SAR (Synthetic Aperture Radar) aircraft target identification method in combination with lightweight attention and deformable convolution

Abstract

The invention relates to the technical field of image processing, in particular to an SAR aircraft target recognition method combining lightweight attention and deformable convolution. According to the SAR aircraft target identification method combining the lightweight attention and the deformable convolution, the application of the lightweight attention mechanism and the deformable convolution can effectively solve the problem of high-precision identification of low-quality and weak dark targets. According to the invention, the attention mechanism enables the model to accurately recognize the aircraft target even in a complex background and a diversified environment through focusing on an important region and a feature channel of an image, and compared with a traditional deep learning method, the mechanism reduces the demand of computing resources while improving the target detection accuracy; and secondly, the deformable convolution provides greater flexibility for the network, so that the network can adapt to the shape change and irregularity of the target in the SAR image, and the convolution technology can capture the characteristics of the aircraft target more accurately by dynamically adjusting the convolution kernel.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a SAR aircraft target recognition method combining lightweight attention and deformable convolution. Background Art

[0002] As an advanced microwave imaging technology, synthetic aperture radar (SAR) plays a vital role in the field of aircraft identification. Compared with traditional optical remote sensing, SAR has all-weather and all-day imaging capabilities and can obtain high-quality images under various climate and lighting conditions, which makes SAR have unique advantages in aircraft identification and monitoring. Especially in cloud-covered or night conditions, SAR can effectively overcome the limitations of optical remote sensing technology and provide more reliable and continuous data support for aircraft monitoring. In addition, SAR images can reflect the unique scattering characteristics of the target, which is particularly important for identifying and distinguishing different types of aircraft. Accurate SAR aircraft identification is crucial to aviation safety, air transportation and other fields. It can provide key information in a timely manner and help track aircraft movements and analyze behavior. With the continuous advancement of technology, the application scope and efficiency of SAR aircraft identification technology have been continuously improved, providing strong technical support for ensuring airport environmental safety monitoring. The development of this technology has also promoted technological progress in related fields such as remote sensing image processing and target recognition, and has had a profound impact on the overall development of remote sensing science and aviation technology.

[0003] SAR aircraft recognition is a complex process involving a variety of advanced image processing and pattern recognition technologies. At present, the commonly used SAR aircraft recognition technologies mainly include feature-based recognition technology, template matching-based recognition technology, and machine learning-based recognition technology. Feature-based recognition technology identifies targets by analyzing features such as the shape, size, and texture of aircraft in SAR images. For example, Zhang Wei et al. proposed a method combining shape and texture features in 2018 to effectively identify aircraft from SAR images. Template matching-based recognition technology determines whether the target is an aircraft by comparing a predefined aircraft model with the target in the SAR image. For example, Chen Li et al. proposed a template matching method based on improved correlation coefficients in 2019, which successfully improved the recognition rate of aircraft targets in SAR images. Finally, machine learning-based recognition technology uses algorithms such as support vector machines (SVM), neural networks, or deep learning models to automatically identify aircraft in SAR images. For example, Li Na et al. applied deep learning technology in 2016 and combined convolutional neural networks (CNN) to perform feature learning on SAR images, which significantly improved the accuracy of aircraft recognition. These methods each have their own advantages and disadvantages, but together they have promoted the development of SAR aircraft identification technology and provided important technical support for fields such as aviation monitoring and military reconnaissance.

[0004] Although these methods have improved the accuracy and efficiency of target detection to a certain extent, they still face some challenges and limitations. First, SAR images usually have high levels of noise and low contrast due to their unique imaging mechanism, which makes it more difficult to accurately extract target features. In addition, due to the characteristics of radar waves, targets in SAR images may appear geometrically distorted, such as stretching or compression, which further increases the complexity of recognition.

[0005] Although deep learning methods perform well in automatic feature extraction, they usually require a large amount of labeled data to train the model, which is often difficult to obtain in the case of SAR images. In addition, these models are often computationally intensive and require high computing resources, which limits their use in resource-constrained real-world application environments. Summary of the invention

[0006] In view of this, the present invention provides a SAR aircraft target recognition method combining lightweight attention and deformable convolution, which can improve the accuracy of target detection while reducing the demand for computing resources and more accurately capture the characteristics of aircraft targets.

