An Identification Method for SAR Multipolarization Wide-Swath Remote Sensing Images Based on GTAF

Abstract

The present invention relates to the technical field of SAR image processing, and more specifically, to a recognition method for SAR multi-polarization wide-swath remote sensing images based on GTAF. The method includes the following steps: S1. Perform sliding window cropping on the SAR image with a size of W×H = 13×13 to construct a SAR three-dimensional rotating ship target training set and a test set; S2. The Gaussian-type activation function (GTAF), which includes an ordinary real-domain activation function and a Gaussian function, is used to jointly activate the real part and the imaginary part of the neuron; through the forward propagation and backward propagation algorithms of GTAF, use the training set to train the CV-CNN, and through the self-learning of the CV-CNN, extract the deep essential features of the samples. After training all the training data, a SAR multi-polarization wide-swath remote sensing image recognition model can be obtained; S3. Input the test set into the model, and through the fully connected layer and the output layer, better realize the recognition of SAR ground objects.

Description

Technical Field

[0001] The present invention relates to the technical field of SAR image processing, and more specifically, to a method for identifying SAR multi-polarization wide-swath remote sensing images based on GTAF. Background Art

[0002] Deep learning has shown excellent performance in a series of tasks, including computer vision, natural language processing, and speech processing. It involves the following: A convolutional neural network (CNN) is a multi-layer network structure composed of cascaded convolutional layers and pooling layers; CNN usually converts the input image into a feature space, in which the input image becomes linearly separable. To achieve this goal, CNN relies on activation functions to achieve non-linearity and map out features; activation functions endow the neural network with the ability to learn complex patterns by introducing non-linearity; the complex-domain activation functions of complex-domain neural networks are extended from real-domain activation functions such as sigmoid, tanh, ReLU, etc.

[0003] In recent years, the most widely used complex-domain activation function is the real-imaginary activation function (RIAF), which applies separate real-domain activation functions to both the real and imaginary parts of the neuron. However, this separate activation method ignores the internal relationship between the real and imaginary parts of the complex input, losing the integrity of activation. Summary of the Invention

[0004] The present invention provides a method for identifying SAR multi-polarization wide-swath remote sensing images based on GTAF, aiming to better complete the identification of SAR multi-polarization wide-swath remote sensing images.

[0005] The above object is achieved by the following technical solutions:

[0006] A method for identifying SAR multi-polarization wide-swath remote sensing images based on GTAF includes the following steps:

[0007] S1. Perform sliding window cropping on the SAR image with a size of W×H = 13×13 to construct a SAR three-dimensional rotating ship target training set and a test set;

[0008] S2. The Gaussian-type activation function (GTAF), including an ordinary real-domain activation function and a Gaussian function, is used to jointly activate the real and imaginary parts of the neuron; through the forward propagation and backward propagation algorithms of GTAF, use the training set to train the CV-CNN, and through the self-learning of the CV-CNN, extract the deep essential features of the samples. After training all the training data, a SAR multi-polarization wide-swath remote sensing image recognition model can be obtained;

[0009] S3. Input the test set into the model, and through the fully connected layer and the output layer, realize the identification of SAR ground objects. Description of the Drawings

[0010] Figure 1 is the forward propagation flowchart of CV-CNN;

[0011] Figure 2 is the backpropagation flowchart of CV-CNN;

[0012] Figure 3 is the network architecture diagram of CV-CNN;

[0013] Figure 4 is the flowchart of ship target recognition;

[0014] Figure 5 is the recognition result of ground objects in the measured SAR image;

[0015] Figure 6 is the confusion matrix of ground object recognition in the measured SAR image. Detailed Implementation Manner

[0016] A recognition method for SAR multi-polarization wide-swath remote sensing images based on GTAF, comprising the following steps:

[0017] S1. Perform sliding window cropping on the SAR image with a size of W×H = 13×13 to construct a SAR three-dimensional rotating ship target training set and a test set; where W represents the width and H represents the height.

[0018] S2. Use the Gaussian-type activation function (GTAF), which includes a common real-domain activation function and a Gaussian function, to jointly activate the real and imaginary parts of neurons; through the forward propagation and backpropagation algorithms of GTAF, use the training set to train CV-CNN, and through CV-CNN self-learning and extracting the deep essential features of samples, a SAR multi-polarization wide-swath remote sensing image recognition model can be obtained after training all the training data; SAR image processing technology and deep learning technology can better meet the needs of SAR multi-polarization wide-swath remote sensing image recognition and maintain a certain degree of stability for perspective changes, affine transformations, and noises.

[0019] Among them, the Gaussian-type activation function.

[0020] Includes a common real-domain activation function and a Gaussian function, applicable to any given complex-domain neural network, used to jointly activate the real and imaginary parts of neurons, and improve the learning ability and learning speed of neurons. Applying the Gaussian-type activation function to the research of SAR ground object classification field, the results show that GTAF has a fast convergence speed and high accuracy.

