04-28 2021-12-07 腾讯云计算（北京）有限责任公司 图像处理方法、设备以及介质
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
CN113283390B (zh) * 2021-06-24 2022-03-08 中国人民解放军国防科技大学 基于门控多尺度匹配网络的sar图像小样本目标识别方法
CN113657181B (zh) * 2021-07-23 2024-01-23 西北工业大学 一种基于平滑标签编码与特征增强的sar图像旋转目标检测方法
CN113657196B (zh) * 2021-07-27 2024-05-28 中国科学院自动化研究所 Sar图像目标检测方法、装置、电子设备和存储介质
CN113569960B (zh) * 2021-07-29 2023-12-26 北京邮电大学 基于域适应的小样本图像分类方法及系统
CN113657252B (zh) * 2021-08-16 2023-07-28 福州大学 基于编码解码器的高效sar图像船舶目标检测方法
CN113688742B (zh) * 2021-08-26 2023-10-20 五邑大学 一种sar图像识别方法、装置及存储介质
CN113705570B (zh) * 2021-08-31 2023-12-08 长沙理工大学 一种基于深度学习的少样本目标检测方法
CN113902978B (zh) * 2021-09-10 2024-06-18 长沙理工大学 基于深度学习的可解释性sar图像目标检测方法及系统
CN113989665B (zh) * 2021-10-25 2023-04-07 电子科技大学 一种基于路由聚合感知fpn的sar船只检测方法
CN114677566B (zh) * 2022-04-08 2023-10-17 北京百度网讯科技有限公司 深度学习模型的训练方法、对象识别方法和装置
CN114792398B (zh) * 2022-06-23 2022-09-27 阿里巴巴（中国）有限公司 图像分类的方法、存储介质、处理器及系统
CN115239946B (zh) * 2022-06-30 2023-04-07 锋睿领创(珠海)科技有限公司 小样本迁移学习训练、目标检测方法、装置、设备和介质
CN115345322B (zh) * 2022-10-19 2023-02-07 电子科技大学长三角研究院(衢州) 一种基于层级化元迁移的小样本雷达目标识别方法
CN116091867B (zh) * 2023-01-12 2023-09-29 北京邮电大学 一种模型训练、图像识别方法、装置、设备及存储介质
CN116544146B (zh) * 2023-05-22 2024-04-09 浙江固驰电子有限公司 功率半导体器件真空烧结设备及方法
CN117115015B (zh) * 2023-08-02 2024-05-28 中国人民解放军61540部队 一种sar海洋图像中海浪抑制方法、系统、设备及介质
CN116912699B (zh) * 2023-09-06 2023-12-05 交通运输部天津水运工程科学研究所 一种基于图像处理的港口溢油扩散趋势预测方法及系统
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
CN117432414B (zh) * 2023-12-20 2024-03-19 中煤科工开采研究院有限公司 顶板磨砂射流成缝的调控方法及系统
CN117789185B (zh) * 2024-02-28 2024-05-10 浙江驿公里智能科技有限公司 基于深度学习的汽车油孔姿态识别系统及方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108710826A (zh) * 2018-04-13 2018-10-26 燕山大学 一种交通标志深度学习模式识别方法
CN109214401A (zh) * 2017-06-30 2019-01-15 清华大学 基于层次化自动编码器的sar图像分类方法及装置
US20190065910A1 (en) * 2017-08-22 2019-02-28 Northrop Grumman Systems Corporation Adaptive real-time detection and examination network (arden)
CN109492556A (zh) * 2018-10-28 2019-03-19 北京化工大学 面向小样本残差学习的合成孔径雷达目标识别方法
CN109584279A (zh) * 2017-09-29 2019-04-05 中国科学院遥感与数字地球研究所 基于sar图像的船舶检测方法、装置及船舶检测系统
CN110472483A (zh) * 2019-07-02 2019-11-19 五邑大学 一种面向sar图像的小样本语义特征增强的方法及装置