[0007] To achieve the above objectives, the present invention provides a SAR aircraft target recognition method combining lightweight attention and deformable convolution. The lightweight attention mechanism is used to focus on important areas and feature channels of the image so that the model can recognize aircraft targets. Deformable convolution is used to make the network adapt to the shape changes and irregularities of targets in SAR images, wherein the characteristics of aircraft targets are captured by dynamically adjusting the convolution kernel.

[0008] The specific implementation method includes the following steps:

[0009] S1: Construct training dataset, validation dataset and test dataset;

[0010] S2: Divide the training data into blocks and enhance the data;

[0011] S3: Construct a SAR airport aircraft detection model based on lightweight attention and deformable convolution. The model is a 10-layer deep learning network model. The first layer is the input layer, the second layer is the convolution layer 1, which is used to extract primary features; the third layer is the batch normalization layer, which is added after the convolution layer to accelerate the training process; the fourth layer is the activation layer; the fifth layer is the convolution layer 2, which is used to extract features of image details; the sixth layer is the lightweight attention layer, which integrates channel and spatial attention mechanisms to emphasize key features in the image while maintaining computational efficiency; the seventh layer is the deformable convolution layer; the eighth layer is the pooling layer, which uses maximum pooling to reduce feature dimensions; the ninth layer is the fully connected layer, which is used to flatten the output of the convolution layer and classify it through the fully connected layer; the last layer is the output layer, which uses the Softmax activation function to output the classification and location information of the target;

[0012] S4: Training a SAR airport aircraft recognition model based on lightweight attention and deformable convolution;

[0013] S5: Testing based on the trained SAR airport aircraft recognition model based on lightweight attention and deformable convolution: loading the optimal model parameters, inputting the data set to be recognized into the input layer of the trained SAR airport aircraft target recognition network, and the network output layer's judgment result on the target type. The output information of the present invention includes the target type label and confidence level;

[0014] S6: Save or output the recognition results as required.

[0015] Wherein, the activation layer uses the ReLU nonlinear activation function to perform feature activation, so as to introduce nonlinearity;

[0016] The lightweight attention layer construction formula is as follows:

[0017] F'=(W ch F)W sp

[0018] Among them, W ch is the channel attention, and the calculation formula is:

[0019] W ch =σ(FC(ReLU(FC(GlobalAvgPool(F)))))

[0020] W sp is the spatial attention, and the calculation formula is:

[0021] W sp =σ(Conv(Concat(MaxPool(F),AvgPool(F))))

[0022] Among them, GlobalAvgPool represents global average pooling, FC represents a fully connected layer, ReLU represents a linear rectification function, σ represents a Sigmoid function, MaxPool and AvgPool represent maximum pooling and average pooling respectively, Conv represents a convolution operation, Concat represents a concatenation operation in the channel dimension, and ⊙ represents a Hadamard product;

[0023] Among them, channel attention W ch The feature map after global average pooling is weighted by the fully connected layer to encode the importance of different channels; spatial attention W sp The pooled feature map is weighted through the convolution layer to emphasize the important spatial areas in the image; finally, the two attentions are fused together by element-wise multiplication to generate a feature map F' that is doubly adjusted in space and channel.

[0024] Among them, the output layer uses the Softmax activation function to output the classification and location information of the target. The probability value p after the Softmax function conversion is i The calculation formula is as follows:

[0025]

[0026] Where N is the number of neurons in the output layer, z i is the original output value of the ith neuron.

[0027] The specific method of constructing the dataset is as follows: collect and download N scenes of Terra SAR and GF3 SAR image data, with a single scene size of no less than 1k×1k pixels, and divide the data into training dataset, validation dataset and test dataset in a ratio of 8:1:1.

[0028] The specific method of dividing the training data into blocks and enhancing the data is as follows: reading the SAR aircraft data in the training set and dividing it into blocks; the blocking adopts an overlapping blocking method, that is, using L×L squares to step horizontally and vertically in the SAR remote sensing image, the step amount is L / 2, and multiple blocks are obtained, called Tiles, and geometric distortion simulation and multi-view imaging simulation are used for data enhancement processing.