[0021] Among them, the forward propagation of the Gaussian-type activation function:

[0022] The complex-domain input of the Gaussian-type activation function After being processed by the Gaussian activation function, the output in the complex domain is obtained. The forward calculation formula is as follows:

[0023]

[0024] Among them, represents the real and imaginary parts of the complex number, j is the imaginary unit, σ(·) represents the activation function in the real number domain, and m, p, q are the learnable parameters in GTAF. Initially, m = 1, p = 0, q = 1.

[0025] Among them, the backpropagation of the Gaussian activation function:

[0026] The error terms of m, p, q are respectively:

[0027]

[0028]

[0029]

[0030] Among them, represents the real and imaginary parts of the error term of the output a of the Gaussian activation function;

[0031] In the backpropagation of the Gaussian activation function, the formulas for updating the parameters of m, p, q are:

[0032] m←m+αδ m

[0033] p←p+αδ p

[0034] q←q+αδ q

[0035] Among them, α is the learning rate.

[0036] Among them, the CV-CNN network architecture design:

[0037] In addition to the input and output layers, the CV-CNN network architecture also has three convolutional layers, one pooling layer, and one fully connected layer. The input layer is 13×13×6, indicating that the size of the SAR slice is 13×13 and the number of channels is 6. Due to convolution and pooling, the size of the feature map will decrease. In the first convolutional layer, the input image is filtered through 6 convolutional filters of size 2×2×6, resulting in 6 feature maps of size 12×12. Then, it passes through an average pooling layer with a pool size of 2×2 and a stride of 1, and the size of the feature map becomes 6×6. The filter size of the second convolutional layer is 2×2×6×12, generating 12 feature maps of size 5×5. The filter size of the third convolutional layer is 2×2×6×12, generating 12 feature maps of size 4×4. Then, the three-dimensional feature is mapped to a one-dimensional vector, generating 192 neurons as the fully connected layer. Finally, the output layer contains c neurons, where c is equal to the number of ground object categories.

[0038] In the training and recognition stage of CV-CNN, through the forward propagation and backpropagation algorithms of the Gaussian activation function, CV-CNN is trained using the training set obtained in the dataset construction stage. By self-learning of CV-CNN and extracting the deep essential features of the samples, the recognition model based on SAR multi-polarization wide-swath remote sensing images can be obtained after training all the training data; the test set is input into the trained CV-CNN model, and the recognition of SAR ground objects is achieved through the fully connected layer and the output layer.

[0039] The specific steps of the forward propagation part of CV-CNN are as follows:

[0040] (1) Initialize the weight matrices and biases of each convolutional layer, fully connected layer, and output layer, and initialize the learnable parameters m = 1, p = 0, q = 1 in each Gaussian activation function;

[0041] (2) Calculate the output of the first convolutional layer without the activation function and apply activation using the Gaussian activation function, i.e.:

[0042]

[0043] where represents the real and imaginary parts of the complex number, j is the imaginary unit, σ(·) represents the real-domain activation function, and m, p, q are the learnable parameters in GTAF, with initial m = 1, p = 0, q = 1.

[0044] (3) Calculate the output of the pooling layer, and the pooling layer uses average pooling;

[0045] (4) According to step (2), calculate the outputs of the second convolutional layer and the third convolutional layer in sequence;

[0046] (5) Stretch the output of the third convolutional layer into a one-dimensional vector as the fully connected layer, multiply it by the weights of the fully connected layer, and then activate it through a Gaussian activation function to obtain the output result of the output layer;

[0047] (6) Calculate the amplitude of each element in the output vector, and the position number of the element with the largest amplitude is the category of the target object.

[0048] The specific steps of the backpropagation part of CV-CNN are as follows:

[0049] (1) Calculate the error terms of each convolutional layer, fully connected layer, and output layer;

[0050] (2) Calculate the parameter gradients of the Gaussian activation functions in the convolutional layer and output layer:

[0051] In the backpropagation of the Gaussian activation function, the error terms of m, p, and q are respectively

[0052]

[0053]

[0054]

[0055] Among them, represent the real and imaginary parts of the error term of the output a of the Gaussian activation function;

[0056] (3) Complete the weight update of each convolutional layer and fully connected layer

[0057] (4) Complete the parameter update of the Gaussian activation functions of each convolutional layer and fully connected layer. In the backpropagation of the Gaussian activation function, the formulas for the parameter updates of m, p, and q are:

[0058] m←m+αδ m

[0059] p←p+αδ p

[0060] q←q+αδ q

[0061] where α is the learning rate.

[0062] Applying this CV-CNN network architecture to the recognition of SAR multi-polarization wide-swath remote sensing images makes full use of the amplitude and phase information of SAR images and has good performance in target object recognition.