Family Cites Families (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB0701869D0 (en) * 2007-01-31 2007-03-14 Cambridge Consultants Adaptive radar
CN106067042B (zh) * 2016-06-13 2019-02-15 西安电子科技大学 基于半监督深度稀疏滤波网络的极化sar分类方法
CN106408562B (zh) * 2016-09-22 2019-04-09 华南理工大学 基于深度学习的眼底图像视网膜血管分割方法及系统
CN107330446B (zh) * 2017-06-05 2020-08-04 浙江工业大学 一种面向图像分类的深度卷积神经网络的优化方法
US11113599B2 (en) * 2017-06-22 2021-09-07 Adobe Inc. Image captioning utilizing semantic text modeling and adversarial learning
US10935982B2 (en) * 2017-10-04 2021-03-02 Huawei Technologies Co., Ltd. Method of selection of an action for an object using a neural network
CN108564097B (zh) * 2017-12-05 2020-09-22 华南理工大学 一种基于深度卷积神经网络的多尺度目标检测方法
CN108647741A (zh) * 2018-05-18 2018-10-12 湖北工业大学 一种基于迁移学习的图像分类方法和系统
CN108986124A (zh) * 2018-06-20 2018-12-11 天津大学 结合多尺度特征卷积神经网络视网膜血管图像分割方法
US10248664B1 (en) * 2018-07-02 2019-04-02 Inception Institute Of Artificial Intelligence Zero-shot sketch-based image retrieval techniques using neural networks for sketch-image recognition and retrieval
CN109145939B (zh) * 2018-07-02 2021-11-02 南京师范大学 一种小目标敏感的双通道卷积神经网络语义分割方法
CN109102005B (zh) * 2018-07-23 2020-08-18 杭州电子科技大学 基于浅层模型知识迁移的小样本深度学习方法
CN109472316B (zh) * 2018-11-16 2021-12-28 南通大学 一种基于深度学习的滤棒装盒质量识别方法
CN109711413B (zh) * 2018-12-30 2023-04-07 陕西师范大学 基于深度学习的图像语义分割方法
CN109934282B (zh) * 2019-03-08 2022-05-31 哈尔滨工程大学 一种基于sagan样本扩充和辅助信息的sar目标分类方法
CN109886359B (zh) * 2019-03-25 2021-03-16 西安电子科技大学 基于卷积神经网络的小目标检测方法及检测系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109214401A (zh) * 2017-06-30 2019-01-15 清华大学 基于层次化自动编码器的sar图像分类方法及装置
US20190065910A1 (en) * 2017-08-22 2019-02-28 Northrop Grumman Systems Corporation Adaptive real-time detection and examination network (arden)
CN109584279A (zh) * 2017-09-29 2019-04-05 中国科学院遥感与数字地球研究所 基于sar图像的船舶检测方法、装置及船舶检测系统
CN108710826A (zh) * 2018-04-13 2018-10-26 燕山大学 一种交通标志深度学习模式识别方法
CN109492556A (zh) * 2018-10-28 2019-03-19 北京化工大学 面向小样本残差学习的合成孔径雷达目标识别方法
CN110472483A (zh) * 2019-07-02 2019-11-19 五邑大学 一种面向sar图像的小样本语义特征增强的方法及装置

