CN114863283A - SAR image target identification method combining transfer learning and attention mechanism - Google Patents
SAR image target identification method combining transfer learning and attention mechanism Download PDFInfo
- Publication number
- CN114863283A CN114863283A CN202210579119.3A CN202210579119A CN114863283A CN 114863283 A CN114863283 A CN 114863283A CN 202210579119 A CN202210579119 A CN 202210579119A CN 114863283 A CN114863283 A CN 114863283A
- Authority
- CN
- China
- Prior art keywords
- attention mechanism
- weight
- sar image
- channel
- input
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000007246 mechanism Effects 0.000 title claims abstract description 108
- 238000000034 method Methods 0.000 title claims abstract description 46
- 238000013526 transfer learning Methods 0.000 title claims abstract description 17
- 238000010586 diagram Methods 0.000 claims abstract description 21
- 238000013508 migration Methods 0.000 claims abstract description 15
- 230000005012 migration Effects 0.000 claims abstract description 15
- 230000006870 function Effects 0.000 claims description 21
- 238000011176 pooling Methods 0.000 claims description 19
- 210000002569 neuron Anatomy 0.000 claims description 12
- 239000013598 vector Substances 0.000 claims description 10
- 239000011159 matrix material Substances 0.000 claims description 8
- 238000012549 training Methods 0.000 claims description 8
- 238000001914 filtration Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 4
- 230000008569 process Effects 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 2
- 239000010410 layer Substances 0.000 description 34
- 238000012216 screening Methods 0.000 description 3
- 238000013136 deep learning model Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 230000035515 penetration Effects 0.000 description 2
- 230000003213 activating effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 239000004973 liquid crystal related substance Substances 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an SAR image target identification method combining transfer learning and attention mechanism, which comprises the following steps of S1, extracting a convolution characteristic graph from an input SAR image through a convolution kernel of 7 × 7; s2, inputting the obtained convolution characteristic diagram into an attention mechanism for focusing to obtain a primary focusing characteristic diagram; s3, inputting the initial focusing feature map into the fine-tuned residual error network to obtain a weight feature map, wherein the residual error network is formed by combining a deep migration learning method, and a weight model trained by the residual error network on a known data set is migrated into the SAR image recognition work; s4, obtaining a weight characteristic diagram containing the weight, inputting the weight characteristic diagram into an attention mechanism again for focusing; s5, inputting the focused target recognition feature map containing the weight into a subsequent convolutional layer for classification to obtain a target recognition result.
Description
Technical Field
The invention relates to a method for recognizing an SAR image target by combining transfer learning and an attention mechanism.
Background
The SAR is used as a high-resolution imaging radar, is not influenced by conditions such as weather, illumination and the like, has certain ground surface penetration capacity, can realize penetration detection of a hidden target, and can realize continuous earth observation all day long and all weather. These excellent properties have led to an increasing range of SAR applications in both civilian and military applications.
In recent years, the technology of SAR has been developed faster and faster, the imaging quality has been better and better, and the image resolution has been higher and higher, but the Automatic Target Recognition (ATR) based on SAR images has been developed relatively slowly. The difficulties of SAR ATR are mainly reflected in two aspects: (1) the influence of noise interference, especially the model performance is reduced under the condition of serious noise, and even the target problem cannot be correctly identified; (2) the information quantity is large, the information is messy and not focused enough, and the related information cannot be identified quickly and accurately.
Disclosure of Invention
The invention aims to solve the technical problems in the prior art and provides an SAR image target identification method combining transfer learning and attention mechanism.