[0063] S3. Input the test set into the model, and through the fully connected layer and the output layer, realize the recognition of SAR target objects.

[0064] Among them, the SAR multi-polarization wide-swath remote sensing image recognition part is specifically divided into two stages: dataset construction, CV-CNN training, and CV-CNN recognition.

[0065] In summary, this application includes the forward propagation of the Gaussian activation function, the backward propagation of the Gaussian activation function, the CV-CNN network architecture design, and the SAR multi-polarization wide-swath remote sensing image recognition. The forward propagation part of the Gaussian activation function gives the specific principle of the forward propagation algorithm of the Gaussian activation function. The backward propagation part of the Gaussian activation function gives the specific principle of the backward propagation algorithm of the Gaussian activation function. The CV-CNN network architecture design part clarifies the specific details of the network design. The SAR multi-polarization wide-swath remote sensing image recognition part is based on the Gaussian activation function and the designed CV-CNN network architecture to train and test SAR ground objects to achieve the recognition function. This algorithm can utilize the advantages of the Gaussian activation function and the CV-CNN network in the SAR earth observation scenario to better complete the SAR multi-polarization wide-swath remote sensing image recognition, meet the actual needs, and is convenient to implement.

Claims

1. A recognition method for SAR multi-polarization wide-swath remote sensing images based on GTAF, characterized in that, It includes the following steps: S1. Perform sliding window cropping on the SAR image to construct a SAR three-dimensional rotating ship target training set and a test set; S2. Through the forward propagation and backpropagation algorithms of GTAF, use the training set to train the CV-CNN. Through the self-learning of the CV-CNN, extract the deep essential features of the samples. After training all the training data, a SAR multi-polarization wide-swath remote sensing image recognition model can be obtained; S3. Input the test set into the model, and through the fully connected layer and the output layer, realize the recognition of SAR ground objects; The GTAF includes a common real-number domain activation function and a Gaussian function, which are used to jointly activate the real part and the imaginary part of the neuron; During the forward propagation of GTAF, the input is in the complex domain After being processed by GTAF, the output is in the complex domain The forward calculation formula is as follows: wherein, represent the real part and the imaginary part of a complex number, j is the imaginary unit, σ(·) represents a real-domain activation function, and m, p, q are learnable parameters in GTAF, with an initial value of m = 1, p = 0, and q = 1; In the backpropagation of GTAF, the error terms of m, p, and q are respectively: Among them, represent the real and imaginary parts of the error term of the output a of the Gaussian activation function; In the backpropagation of the Gaussian activation function, the formulas for updating the parameters of m, p, and q are: m ← m + αδ m ; p ← p + αδ p ; q ← q + αδ q ; where α is the learning rate; The CV-CNN includes an input layer, an output layer, three convolutional layers, a pooling layer, and a fully connected layer; Among them, the input layer is 13×13×6, indicating that the size of the SAR slice is 13×13 and the number of channels is 6; In the first convolutional layer, the input image is filtered through 6 convolutional filters with a size of 2×2×6 to obtain 6 feature maps with a size of 12×12; Then, it passes through an average pooling layer with a pool size of 2×2 and a stride of 1, and the size of the feature map becomes 6×6; The filtering size of the second convolutional layer is 2×2×6×12, generating 12 feature maps with a size of 5×5; The filtering size of the third convolutional layer is 2×2×6×12, generating 12 feature maps with a size of 4×4; Then, map the three-dimensional features to a one-dimensional vector to generate 192 neurons as the fully connected layer; Finally, the output layer contains c neurons, where c is equal to the number of ground object categories; The specific steps of the forward propagation part of the CV-CNN are: T1. Initialize the weight matrices and biases of each convolutional layer, fully connected layer, and output layer, and initialize the learnable parameters m = 1, p = 0, q = 1 in each Gaussian activation function; T2. Calculate the output of the first convolutional layer without passing through the activation function: and activate it with GTAF; T3. Calculate the output of the pooling layer, and the average pooling is used in the pooling layer; T4. Calculate the outputs of the second convolutional layer and the third convolutional layer in sequence according to T2; T5. Stretch the output of the third convolutional layer into a one-dimensional vector as the fully connected layer, multiply it by the weights of the fully connected layer, and then activate it through the Gaussian activation function to obtain the output result of the output layer; T6. Calculate the amplitude of each element in the output vector, and the position number of the element with the largest amplitude is the category of the ground object to which it belongs; The specific steps of the backpropagation part of the CV-CNN are: U1. Calculate the error terms of each convolutional layer, fully connected layer, and output layer; U2. Calculate the parameter gradients of the Gaussian activation function in the convolutional layer and the output layer; U3. Complete the weight update of each convolutional layer and fully connected layer; U4. Complete the parameter update of the Gaussian activation function of each convolutional layer and fully connected layer.

2. According to the recognition method described in claim 1, the size of the sliding window clipping is W×H = 13×13.

Images