[0029] Among them, the specific method of training the SAR airport aircraft recognition model based on lightweight attention and deformable convolution is: input the training data set and label set into the built aircraft recognition model, train the lightweight attention module and the deformable convolution module respectively, obtain the network model architecture of the Mth generation of training, and save the optimal model parameters for training.

[0030] Among them, the specific method of testing the trained SAR airport aircraft recognition model based on lightweight attention and deformable convolution is: loading the optimal model parameters, inputting the data set to be identified into the input layer of the trained SAR airport aircraft target recognition network, and the network output layer's judgment result on the target type, and the output information includes the target type label and confidence.

[0031] Beneficial effects:

[0032] 1. The SAR aircraft target recognition method combining lightweight attention and deformable convolution of the present invention has proposed significant innovations and advantages on the basis of the existing technology, taking into account the balance between processing performance and processing efficiency, and can be quickly migrated and deployed on embedded hardware platforms; the application of lightweight attention mechanism and deformable convolution can effectively solve the problem of high-precision recognition of low-quality, weak and dark targets.

[0033] 2. The present invention is a remote sensing land image target recognition method, which can quickly and accurately identify targets under the large field of view of remote sensing. First, the lightweight attention mechanism effectively improves the model's ability to recognize key features while maintaining the computational efficiency of the network. This attention mechanism focuses on important areas and feature channels of the image, so that the model can accurately identify aircraft targets even in complex backgrounds and diverse environments. Compared with traditional deep learning methods, this mechanism improves target detection accuracy while reducing the demand for computing resources; secondly, deformable convolution provides greater flexibility for the network, enabling it to adapt to the shape changes and irregularities of targets in SAR images. This convolution technology can more accurately capture the characteristics of aircraft targets by dynamically adjusting the convolution kernel, especially in the case of changes in perspective and geometric distortion, which is difficult to achieve with traditional convolutional networks because they are usually more sensitive to changes in the shape and size of the target.

[0034] 3. The method of the present invention can more effectively extract useful information from complex SAR data, improve the detection accuracy of aircraft targets, and reduce the dependence on large amounts of labeled data and high computing resources.

[0035] 4. The method of the present invention is technically innovative and is mainly used in SAR aircraft recognition processing, but is not limited to this. The method is versatile in solving similar weak, low-significance targets and scenes, and has significant advantages in practical applications, especially in scenes that require real-time or near real-time processing.

[0036] 5. The method of the present invention also shows advantages in processing the noise and low contrast problems of SAR images. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the process of the present invention.

[0038] Figure 2 This is the network structure diagram of SAR airport aircraft recognition based on lightweight attention and deformable convolution in the present invention.

[0039] Figure 3 Schematic diagram of Boeing 737 recognition results at different angles in an embodiment of the present invention.

[0040] Figure 4 Schematic diagram of A330 recognition results at different angles in an embodiment of the present invention.

[0041] Figure 5 Schematic diagram of the recognition results of A220 at different angles in an embodiment of the present invention.

[0042] Figure 6 Schematic diagram of the whole scene target detection and recognition result 1 in an embodiment of the present invention.

[0043] Figure 7 Schematic diagram of the whole scene target detection and recognition result 2 in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0045] The present invention provides a SAR aircraft target recognition method combining lightweight attention and deformable convolution, the process is as follows Figure 1 As shown, the following steps are included:

[0046] S1: Constructing data sets: Collect and download Terra SAR and GF3 SAR image data with N scenes in total, with the size of each scene not less than 1k×1k pixels, and divide the data into training data set, validation data set and test data set in a ratio of 8:1:1;

[0047] S2: Blocking and data enhancement of training data: Read the SAR aircraft data in the training set and block them; the block is divided into overlapping blocks, that is, L×L squares are used to step horizontally and vertically in the SAR remote sensing image, and the step amount is L / 2 to obtain multiple blocks, called Tiles, and use geometric distortion simulation, multi-view imaging simulation and other data enhancement processing;

[0048] S3: Construct a SAR airport aircraft detection model based on lightweight attention and deformable convolution. The model is a 10-layer deep learning network model:

[0049] (3a) The first layer is the input layer, which mainly receives SAR image data and performs fixed-size downsampling processing on the input data. The present invention uses mean downsampling to sample the input image to a size of 1024 pixels × 1024 pixels;

[0050] (3b) The second layer is the convolution layer 1, which uses a larger convolution kernel (a 5×5 convolution kernel is used in this patent) to extract primary features such as edges and textures;

[0051] (3c) The third layer is the batch normalization layer, which is added after the convolutional layer to speed up the training process and improve the stability of the model;

[0052] (3d) The fourth layer is an activation layer. The present invention uses a ReLU nonlinear activation function for feature activation, which is used to introduce nonlinearity so that the network can capture complex features.