In order to achieve the purpose, the technical scheme provided by the invention is as follows: a SAR image target identification method combining transfer learning and attention mechanism comprises the following steps:
s1, extracting a convolution feature map from the input SAR image through 7 × 7 convolution kernels;
s2, inputting the obtained convolution characteristic diagram into an attention mechanism for focusing to obtain a primary focusing characteristic diagram, wherein the attention mechanism is used for filtering interest points from a large amount of digital image information, selectively filtering important information and focusing;
s3, inputting the initial focusing feature map into the fine-tuned residual error network to obtain a weight feature map, wherein the residual error network is formed by utilizing a depth convolution network to complete an end-to-end target identification task and then combining a depth migration learning method, a weight model trained by the residual error network on a known data set is migrated into the SAR image identification work, and the SAR image identification model training time is accelerated;
s4, obtaining a weight characteristic diagram containing the weight, inputting the weight characteristic diagram into an attention mechanism again for focusing;
and S5, inputting the focused target recognition feature map containing the weight into a subsequent convolutional layer for classification to obtain a target recognition result.
Preferably, the deep migration learning method in S3 is fine tuning of a deep network, and a weight model trained by a residual error network on an ImageNet data set is migrated to an SAR image recognition operation.
Preferably, the attention mechanism in S2 is a CBAM hybrid attention module, and the operation method of the hybrid attention module is a combination of a channel attention mechanism and a spatial attention mechanism.
Preferably, the channel attention mechanism is as follows: the channel sub-module utilizes the maximum convergent output and the average convergent output of the shared network, and the channel attention mechanism calculation formula is as follows:
wherein A is c (F) Representing the channel attention mechanism function, F representing the input of the channel attention mechanism, and calculating in a matrix form, sigma representing the sigmoid function, W 0 Representing the multi-layer perceptron built-in hidden layer weight computation,W 1 representing the multi-layer perceptron built-in output layer weight computation,indicating that input F is averaged pooled within the channel attention mechanism;indicating the maximum pooling of input F within the channel attention mechanism.
Preferably, said W 0 And W 1 For two input sharing, the ReLU activation function is followed by W 0 ;W 0 ∈R C/r×C ,W 1 ∈R C×C/r ,W 0 ∈R C/r×C Represents W 0 The method comprises the steps that weight values between C neurons of an input layer and C/R neurons of a hidden layer are input, C represents the channel number of SAR image data, R represents a hyper-parameter of a multilayer perceptron, and R represents picture data; w 1 ∈R C×C/r Represents W 1 Is the weight value between the C/r neurons of the hidden layer to the C neurons of the output layer.
Preferably, the spatial attention mechanism is as follows: the spatial submodule takes two similar outputs collected along the channel axis and forwards them to the convolutional layer, the formula for the spatial attention mechanism being:
wherein A is s (F) Representing a function of a spatial attention mechanism, and F representing an input of the spatial attention mechanism; σ denotes sigmoid function, f 7×7 Represents a convolution operation with a convolution kernel size of 7 x 7,representing the average pooling of image data along the channel dimension in a spatial attention mechanism;indicating doing image data in a spatial attention mechanismMaximal pooling is done along the channel dimension.
Preferably, the first and second liquid crystal materials are,representing the feature map dimensions obtained by stitching together the average pooled data and the maximum pooled tree along the channel dimensionsRepresenting that the feature map data dimension becomes 1 × H × W after the input feature map data is subjected to the average pooling process of the spatial attention mechanism;indicating that the feature map data dimension becomes 1 × H × W after maximum pooling processing of the input feature map data in the spatial attention mechanism.
Preferably, the target recognition result obtained in S5 is used as the target recognition vector M; the target identification vector M is obtained by multiplying the weight A obtained by attention learning by the input H:
M=AH (3)
in the above formula, H is the attention layer input, and A is the attention coding output;
a is calculated as follows:
wherein,is a matrix multiplication element; a. the c Attention is drawn to one-dimensional channels, A s Attention is drawn to two-dimensional space.
The invention has the beneficial effects that:
compared with the traditional SAR image target identification method, the SAR ATR method based on the deep migration learning has stronger expression capability, can quickly extract image characteristics by utilizing a fine-tuned residual error network, realizes end-to-end learning, combines an attention mechanism, focuses information containing target characteristics, improves the capability of characteristic screening, further improves the target identification capability of the SAR image, and effectively improves the identification performance.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention and not to limit the invention.
