CN-121982575-A - SAR image change detection method combining convolution and mixed attention

CN121982575ACN 121982575 ACN121982575 ACN 121982575ACN-121982575-A

Abstract

The invention discloses a SAR image change detection method combining convolution and mixed attention, which comprises the implementation steps of firstly generating a difference map for two SAR images by using a composite neighborhood intensity difference method. And performing pre-classification processing on the difference map by using a hierarchical FCM clustering algorithm to generate a pseudo-label matrix, selecting high-probability sample pixels with variation and invariable in pseudo-label pixels, extracting the spatial positions of the pixels, taking a pixel block taking the pixel points as the center as a training set on the pixels of the corresponding spatial positions of the original two SAR images, and extracting pixel blocks taking all the pixel points as the centers from the original two SAR images as a test set. And training the training sample set by using the neural network combining convolution and mixed attention, and then performing change detection analysis on the test set by using the trained network to generate a final change detection result graph. The method has clear advantages in inhibiting SAR speckle noise and variation detection precision.

Inventors

LI CHUNYANG
WANG JUN
Fu Qiongjun
CAO RUIXIU
NIU SANKU

Assignees

浙江工业大学

Dates

Publication Date: 20260505
Application Date: 20260204

Claims (7)

1. The SAR image change detection method combining convolution and mixed attention is characterized by comprising the following steps of: step 1, acquiring two SAR images at different moments shot in the same geographic area; Step 2, analyzing the two SAR images obtained in the step 1 by a composite neighborhood intensity difference method to generate a difference image; step 3, pre-classifying the difference image obtained in the step 2, dividing the difference image into three categories of variation, non-variation and uncertainty, selecting a certain proportion of sample points from the divided variation and non-variation categories to construct a training data set, and constructing a test data set by using all pixel points of the two SAR images; Step 4, constructing a neural network combining convolution and mixed attention; the built neural network combining convolution and mixed attention consists of a convolution module and a mixed attention module, wherein the mixed attention consists of self-attention and cross boundary attention through an adaptive weight learning fusion module; Step 5, using the training data set obtained in the step 3 for training the neural network constructed in the step 4 by combining convolution and mixed attention; and 6, testing the test data set by using a trained neural network combining convolution and mixed attention to obtain a predictive label of the test data set, and further generating a final change detection result graph.
2. The SAR image change detection method combining convolution and mixed attention according to claim 1, wherein in step 1, the original SAR image is subjected to radiation correction and geometric registration preprocessing to obtain two SAR images at different moments.
3. The method for detecting the change of the SAR image by combining convolution and mixed attention according to claim 1, wherein the composite neighborhood intensity difference method adopted in the step 2 is to fuse a neighborhood difference operator and a relative intensity difference value in a texture adaptive weighting mechanism mode so as to generate a smoother difference map.
4. The SAR image variation detection method combining convolution and mixed attention according to claim 1, wherein said step 3 specifically comprises the steps of: (3.1) pre-classifying the difference map by using a hierarchical FCM clustering algorithm to obtain pseudo tag matrixes with pseudo tag values of 0,1,0.5 respectively, namely, generating high-probability samples and uncertain samples which are judged to be unchanged or changed after pre-classifying; (3.2) selecting 5-15% of pixels from the pixels with pseudo tag values of 0 and 1, and extracting the periphery of the pixel points on the original two SAR images according to the spatial positions of the pixels As a training set. The image block is extracted by field filling 0 for the edge pixel point, The value of (2) is an odd number not less than 3; (3.3) extracting all the pixel points around on the original two SAR images As a test set. The image block is extracted by field filling 0 for the edge pixel point, The value of (2) is an odd number not less than 3.
5. The SAR image change detection method combining convolution and mixed attention according to claim 1, wherein said step 4 is implemented by constructing a neural network combining convolution and mixed attention by: (4.1) self-attention module construction: dividing the training sample set obtained in the step 3 into non-overlapping parts The image block is then processed to determine, Applying a linear embedding layer to each image block, converting it into a token with fixed dimension, embedding the embedded token, and performing linear transformation to calculate the query Key and key Sum value The three input features obtained by the linear transformation are input into self-attention calculation to obtain a self-attention output sequence, and the self-attention output sequence is converted into output features through an output projection layer; (4.2) Cross-boundary attention module construction: The cross boundary attention consists of a boundary attention module and a cross attention self-adaptive weight learning fusion module, wherein the two SAR images obtained in the step 1 are divided into Is used for the image blocks of the (c), Respectively applying boundary attention modules to each image to generate corresponding boundary weights, adopting two depth separable convolution layers and performing differential edge detection to obtain an edge intensity image, performing boundary image normalization on the edge intensity image to obtain an edge space weight image, adopting the self-adaptability of a lightweight channel attention enhancement channel dimension, performing element-by-element multiplication operation on the channel attention weight image and the edge space weight image to obtain boundary weights, performing element-by-element multiplication on the boundary weights and input features to obtain boundary enhancement features, flattening the boundary enhancement features into a space sequence form, and using a first SAR image Boundary enhancement of (c) is characterized by With a second SAR image Boundary enhancement of (c) is characterized by And Calculating multiple head attention, linearly projecting the attention output, remodelling to spatial tensor, and combining with The characteristics are fused through a convolution layer, a batch normalization layer and an activation function to obtain enhanced first time phase characteristics Second time phase characteristics Maintaining unchanged, combining the enhanced features with the second phase features Performing channel dimension connection to obtain enhanced cross boundary attention characteristics; (4.3) constructing an adaptive weight learning fusion module: The global feature of the self-attention output and the enhanced cross boundary feature obtained by the cross boundary attention are connected in the channel dimension by adopting Concate operation, a fusion weight matrix is generated by adopting a convolution layer and a Sigmoid activation function, the global feature and the enhanced cross boundary feature are subjected to weighted fusion according to the weight matrix, and the fusion feature is obtained by learning residual connection; And (4.4) classifying the fusion features and the local features through a fully-connected network for three times after carrying out Concate operation on the channel dimension to obtain a changed and unchanged prediction label result, and optimizing parameters of the neural network combining convolution and mixed attention by adopting an Adam optimizer according to a loss function of a prediction label result calculation model.
6. The SAR image variation detection method combining convolution and mixed attention according to claim 1, wherein the channel attention weight map calculation formula is as follows: ; Wherein, the The input characteristics are represented as such, Representing a global average pooling of the data, The activation function is represented as a function of the activation, And Indicating that the full-link layer is to be formed, Representation of The function is activated.
7. The SAR image change detection method combining convolution and mixed attention of claim 1, wherein the convolution module performs feature transformation and channel compression using a large convolution kernel input feature, normalized and After activating the function, obtaining local features, recovering the number of channels to original dimension by using a convolution kernel with the same size, and using normalization and again Activating the function, resulting in enhanced local features.