[0053] (3e) The fifth layer is the convolution layer 2, which uses a smaller convolution kernel (the present invention uses a 3x3 convolution kernel) to further extract more detailed features of the image;

[0054] (3f) The sixth layer is a lightweight attention layer, which integrates channel and spatial attention mechanisms to emphasize key features in the image while maintaining computational efficiency. The lightweight attention layer construction formula of the present invention is as follows:

[0055] F'=(W ch F)W sp

[0056] In the above formula, W ch is the channel attention, and the calculation formula is:

[0057] W ch =σ(FC(ReLU(FC(GlobalAvgPool(F)))))

[0058] W sp is the spatial attention, and the calculation formula is:

[0059] W sp =σ(Conv(Concat(MaxPool(F),AvgPool(F))))

[0060] Among them, GlobalAvgPool represents global average pooling, FC represents the fully connected layer, ReLU represents the linear rectification function (activation function), σ represents the Sigmoid function, MaxPool and AvgPool represent maximum pooling and average pooling respectively, Conv represents the convolution operation, Concat represents the concatenation operation in the channel dimension, and ⊙ represents the Hadamard product (element-wise product).

[0061] In this formula, the channel attention W ch The feature map after global average pooling is weighted by a fully connected layer to encode the importance of different channels. Spatial Attention W spThe pooled feature map is weighted through the convolution layer to emphasize the important spatial regions in the image. Finally, the two attentions are fused together by element-wise multiplication to generate a feature map F' that is spatially and channel-wise adjusted.

[0062] (3g) The seventh layer is a deformable convolution layer. This layer uses deformable convolution to increase the adaptability of the network to the shape changes of SAR aircraft targets. The specific operation is as follows: on the basis of traditional convolution, a learnable offset is added. These offsets are generated by the additional convolution layer and are used to adjust the sampling points of each convolution kernel position;

[0063] (3h) The eighth layer is a pooling layer. The present invention uses maximum pooling to reduce feature dimensions and computational complexity while retaining important features.

[0064] (3i) The ninth layer is a fully connected layer: the output of the convolutional layer is flattened and classified through the fully connected layer;

[0065] (3j) The output layer is defined as follows: The last layer uses the Softmax activation function to output the classification and location information of the target. The probability value p after the Softmax function conversion i The calculation formula is as follows:

[0066]

[0067] Where N is the number of neurons in the output layer, z i is the original output value of the i-th neuron;

[0068] S4: Training the SAR airport aircraft recognition model based on lightweight attention and deformable convolution: Input the training data set and label set into the built aircraft recognition model, train the lightweight attention module and deformable convolution module respectively, obtain the network model architecture of the Mth generation of training, and save the optimal model parameters of the training;

[0069] S5: Testing based on the trained SAR airport aircraft recognition model based on lightweight attention and deformable convolution: loading the optimal model parameters, inputting the data set to be recognized into the input layer of the trained SAR airport aircraft target recognition network, and the network output layer's judgment result on the target type. The output information of the present invention includes the target type label and confidence level;

[0070] S6: Save or output the recognition results as required.

[0071] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A SAR aircraft target recognition method combining lightweight attention and deformable convolution, characterized in that: The lightweight attention mechanism is used to focus on important areas and feature channels of the image, so that the model can recognize aircraft targets. The deformable convolution is used to make the network adapt to the shape changes and irregularities of targets in SAR images, and the characteristics of aircraft targets are captured by dynamically adjusting the convolution kernel.