FIG. 1 is a schematic view of a depth migration model according to the present invention;
FIG. 2 is a structural diagram of a CBAM according to the present invention;
FIG. 3 is a schematic diagram of a channel attention model according to the present invention;
FIG. 4 is a schematic diagram of a spatial attention model according to the present invention;
FIG. 5 is a schematic diagram of the SAR image recognition algorithm structure combining the attention mechanism and the migration model according to the present invention;
fig. 6 is a graph of the recognition rate versus the line of the SAR image under the influence of different degrees of noise.
Detailed Description
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention.
Referring to fig. 1-6, a preferred embodiment of the present invention, a method for target recognition of SAR images by combining transfer learning and attention mechanism, the method comprises the following steps:
s1, extracting a convolution feature map from the input SAR image through 7 × 7 convolution kernels;
s2, inputting the obtained convolution characteristic diagram into an attention mechanism for focusing to obtain a primary focusing characteristic diagram, wherein the attention mechanism is used for filtering out interest points from a large amount of digital image information, selectively screening out important information and focusing to further improve the performance;
s3, inputting the initial focusing feature map into the fine-tuned residual error network to obtain a weight feature map, wherein the residual error network is formed by utilizing a depth convolution network to complete an end-to-end target identification task and then combining a depth migration learning method, a weight model trained on a known data set by the residual error network is migrated into the SAR image identification work, the SAR image identification model training time is accelerated, and the identification efficiency is improved;
s4, obtaining a weight characteristic diagram containing the weight, inputting the weight characteristic diagram into an attention mechanism again for focusing;
and S5, inputting the focused target identification feature map containing the weight into a subsequent convolutional layer for classification to obtain a target identification result.
Specifically, the adopted migration model is a model obtained by training an ImageNet data set, the attention mechanism method is adopted for channel attention and space attention, and training is carried out by combining the attention mechanism under the condition that an original deep convolution network structure is not changed; the functional block diagram of the ATR algorithm of SAR with a combination of deep migration learning and attention mechanism is shown in fig. 5.
As a preferred embodiment of the present invention, it may also have the following additional technical features:
in this embodiment, the deep migration learning method in S3 is fine tuning of a deep network, and a weight model trained on an ImageNet data set by a residual error network is migrated to an SAR image recognition work.
In this embodiment, the attention mechanism in S2 is a hybrid attention module, and the operation method of the hybrid attention module is a combination of a channel attention mechanism and a spatial attention mechanism.
In this embodiment, the channel attention mechanism is: the channel submodule utilizes the maximum convergent output and the average convergent output of the shared network, and the channel attention mechanism calculation formula is as follows:
wherein, A c (F) Representing the channel attention mechanism function, F representing the input to the channel attention mechanism,calculated in matrix form, sigma denotes sigmoid function, W 0 Representing the computation of weights of hidden layers within a multi-layer perceptron, W 1 Representing the multi-layer perceptron built-in output layer weight calculation,indicating that input F is averaged pooled within the channel attention mechanism;indicating the maximum pooling of input F within the channel attention mechanism.
In this embodiment, W is as defined above 0 And W 1 For two input sharing, the ReLU activation function is followed by W 0 ;W 0 ∈R C/r×C ,W 1 ∈R C ×C/r ,W 0 ∈R C/r×C Represents W 0 The method comprises the steps that weight values between C neurons of an input layer and C/R neurons of a hidden layer are input, C represents the channel number of SAR image data, R represents a hyper-parameter of a multilayer perceptron, and R represents picture data; w is a group of 1 ∈R C×C/r Represents W 1 Is the weight value between the C/r neurons of the hidden layer to the C neurons of the output layer.