Description

SAR image change detection method combining convolution and mixed attention Technical Field The invention relates to the field of synthetic aperture radar image detection methods, in particular to a SAR image change detection method combining convolution and mixed attention. Background With the wide application of satellite imaging technology in daily life, the remote sensing image data of various terrains are increasingly abundant. In the field of computer vision and in research of intelligent recognition of remote sensing images, remote sensing image change detection has become an important direction. Although most of the related researches are mainly focused on the overall feature extraction and processing of the change detection region, the extraction and analysis of the boundary features are relatively insufficient. In fact, the ground object boundary of remote sensing often contains rich semantic information. Therefore, the extraction and processing of the boundary information have important significance for improving the accuracy of remote sensing image change detection. In recent years, deep learning techniques, particularly Convolutional Neural Networks (CNNs), have been widely used in the task of change detection. Typical implementations include dual stream encoder-decoder architecture, twin networks, and U-Net based architectures, among others. The method can automatically extract multi-level spatial features and improve detection performance to a certain extent. However, since CNN focuses mainly on information processing in a local receptive field, modeling capability of the CNN on long-distance dependency is limited, and it is difficult to effectively capture the change of a wide-range ground feature structure or the consistency of global semantics in a remote sensing image. To overcome these limitations, researchers have begun exploring the introduction of visual transducers (Vision Transformer, viT) into the change detection task. ViT utilize a self-attention mechanism to model global interactions between image blocks, significantly enhancing the model's ability to understand context information. However, viT is subject to large computational overhead when processing high-resolution remote sensing images and does not perform well in capturing local detail, particularly boundary areas, where actual feature changes tend to manifest themselves on these edges or contours first. Therefore, how to effectively extract the local detail information and the global information and realize effective fusion becomes a key challenge of the change detection effect. Disclosure of Invention The invention provides a SAR image change detection method combining convolution and mixed attention, which aims to solve the problems in the background technology. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: The SAR image change detection method combining convolution and mixed attention comprises the following steps: step 1, acquiring two SAR images at different moments; Step 2, analyzing the two SAR images obtained in the step 1 by a composite neighborhood intensity difference method to obtain a difference image; step 3, pre-classifying the difference image obtained in the step 2, dividing the difference image into three categories of variation, non-variation and uncertainty, selecting a certain proportion of sample points from the divided variation and non-variation categories to construct a training data set, and constructing a test data set by using all pixel points of the two SAR images; Step 4, constructing a neural network combining convolution and mixed attention; Step 5, using the training data set obtained in the step 3 for training the neural network constructed in the step 4 by combining convolution and mixed attention; And 6, testing the test data set by using a trained neural network combining convolution and mixed attention to obtain a predictive label of the test data set, and further obtaining a change detection result of the whole image. Furthermore, the step 1 further comprises the steps of carrying out radiation correction and geometric registration preprocessing on the original SAR image to obtain SAR images at two different moments. In the further step 2, the composite neighborhood intensity difference method adopted in the step 2 is to fuse the neighborhood difference operator and the relative intensity difference value in a texture self-adaptive weighting mechanism mode, so that a smoother difference map is generated. Further, the neural network constructed in the step 4 and combining convolution and mixed attention consists of a convolution module and a mixed attention module, and the mixed attention consists of self-attention and cross-boundary attention through an adaptive weight learning fusion module. Further, the self-attention module in step 5 divides the training sample set obtained in step 3 into non-overlapping setsThe image block is