2. The method according to claim 1, characterized in that The specific implementation method includes the following steps: S1: Construct training dataset, validation dataset and test dataset; S2: Divide the training data into blocks and enhance the data; S3: Construct a SAR airport aircraft detection model based on lightweight attention and deformable convolution. The model is a 10-layer deep learning network model. The first layer is the input layer, the second layer is the convolution layer 1, which is used to extract primary features; the third layer is the batch normalization layer, which is added after the convolution layer to accelerate the training process; the fourth layer is the activation layer; the fifth layer is the convolution layer 2, which is used to extract features of image details; the sixth layer is the lightweight attention layer, which integrates channel and spatial attention mechanisms to emphasize key features in the image while maintaining computational efficiency; the seventh layer is the deformable convolution layer; the eighth layer is the pooling layer, which uses maximum pooling to reduce feature dimensions; the ninth layer is the fully connected layer, which is used to flatten the output of the convolution layer and classify it through the fully connected layer; the last layer is the output layer, which uses the Softmax activation function to output the classification and location information of the target; S4: Training a SAR airport aircraft recognition model based on lightweight attention and deformable convolution; S5: Testing based on the trained SAR airport aircraft recognition model based on lightweight attention and deformable convolution: loading the optimal model parameters, inputting the data set to be recognized into the input layer of the trained SAR airport aircraft target recognition network, and the network output layer's judgment result on the target type. The output information of the present invention includes the target type label and confidence level; S6: Save or output the recognition results as required.

3. The method according to claim 2, characterized in that ,The activation layer uses the ReLU nonlinear activation function to perform feature activation, which is used to introduce nonlinearity; The lightweight attention layer construction formula is as follows: F'=(W ch F)W sp Among them, W ch is the channel attention, and the calculation formula is: W ch =σ(FC(ReLU(FC(GlobalAvgPool(F))))) W sp is the spatial attention, and the calculation formula is: W sp =σ(Conv(Concat(MaxPool(F),AvgPool(F)))) Among them, GlobalAvgPool represents global average pooling, FC represents a fully connected layer, ReLU represents a linear rectification function, σ represents a Sigmoid function, MaxPool and AvgPool represent maximum pooling and average pooling respectively, Conv represents a convolution operation, Concat represents a concatenation operation in the channel dimension, and ⊙ represents a Hadamard product; Among them, channel attention W ch The feature map after global average pooling is weighted by the fully connected layer to encode the importance of different channels; spatial attention W sp The pooled feature map is weighted through the convolution layer to emphasize the important spatial areas in the image; finally, the two attentions are fused together by element-wise multiplication to generate a feature map F' that is doubly adjusted in space and channel.

4. The method according to claim 3, characterized in that ,The output layer uses the Softmax activation function to output the classification and location information of the target. The probability value p after the Softmax function conversion i The calculation formula is as follows: Where N is the number of neurons in the output layer, z i is the original output value of the ith neuron.

5. The method according to any one of claims 2 to 4, characterized in that: The specific method of constructing the dataset is as follows: collect and download Terra SAR and GF3 SAR image data for a total of N scenes, with a single scene size of no less than 1k×1k pixels, and divide the data into training dataset, validation dataset, and test dataset in a ratio of 8:1:

1.

6. The method according to claim 5, characterized in that The specific method of dividing the training data into blocks and enhancing the data is as follows: read the SAR aircraft data in the training set and divide it into blocks; the blocking adopts an overlapping blocking method, that is, L×L squares are used to step horizontally and vertically in the SAR remote sensing image, and the step amount is L / 2 to obtain multiple blocks, called Tiles, and use geometric distortion simulation and multi-view imaging simulation to perform data enhancement processing.

7. The method according to claim 6, characterized in that The specific method of training the SAR airport aircraft recognition model based on lightweight attention and deformable convolution is: input the training data set and label set into the built aircraft recognition model, train the lightweight attention module and the deformable convolution module respectively, obtain the network model architecture of the Mth generation of training, and save the optimal model parameters for training.

8. The method according to claim 6 or 7, characterized in that The specific method of testing the trained SAR airport aircraft recognition model based on lightweight attention and deformable convolution is as follows: load the optimal model parameters, input the data set to be identified into the input layer of the trained SAR airport aircraft target recognition network, and the network output layer determines the target type. The output information includes the target type label and confidence.

Images