In this embodiment, the spatial attention mechanism is: the spatial submodule takes two similar outputs collected along the channel axis and forwards them to the convolutional layer, the formula for the spatial attention mechanism being:
wherein A is s (F) Representing a function of a spatial attention mechanism, and F representing an input of the spatial attention mechanism; σ denotes sigmoid function, f 7×7 Represents a convolution operation with a convolution kernel size of 7 x 7,representing the average pooling of image data along the channel dimension in a spatial attention mechanism;indicating that the image data is maximally pooled along the channel dimension in the spatial attention mechanism.
In the present embodiment, the first and second electrodes are,representing the feature map dimensions obtained by stitching together the average pooled data and the maximum pooled tree along the channel dimensionsRepresenting that the feature map data dimension becomes 1 × H × W after the input feature map data is subjected to the average pooling process of the spatial attention mechanism;indicating that the feature map data dimensions become 1 × H × W after maximum pooling of the input feature map data in the spatial attention mechanism.
In this embodiment, the target recognition result obtained in S5 is used as the target recognition vector M; the target identification vector M is obtained by multiplying the weight A obtained by attention learning by the input H:
M=AH (3)
in the above formula, H is the attention layer input, and A is the attention coding output;
a is calculated as follows:
wherein,is a matrix multiplication element; a. the c Attention is drawn to one-dimensional channels, A s Attention is drawn to two-dimensional space.
DETAILED DESCRIPTION OF EMBODIMENT (S) OF INVENTION
Fig. 1 shows a deep migration model, and fine-tuning (fine-tuning) of a deep network is the most common deep network migration method, and network fine-tuning adjusts a network according to a task of the network by using a network that is trained by others. The method has the advantages that the network does not need to be trained from the beginning aiming at a new task, so that the time cost is saved; the pre-trained model is usually performed on a large data set, so that training data is expanded, the model has higher robustness, and the generalization capability of the model is effectively improved.
The specific process is embodied in that a model structure and partial parameters of an ATR deep learning model of the trained A target SAR are directly transferred to an ATR deep learning model of the B target SAR, and a small amount of training data of a B target SAR data set is used for fine tuning; the method is equivalent to performing output layer conversion on a pre-trained model, namely converting an original output layer into a new output layer with randomly initialized parameters, and then training the output layer by using a smaller data set on the basis of the model.
FIG. 2 shows a CBAM structure. After an SAR image is input, the CBAM considers two problems of "what" the target is and "where" the target is, and the specific operation method is as follows: the channel sub-module utilizes the maximum convergent output and the average convergent output of the shared network; the spatial submodule takes two similar outputs, which converge along the channel axis, and forwards them to the convolutional layer.
In order to solve the problem that the nonlinear classification cannot be well solved by using a Softmax single-layer full-link mode in the conventional CNN, we adopt a Attention mechanism as a hybrid Attention Module (CBAM), and compared with the original Attention method, the CBAM includes a channel Attention mechanism and a spatial Attention mechanism, and the Attention mechanism can flexibly capture the link between global information and local information. The purpose of the attention mechanism is to enable the model to obtain a target region needing important attention, put more weight into the target region, highlight remarkable useful features and suppress and ignore irrelevant features.
FIG. 3 shows a channel attention mechanism, which is formulated as follows:
wherein A is c (F) Representing the channel attention mechanism function, F representing the input of the channel attention mechanism, and calculating in a matrix form, sigma representing the sigmoid function, W 0 Representing the computation of weights of hidden layers within a multi-layer perceptron, W 1 Representing the multi-layer perceptron built-in output layer weight computation,indicating that input F is averaged pooled within the channel attention mechanism;indicating the maximum pooling of input F within the channel attention mechanism.