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112801166A (zh) * 2021-01-25 2021-05-14 上海海事大学 一种基于特征数据及维度缩放的黑色素瘤分类方法
CN112801166B (zh) * 2021-01-25 2024-03-01 上海海事大学 一种基于特征数据及维度缩放的黑色素瘤分类方法
CN113111909B (zh) * 2021-03-04 2024-03-12 西北工业大学 一种面向训练目标视角不完备的sar目标识别的自学习方法
CN113111909A (zh) * 2021-03-04 2021-07-13 西北工业大学 一种面向训练目标视角不完备的sar目标识别的自学习方法
CN112990334A (zh) * 2021-03-29 2021-06-18 西安电子科技大学 基于改进原型网络的小样本sar图像目标识别方法
CN113095416A (zh) * 2021-04-16 2021-07-09 西安电子科技大学 基于混合损失与图注意力的小样本sar目标分类方法
CN113095416B (zh) * 2021-04-16 2023-08-18 西安电子科技大学 基于混合损失与图注意力的小样本sar目标分类方法
CN115169499A (zh) * 2022-08-03 2022-10-11 中国电子科技集团公司信息科学研究院 资产数据降维方法、装置、电子设备和计算机存储介质
CN115169499B (zh) * 2022-08-03 2024-04-05 中国电子科技集团公司信息科学研究院 资产数据降维方法、装置、电子设备和计算机存储介质
CN116012364A (zh) * 2023-01-28 2023-04-25 北京建筑大学 Sar图像变化检测方法和装置
CN116012364B (zh) * 2023-01-28 2024-01-16 北京建筑大学 Sar图像变化检测方法和装置
CN116071609A (zh) * 2023-03-29 2023-05-05 中国科学技术大学 基于目标特征动态自适应提取的小样本图像分类方法
CN117422855A (zh) * 2023-12-19 2024-01-19 浙江省北大信息技术高等研究院 面向机器视觉的图像预处理方法、装置、设备及存储介质
CN117422855B (zh) * 2023-12-19 2024-05-03 浙江省北大信息技术高等研究院 面向机器视觉的图像预处理方法、装置、设备及存储介质

Also Published As

Publication number Publication date
CN110472483A (zh) 2019-11-19
CN110472483B (zh) 2022-11-15
US11402496B2 (en) 2022-08-02
US20210003700A1 (en) 2021-01-07

Similar Documents

Publication Publication Date Title
WO2021000906A1 (zh) 一种面向sar图像的小样本语义特征增强的方法及装置
CN110472627B (zh) 一种端到端的sar图像识别方法、装置及存储介质
Zhu et al. Deep learning meets SAR: Concepts, models, pitfalls, and perspectives
Chen et al. Deep learning for autonomous ship-oriented small ship detection
Han et al. Underwater image processing and object detection based on deep CNN method
US11610115B2 (en) Learning to generate synthetic datasets for training neural networks
US10733755B2 (en) Learning geometric differentials for matching 3D models to objects in a 2D image
CN111754596B (zh) 编辑模型生成、人脸图像编辑方法、装置、设备及介质
WO2017096758A1 (zh) 图像分类方法、电子设备和存储介质
DE102021119726A1 (de) Dreidimensionale objektrekonstruktion aus einem video
CN107169954B (zh) 一种基于并行卷积神经网络的图像显著性检测方法
CN113743417B (zh) 语义分割方法和语义分割装置
US11270425B2 (en) Coordinate estimation on n-spheres with spherical regression
CN116110022B (zh) 基于响应知识蒸馏的轻量化交通标志检测方法及系统
DE102022100360A1 (de) Framework für maschinelles lernen angewandt bei einer halbüberwachten einstellung, um instanzenverfolgung in einer sequenz von bildframes durchzuführen
DE102022113244A1 (de) Gemeinsame Form- und Erscheinungsbildoptimierung durch Topologieabtastung
Abbott et al. Deep object classification in low resolution lwir imagery via transfer learning
Yang et al. Multiple heterogeneous P-DCNNs ensemble with stacking algorithm: A novel recognition method of space target ISAR images under the condition of small sample set
CN113536920B (zh) 一种半监督三维点云目标检测方法
Xie et al. High-efficiency and low-energy ship recognition strategy based on spiking neural network in SAR images
CN111209813B (zh) 基于迁移学习的遥感图像语义分割方法
Luo et al. An improved gesture segmentation method for gesture recognition based on CNN and YCbCr
Huang et al. EST-YOLOv5s: SAR image aircraft target detection model based on improved YOLOv5s
CN116503602A (zh) 基于多层级边缘增强的非结构化环境三维点云语义分割方法
Noor et al. A hybrid deep learning model for UAVs detection in day and night dual visions

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: 20835544

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20835544

Country of ref document: EP

Kind code of ref document: A1