For a given SAR image profile input F ∈ R H×W×C Where F denotes the input of SAR image data, R H×W×C Three dimensional data representing an input feature map are respectively H image height, W image width and C image channel number; simultaneously, through Global Average Pooling (GAP) and Global Maximum Pooling (GMP), operators with different space semantic descriptions are respectively obtained, the operators and the operators are fused through a Multilayer Perceptron (MPL), then, two feature vectors are fused in an addition mode, and finally, through activating a sigmoid function, a channel attention vector A is obtained c ∈R C×1×1 。A c ∈R C×1×1 The three dimensional data after the image is processed by the channel attention mechanism are respectively H image height 1, W image width 1 and C image channel number C.
Shown in fig. 4 is a spatial attention mechanism, where the spatial submodule takes two similar outputs collected along the channel axis and forwards them to the convolutional layer. The formula is as follows:
wherein A is s (F) Representing a function of a spatial attention mechanism, and F representing an input of the spatial attention mechanism; sigma denotes sigmoid function, f 7×7 Represents a convolution operation with a convolution kernel size of 7 x 7,representing the average pooling of image data along the channel dimension in a spatial attention mechanism;indicating that the image data is maximally pooled along the channel dimension in the spatial attention mechanism.
For a given input: f is belonged to R H×W×C Simultaneously carrying out Global Average Pooling (GAP) and Global Maximum Pooling (GMP) operations along channel dimensions to respectively obtain two different channel feature description operators, splicing the two channel feature description operators to obtain a two-dimensional feature map, carrying out 7-by-7 convolution operation on the two channel feature description operators and a sigmoid activation function to finally obtain a spatial attention vector A s ∈R 1×H×W 。A s ∈R 1×H×W The three dimensional data after the image is processed by the channel attention mechanism are respectively H image height, W image width and C image channel number of 1.
Combining the fig. 3 and fig. 4, inputting a feature map extracted from the SAR image through a convolutional layer into a channel attention mechanism shown in fig. 3, correcting an original feature map through the channel attention mechanism to obtain an intermediate feature map, inputting the intermediate feature map into a space attention mechanism module for correction, and finally obtaining a feature map subjected to attention focusing as shown in fig. 2, wherein the CBAM mixed attention mechanism is a serial mixing mechanism which is obtained by firstly passing the feature map through the channel attention mechanism and then passing the space attention mechanism.
Fig. 5 shows a network framework of SAR images combining deep migration learning and attention mechanism, where a target recognition vector M represents a final target recognition result, and M is obtained by multiplying a weight a obtained by attention learning by an input H:
M=AH (3)
in the above formula, H is the attention layer input, and a is the encoded output of attention;
a is calculated as follows:
wherein,is a matrix multiplication element; a. the c Attention is drawn to one-dimensional channels, A s Attention is drawn to two-dimensional space.
And finally, an attention mechanism and a transfer learning method are effectively combined, so that the integral algorithm of the SAR ATR is realized.
Fig. 6 shows the SAR image recognition accuracy under 0% to 15% random noise, and the SAR image recognition rate of the invention for the MSTART data set containing 15% random noise is much higher than that of the MSTART data set containing 15% random noise by other methods.
Compared with the traditional SAR image target identification method, the SAR ATR method based on the deep migration learning has stronger expression capability, can quickly extract image characteristics by utilizing a fine-tuned residual error network, realizes end-to-end learning, combines an attention mechanism, focuses information containing target characteristics, improves the capability of characteristic screening, further improves the target identification capability of the SAR image, and effectively improves the identification performance.
The above additional technical features can be freely combined and used in superposition by those skilled in the art without conflict.
The above description is only a preferred embodiment of the present invention, and the technical solutions that achieve the objects of the present invention by basically the same means are all within the protection scope of the present invention.
Claims (8)
1. A SAR image target identification method combining transfer learning and attention mechanism is characterized in that: the method comprises the following steps:
s1, extracting a convolution feature map from the input SAR image through 7 × 7 convolution kernels;
s2, inputting the obtained convolution characteristic diagram into an attention mechanism for focusing to obtain a primary focusing characteristic diagram, wherein the attention mechanism is used for filtering interest points from a large amount of digital image information, selectively filtering important information and focusing;
s3, inputting the initial focusing feature map into the fine-tuned residual error network to obtain a weight feature map, wherein the residual error network is formed by utilizing a depth convolution network to complete an end-to-end target identification task and then combining a depth migration learning method, a weight model trained by the residual error network on a known data set is migrated into the SAR image identification work, and the SAR image identification model training time is accelerated;
s4, obtaining a weight characteristic diagram containing the weight, inputting the weight characteristic diagram into an attention mechanism again for focusing;
and S5, inputting the focused target recognition feature map containing the weight into a subsequent convolutional layer for classification to obtain a target recognition result.
2. The SAR image target recognition method combining transfer learning and attention mechanism as claimed in claim 1, wherein: the deep transfer learning method in the step S3 is fine tuning of a deep network, and a weight model trained on an ImageNet data set by a residual error network is transferred to SAR image recognition work.
3. The SAR image target recognition method combining transfer learning and attention mechanism as claimed in claim 1, wherein: the attention mechanism in S2 is a CBAM hybrid attention module, and the method of operation of the hybrid attention module is a combination of a channel attention mechanism and a spatial attention mechanism.
4. The SAR image target recognition method combining transfer learning and attention mechanism as claimed in claim 3, wherein: the channel attention mechanism is as follows: the channel submodule utilizes the maximum convergent output and the average convergent output of the shared network, and the channel attention mechanism calculation formula is as follows:
wherein A is c (F) Representing the channel attention mechanism function, F representing the input of the channel attention mechanism, and calculating in a matrix form, sigma representing the sigmoid function, W 0 Representing the computation of weights of hidden layers within a multi-layer perceptron, W 1 Representing the multi-layer perceptron built-in output layer weight computation,indicating that input F is averaged pooled within the channel attention mechanism;indicating the maximum pooling of input F within the channel attention mechanism.
5. The SAR image target recognition method combining transfer learning and attention mechanism as claimed in claim 4, wherein: the W is 0 And W 1 For two input sharing, the ReLU activation function is followed by W 0 ;W 0 ∈R C/r×C ,W 1 ∈R C×C/r ,W 0 ∈R C/r×C Represents W 0 The method comprises the steps that weight values between C neurons of an input layer and C/R neurons of a hidden layer are input, C represents the channel number of SAR image data, R represents a hyper-parameter of a multilayer perceptron, and R represents picture data; w 1 ∈R C×C/r Represents W 1 Is the weight value between the C/r neurons of the hidden layer to the C neurons of the output layer.
6. The SAR image target recognition method combining transfer learning and attention mechanism as claimed in claim 3, wherein: the spatial attention mechanism is as follows: the spatial submodule takes two similar outputs collected along the channel axis and forwards them to the convolutional layer, the formula for the spatial attention mechanism being:
wherein A is s (F) Representing a function of a spatial attention mechanism, and F representing an input of the spatial attention mechanism; σ denotes sigmoid function, f 7×7 Represents a convolution operation with a convolution kernel size of 7 x 7,representing the average pooling of image data along the channel dimension in a spatial attention mechanism;indicating that the image data is maximally pooled along the channel dimension in the spatial attention mechanism.
7. The SAR image target recognition method combining transfer learning and attention mechanism as claimed in claim 6, wherein: representing the feature map dimensions obtained by stitching together the average pooled data and the maximum pooled tree along the channel dimensionsRepresenting that the feature map data dimension becomes 1 × H × W after the input feature map data is subjected to the average pooling process of the spatial attention mechanism;indicating that the feature map data dimension becomes 1 × H × W after maximum pooling processing of the input feature map data in the spatial attention mechanism.
8. The SAR image target recognition method combining transfer learning and attention mechanism as claimed in claim 1, wherein: using the target recognition vector M to obtain a target recognition result in the S5; the target identification vector M is obtained by multiplying the weight A obtained by attention learning by the input H:
M=AH (3)
in the above formula, H is the attention layer input, and A is the attention coding output;
a is calculated as follows:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210579119.3A CN114863283A (en) | 2022-05-25 | 2022-05-25 | SAR image target identification method combining transfer learning and attention mechanism |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210579119.3A CN114863283A (en) | 2022-05-25 | 2022-05-25 | SAR image target identification method combining transfer learning and attention mechanism |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114863283A true CN114863283A (en) | 2022-08-05 |
Family
ID=82639805
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210579119.3A Pending CN114863283A (en) | 2022-05-25 | 2022-05-25 | SAR image target identification method combining transfer learning and attention mechanism |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114863283A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116228797A (en) * | 2023-05-09 | 2023-06-06 | 中国石油大学(华东) | Shale scanning electron microscope image segmentation method based on attention and U-Net |
CN116559949A (en) * | 2023-05-19 | 2023-08-08 | 北京宸宇金源科技有限公司 | Carbonate reservoir prediction method, system and equipment based on deep learning |
-
2022
- 2022-05-25 CN CN202210579119.3A patent/CN114863283A/en active Pending
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116228797A (en) * | 2023-05-09 | 2023-06-06 | 中国石油大学(华东) | Shale scanning electron microscope image segmentation method based on attention and U-Net |
CN116228797B (en) * | 2023-05-09 | 2023-08-15 | 中国石油大学(华东) | Shale scanning electron microscope image segmentation method based on attention and U-Net |
CN116559949A (en) * | 2023-05-19 | 2023-08-08 | 北京宸宇金源科技有限公司 | Carbonate reservoir prediction method, system and equipment based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109753903B (en) | Unmanned aerial vehicle detection method based on deep learning | |
CN111325165B (en) | Urban remote sensing image scene classification method considering spatial relationship information | |
CN113065558A (en) | Lightweight small target detection method combined with attention mechanism | |
CN114863283A (en) | SAR image target identification method combining transfer learning and attention mechanism | |
CN109801215B (en) | Infrared super-resolution imaging method based on countermeasure generation network | |
CN110717856A (en) | Super-resolution reconstruction algorithm for medical imaging | |
Mostofa et al. | Joint-SRVDNet: Joint super resolution and vehicle detection network | |
CN113361485B (en) | Hyperspectral image classification method based on spectrum space attention fusion and deformable convolution residual error network | |
CN111612711A (en) | Improved picture deblurring method based on generation countermeasure network | |
CN109034184B (en) | Grading ring detection and identification method based on deep learning | |
Hu et al. | A two-stage unsupervised approach for low light image enhancement | |
CN112241939B (en) | Multi-scale and non-local-based light rain removal method | |
CN113920043A (en) | Double-current remote sensing image fusion method based on residual channel attention mechanism | |
CN116758130A (en) | Monocular depth prediction method based on multipath feature extraction and multi-scale feature fusion | |
CN115331104A (en) | Crop planting information extraction method based on convolutional neural network | |
CN113240697B (en) | Lettuce multispectral image foreground segmentation method | |
CN113763417B (en) | Target tracking method based on twin network and residual error structure | |
CN110751271B (en) | Image traceability feature characterization method based on deep neural network | |
CN113034361A (en) | Remote sensing image super-resolution reconstruction method based on improved ESRGAN | |
CN114937202A (en) | Double-current Swin transform remote sensing scene classification method | |
CN113538243A (en) | Super-resolution image reconstruction method based on multi-parallax attention module combination | |
Yanshan et al. | OGSRN: Optical-guided super-resolution network for SAR image | |
CN114463176B (en) | Image super-resolution reconstruction method based on improved ESRGAN | |
CN116843975A (en) | Hyperspectral image classification method combined with spatial pyramid attention mechanism | |
CN113627487B (en) | Super-resolution reconstruction method based on deep attention mechanism |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |