A remote sensing image authenticity identification method and system

By extracting features from image domain and frequency domain encoders and performing deep fusion, the problem of insufficient feature fusion in the authentication of remote sensing images is solved, and higher authentication accuracy and robustness are achieved.

CN118608975BActive Publication Date: 2024-11-05深圳市规划和自然资源数据管理中心(深圳市空间地理信息中心)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411066838.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-06
Publication Date
2024-11-05
Estimated Expiration
2044-08-06

AI Technical Summary

Technical Problem

Existing remote sensing image authenticity identification techniques are insufficient in dealing with high-frequency artifact pattern differences and cross-dataset generalization, and are difficult to effectively combine image domain and frequency domain features for synergistic complementarity, resulting in insufficient identification accuracy.

Method used

Features are extracted using an image domain encoder and a frequency domain encoder, initially fused using a multi-domain representation unified mapping module, and then optimized using a hybrid domain feature optimization module. Finally, a simple linear classifier is used to distinguish between true and false information.

Benefits of technology

It achieves higher accuracy and generalization in identifying genuine and fake remote sensing images, and can more accurately capture the features of different fake images, thus improving the robustness of identification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118608975B_ABST
    Figure CN118608975B_ABST
Patent Text Reader

Abstract

The present invention discloses a remote sensing image authenticity identification method and system, which relates to the field of image processing technology. The method comprises: acquiring a remote sensing image, performing image domain feature extraction processing on the remote sensing image based on an image domain encoder to obtain image domain features; performing frequency domain feature extraction processing on the remote sensing image based on a frequency domain encoder to obtain frequency domain features; performing preliminary fusion of the image domain features and frequency domain features based on a multi-domain representation unified mapping module to obtain preliminary fusion features; optimizing the preliminary fusion features based on a mixed domain feature optimization module to obtain target fusion features; performing authenticity identification processing of remote sensing images based on a linear classifier using the target fusion features to obtain authenticity identification results. The present invention achieves more advanced remote sensing image authenticity identification accuracy and captures generalization features about different false images.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a remote sensing image authenticity identification method and system. Background Art

[0002] Existing remote sensing image authenticity identification technologies are mainly designed for natural images and can be roughly divided into methods based on image domain features and frequency domain features. Methods based on image domain features capture the forgery traces of false images in the image domain by designing manual features or modifying neural networks with powerful feature extraction capabilities. As an early mainstream method, the key to its success often relies on the designer's deep insight and accurate understanding of the false features of the image. However, methods based on image domain features usually only achieve good results on low-quality generated images with obvious traces of forgery. With the rapid progress of adversarial generative models, it is increasingly difficult to distinguish the authenticity of images with the naked eye.

[0003] In this regard, some studies have found that high-frequency artifacts can be used as key clues to identify the authenticity of images. This high-frequency artifact is essentially caused by the upsampling defects of the adversarial generative model decoder and the smoothing effect of the model training loss function. Based on this prior law, a large number of image authenticity identification technologies have begun to focus on capturing unique forgery traces in the frequency domain and have achieved impressive results. However, there are differences in the patterns of high-frequency artifacts in images generated by different adversarial generative models, and forged image detection algorithms based on frequency domain features can easily overfit to specific patterns in the training set and are difficult to generalize across data sets. In addition, with the further breakthroughs in AIGC technology represented by diffusion models, the reliability of image authenticity identification methods that rely solely on frequency domain priors is also facing severe challenges. Although some image authenticity identification technologies based on mixed domain features have achieved certain improvements in generalization by combining image domain and frequency domain features, they are insufficient in frequency domain feature extraction. Specifically, these methods often fail to fully decouple and highlight high-frequency information that is critical to identifying forgery traces due to the lack of targeted frequency domain feature extractors. The lack of a mechanism for deep fusion is not conducive to giving full play to the synergistic and complementary role of the two features in the detection of false remote sensing images. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art. The present invention provides a remote sensing image authenticity identification method and system, which achieves more advanced remote sensing image authenticity identification accuracy and captures generalized features about different false images.

[0005] In order to solve the above technical problems, the present invention provides a remote sensing image authenticity identification method, characterized in that the method comprises:

[0006] Acquire a remote sensing image, and perform image domain feature extraction processing on the remote sensing image based on an image domain encoder to obtain image domain features;

[0007] Performing frequency domain feature extraction processing on the remote sensing image based on a frequency domain encoder to obtain frequency domain features;

[0008] Based on a multi-domain representation unified mapping module, the image domain features and the frequency domain features are preliminarily fused to obtain a preliminary fusion feature;

[0009] Optimizing the preliminary fusion features based on a hybrid domain feature optimization module to obtain target fusion features;

[0010] Based on the target fusion features, the authenticity identification processing of the remote sensing image is performed to obtain the authenticity identification result.

[0011] Optionally, the expression of the image domain encoder is:

[0012] ,

[0013] Among them, f is the output of the image domain encoder, x is the input remote sensing image, is a convolution operation, BN is a batch normalization operation, ReLU is an activation function, and Maxpool is a maximum pooling operation.

[0014] Optionally, the frequency domain encoder includes an image frequency domain representation module, a high-frequency feature extraction module and a frequency convolution layer.

[0015] Optionally, the image frequency domain representation module is used to perform medium and low frequency information filtering processing on the remote sensing image, wherein the process of medium and low frequency information filtering processing is:

[0016] ,

[0017] in, is the remote sensing image after filtering the medium and low frequency information, iFFT() is the inverse fast Fourier transform, B() is the high pass filter, FFT() is the fast Fourier transform, x is the input remote sensing image,

[0018] The operation rule of the high-pass filter is:

[0019] ,

[0020] Among them, B() is a high-pass filter, It is the element in the i-th row and j-th column of the input feature x in the remote sensing image.

[0021] Optionally, the high-frequency feature extraction module is used to perform high-frequency enhancement on the remote sensing image after the medium- and low-frequency information filtering process to obtain a remote sensing image after high-frequency enhancement, wherein the process of high-frequency enhancement is:

[0022] ,

[0023] in, is the remote sensing image after high frequency enhancement. To perform the inverse fast Fourier transform operation across space, B() is a high-pass filter, To perform a fast Fourier transform operation across space, is the inverse fast Fourier transform operation across the channel dimension, is the fast Fourier transform operation across channel dimensions, f is the input feature, For high frequency enhancement in the spatial dimension, It is the high frequency enhancement in the channel dimension.

[0024] Optionally, the frequency convolution layer is used to perform feature analysis on the high-frequency enhanced remote sensing image to obtain frequency domain features, wherein the feature analysis process is:

[0025] ,

[0026] ,

[0027] ,

[0028] in, is the frequency domain feature, is the fast Fourier transform, f is the input feature, is the convolutional layer applied to the phase spectrum, is the convolutional layer applied to the amplitude spectrum, is the frequency domain feature after spectrum decomposition, is the activation function, is the frequency domain feature, is the inverse fast Fourier transform operation.

[0029] Optionally, the multi-domain representation-based unified mapping module performs preliminary fusion of the image domain features and the frequency domain features to obtain preliminary fusion features, including:

[0030] A shortcut connection across projection units is set, and feature mapping is performed on the image domain features and the frequency domain features based on the shortcut connection across projection units and a plurality of feature projection units to obtain preliminary fusion features, wherein each feature projection unit is generated by stacking a plurality of 3*3 convolutional layers and 1*1 convolutional layers, and each convolutional layer includes a nonlinear activation function.

[0031] Optionally, the optimizing the preliminary fusion feature based on the hybrid domain feature optimization module to obtain the target fusion feature includes:

[0032] Aggregating spatial information of the preliminary fusion features based on average pooling and maximum pooling operations to obtain a spatial context descriptor;

[0033] Generate a channel attention map using the spatial context descriptor based on a shared network;

[0034] Based on the channel attention map, channel attention processing is performed on the preliminary fusion feature to obtain the preliminary fusion feature after the channel attention processing, wherein the process of the channel attention processing is:

[0035] ,

[0036] in, is the result of channel attention processing, is the channel attention mechanism operation, f is the initial fusion feature, It is a pixel-level multiplication operation;

[0037] Based on the pooling operation, the channel information of the preliminary fusion features after the channel attention processing is summarized to obtain a two-dimensional graph;

[0038] Based on the convolution layer, the two-dimensional graph is concatenated and convolved to obtain a two-dimensional spatial attention graph;

[0039] Based on the two-dimensional spatial attention map, spatial attention processing is performed on the preliminary fusion features after channel attention processing to obtain target fusion features, wherein the process of spatial attention processing is:

[0040] ,

[0041] in, is the result of spatial attention processing, is the result of channel attention processing, Operates for the spatial attention mechanism.

[0042] Optionally, the channel attention mechanism operates as follows:

[0043] ,

[0044] in, is the channel attention mechanism operation, is the sigmoid function, is a multi-layer perceptron, is the average pooling operation, f is the initial fusion feature, is the maximum pooling operation, and are the learnable weight parameters of the multilayer perceptron, is the average pooling feature, is the maximum pooling feature;

[0045] The spatial attention mechanism operates as follows:

[0046] ,

[0047] in, For the spatial attention mechanism operation, is the sigmoid function, is a convolution operation with a filter size of 7×7, is the average pooling operation, f is the initial fusion feature, is the maximum pooling operation, is the average pooling feature, is the maximum pooling feature.

[0048] In addition, the present invention also provides a remote sensing image authenticity identification system, the system comprising:

[0049] Image domain feature extraction module: used to obtain remote sensing images, perform image domain feature extraction processing on the remote sensing images based on the image domain encoder, and obtain image domain features;

[0050] Frequency domain feature extraction module: used for performing frequency domain feature extraction processing on the remote sensing image based on a frequency domain encoder to obtain frequency domain features;

[0051] A feature preliminary fusion module: used to perform preliminary fusion of the image domain features and the frequency domain features based on a multi-domain representation unified mapping module to obtain preliminary fusion features;

[0052] Feature optimization module: used to optimize the preliminary fusion features based on the hybrid domain feature optimization module to obtain target fusion features;

[0053] Image authenticity identification module: used to perform authenticity identification processing of remote sensing images based on the target fusion features to obtain authenticity identification results.

[0054] In the embodiment of the present invention, taking into account the differences between the image domain and the frequency domain and the key significance of high-frequency features for the authenticity identification of images, the remote sensing image is subjected to frequency domain feature extraction processing based on the frequency domain encoder to more accurately capture the forgery traces of the false image in the frequency domain. The image domain features and the frequency domain features are initially fused based on the multi-domain representation unified mapping module, and then the initial fusion features are optimized based on the hybrid domain feature optimization module. This can achieve deep fusion of features to address the problem of hindering the synergistic complementation of differentiated image domain and spatial features due to the lack of an effective feature fusion mechanism, achieve stronger generalization and robustness by capturing richer and more comprehensive features, achieve more advanced remote sensing image authenticity identification accuracy, and capture generalization features about different false images. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0056] Figure 1 is a schematic diagram of a flow chart of a remote sensing image authenticity identification method in an embodiment of the present invention;

[0057] Figure 2 It is a schematic diagram of the structure of the remote sensing image authenticity identification system in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] Embodiment 1

[0060] See also Figure 1 , Figure 1 : is a flow chart of a remote sensing image authenticity identification method in an embodiment of the present invention, the method comprising:

[0061] S11: Acquire a remote sensing image, and perform image domain feature extraction processing on the remote sensing image based on an image domain encoder to obtain image domain features;

[0062] In the specific implementation process of the present invention, the expression of the image domain encoder is:

[0063] ,

[0064] Among them, f is the output of the image domain encoder, x is the input remote sensing image, is a convolution operation, BN is a batch normalization operation, ReLU is an activation function, and Maxpool is a maximum pooling operation.

[0065] Specifically, remote sensing images are acquired based on earth observation equipment, and image domain feature extraction is performed on the remote sensing images based on an image domain encoder to obtain image domain features. In order to extract the forgery pattern in the image domain, a convolutional layer with good spatial feature extraction capability is introduced to learn and characterize the geometric, structural features and local texture features of the remote sensing images. It is worth noting that the features extracted by the shallow convolutional network are more generalized in the authenticity identification of images of different data. The main reason behind this is that high-level semantic features are usually not affected by local features because they are too abstract, and are shared in true and false images, which is not conducive to the characterization of subtle traces of forgery in the image. Inspired by this, a simple single-layer convolution is used to construct an image domain encoder to capture finer image domain features and enhance the separability of true and false images. The expression of the image domain encoder is:

[0066] ,

[0067] Among them, f is the output of the image domain encoder, x is the input remote sensing image, is a convolution operation, BN is a batch normalization operation, ReLU is an activation function, and Maxpool is a maximum pooling operation.

[0068] S12: performing frequency domain feature extraction processing on the remote sensing image based on a frequency domain encoder to obtain frequency domain features;

[0069] In the specific implementation process of the present invention, the frequency domain encoder includes an image frequency domain representation module, a high-frequency feature extraction module and a frequency convolution layer.

[0070] Furthermore, the image frequency domain representation module is used to perform medium and low frequency information filtering processing on the remote sensing image, wherein the process of medium and low frequency information filtering processing is:

[0071] ,

[0072] in, is the remote sensing image after filtering the medium and low frequency information, iFFT() is the inverse fast Fourier transform, B() is the high pass filter, FFT() is the fast Fourier transform, x is the input remote sensing image,

[0073] The operation rule of the high-pass filter is:

[0074] ,

[0075] Among them, B() is a high-pass filter, It is the element in the i-th row and j-th column of the input feature x in the remote sensing image.

[0076] Furthermore, the high-frequency feature extraction module is used to perform high-frequency enhancement on the remote sensing image after the medium and low frequency information filtering process to obtain the remote sensing image after high-frequency enhancement, wherein the process of high-frequency enhancement is:

[0077] ,

[0078] in, is the remote sensing image after high frequency enhancement. To perform the inverse fast Fourier transform operation across space, B() is a high-pass filter, To perform a fast Fourier transform operation across space, is the inverse fast Fourier transform operation across the channel dimension, is the fast Fourier transform operation across channel dimensions, f is the input feature, For high frequency enhancement in the spatial dimension, It is the high frequency enhancement in the channel dimension.

[0079] Furthermore, the frequency convolution layer is used to perform feature analysis on the high-frequency enhanced remote sensing image to obtain frequency domain features, wherein the feature analysis process is:

[0080] ,

[0081] ,

[0082] ,

[0083] in, is the frequency domain feature, is the fast Fourier transform, f is the input feature, is the convolutional layer applied to the phase spectrum, is the convolutional layer applied to the amplitude spectrum, is the frequency domain feature after spectrum decomposition, is the activation function, is the frequency domain feature, is the inverse fast Fourier transform operation.

[0084] Specifically, high-frequency information is crucial for detecting counterfeit images obtained by most generation methods, but in the original image domain, different frequency information is coupled with each other, resulting in the high-frequency components that can provide key information for authenticity detection being masked and difficult to fully mine. In this regard, a frequency domain encoder is designed. First, frequency decoupling and medium and low frequency filtering are performed at different levels such as images and features to fully extract the high-frequency information of the image. Then a special frequency feature convolution module is designed to directly perform deep feature mining on the phase spectrum and amplitude spectrum of high-frequency information. Frequency domain feature extraction processing is performed on the remote sensing image based on the frequency domain encoder. The frequency domain encoder includes an image frequency domain representation module, a high-frequency feature extraction module and a frequency convolution layer. The image frequency domain representation module is used to filter the medium and low frequency information of the remote sensing image to highlight the high-frequency information that is more critical to identifying traces of counterfeiting. The process of medium and low frequency information filtering processing is as follows:

[0085] ,

[0086] in, is the remote sensing image after filtering the medium and low frequency information, iFFT() is the inverse fast Fourier transform, B() is the high pass filter, FFT() is the fast Fourier transform, x is the input remote sensing image, the fast Fourier transform is the conversion of the remote sensing image from the spatial domain to the frequency domain, and the zero frequency component is moved to the center, the inverse fast Fourier transform is the conversion of the extracted high frequency information back to the spatial domain, and the high pass filter is used to extract high frequency information from the features.

[0087] The operation rule of the high-pass filter is:

[0088] ,

[0089] Among them, B() is a high-pass filter, It is the element in the i-th row and j-th column of the input feature x in the remote sensing image.

[0090] The high-frequency feature extraction module is used to perform high-frequency enhancement on the remote sensing image after filtering the medium and low frequency information, and obtain the remote sensing image after high-frequency enhancement. The frequency patterns of forged images by different methods are often different, which may cause the model to overfit only the frequency pattern of the current training data, thereby damaging the generalization ability. In this regard, the high-frequency feature extraction module forces the frequency encoder to give priority to the high-frequency information in the feature space to alleviate the tendency of the model to overfit the frequency pattern unique to the current data, that is, high-frequency enhancement is performed in the spatial dimension and the channel dimension respectively, wherein the process of high-frequency enhancement is:

[0091] ,

[0092] in, is the remote sensing image after high frequency enhancement. To perform the inverse fast Fourier transform operation across space, B() is a high-pass filter, To perform a fast Fourier transform operation across space, is the inverse fast Fourier transform operation across the channel dimension, is the fast Fourier transform operation across channel dimensions, f is the input feature, For high frequency enhancement in the spatial dimension, It is a high-frequency enhancement in the channel dimension. The high-frequency feature extraction module enables the high-frequency information of the intermediate features to be further purified and enhanced, thus better supporting the identification of forgery traces.

[0093] The frequency convolution layer is used to perform feature analysis on the high-frequency enhanced remote sensing image to obtain frequency domain features. The frequency convolution layer directly performs feature analysis on the phase and amplitude spectrum of the high-frequency component of the image obtained by the above module, thereby avoiding overfitting of insignificant low-frequency information in the data. The feature analysis process is as follows:

[0094] ,

[0095] ,

[0096] ,

[0097] in, is the frequency domain feature, is the fast Fourier transform, f is the input feature, is the convolutional layer applied to the phase spectrum, is the convolutional layer applied to the amplitude spectrum, is the frequency domain feature after spectrum decomposition, is the activation function, is the frequency domain feature, It is the inverse fast Fourier transform operation; the fast Fourier transform converts the input features into the frequency domain space, so that the obtained frequency domain features are complex tensors, which can be further decomposed into phase and amplitude spectra. Taking into account the fundamental difference between the two features in the physical sense, two convolutional layers with the same structure but non-shared parameters are set and They are applied to the phase spectrum and amplitude spectrum respectively to fully learn the frequency domain features. Finally, the learned frequency information is converted back to the standard feature space using the inverse fast Fourier transform to facilitate subsequent fusion and complementation with the image domain features.

[0098] S13: Preliminarily fusing the image domain features and the frequency domain features based on a multi-domain representation unified mapping module to obtain preliminary fusion features;

[0099] In the specific implementation process of the present invention, the unified mapping module based on multi-domain representation performs preliminary fusion of the image domain features and the frequency domain features to obtain preliminary fusion features, including: setting shortcut connections across projection units, performing feature mapping on the image domain features and the frequency domain features based on shortcut connections across projection units and a number of feature projection units to obtain preliminary fusion features, wherein each feature projection unit is generated by stacking a number of 3*3 convolutional layers and 1*1 convolutional layers, and each convolutional layer includes a nonlinear activation function.

[0100] Specifically, in order to give full play to the complementary role of image domain and frequency domain features, a multi-domain representation unified mapping module is set up. It aims to overcome the significant differences between image domain and frequency domain features, and achieve the initial fusion of the two by uniformly mapping multi-domain features to a common feature subspace. The differences in content and physical meaning between image domain and frequency domain features are also the basis for their complementarity. If the two features are simply strictly aligned and fused, it will inevitably destroy their unique key information, and ultimately lose the meaning of fusion due to the damage to complementarity. Therefore, how to ensure that common ground is sought while reserving differences is the core issue for achieving effective fusion of multi-domain features. In this regard, the design process of the multi-domain representation unified mapping module mainly follows the following two principles. The first principle is a higher degree of nonlinearity to enhance the ability of feature transformation. In the multi-domain representation unified mapping module, a total of 3 feature projection units are set up, each of which is composed of several 3*3 and 1*1 convolutions stacked to achieve complex feature mapping. In addition, a nonlinear activation function ReLU is introduced between the convolutional layers to increase the overall nonlinear feature transformation capability. The second principle is to provide quick connections to ensure that key information is fully preserved: to avoid the loss of key information in complex feature transformations, quick connections across projection units are set up so that some underlying features have the opportunity to be fully preserved. Based on the above two principles, two branches with the same structure but non-shared parameters are set for image domain and frequency domain features.

[0101] S14: Optimizing the preliminary fusion features based on the hybrid domain feature optimization module to obtain target fusion features;

[0102] In the specific implementation process of the present invention, the hybrid domain feature optimization module is used to optimize the preliminary fusion features to obtain target fusion features, including: performing spatial information aggregation on the preliminary fusion features based on average pooling and maximum pooling operations to obtain a spatial context descriptor; generating a channel attention map based on a shared network using the spatial context descriptor; performing channel attention processing on the preliminary fusion features based on the channel attention map to obtain preliminary fusion features after channel attention processing, wherein the process of channel attention processing is:

[0103] ,

[0104] in, is the result of channel attention processing, is the channel attention mechanism operation, f is the initial fusion feature, It is a pixel-level multiplication operation;

[0105] Based on the pooling operation, the channel information of the preliminary fusion features after the channel attention processing is summarized to obtain a two-dimensional graph; based on the convolution layer, the two-dimensional graph is concatenated and convolved to obtain a two-dimensional spatial attention graph; based on the two-dimensional spatial attention graph, the preliminary fusion features after the channel attention processing are subjected to spatial attention processing to obtain target fusion features, wherein the process of spatial attention processing is:

[0106] ,

[0107] in, is the result of spatial attention processing, is the result of channel attention processing, Operates for the spatial attention mechanism.

[0108] Furthermore, the channel attention mechanism operates as follows:

[0109] ,

[0110] in, is the channel attention mechanism operation, is the sigmoid function, is a multi-layer perceptron, is the average pooling operation, f is the initial fusion feature, is the maximum pooling operation, and are the learnable weight parameters of the multilayer perceptron, is the average pooling feature, is the maximum pooling feature.

[0111] The spatial attention mechanism operates as follows:

[0112] ,

[0113] in, For the spatial attention mechanism operation, is the sigmoid function, is a convolution operation with a filter size of 7×7, is the average pooling operation, f is the initial fusion feature, is the maximum pooling operation, is the average pooling feature, is the maximum pooling feature.

[0114] Specifically, the hybrid domain feature optimization module is constructed by introducing the attention mechanism on the basis of the standard residual convolutional network. On the one hand, the low-level image and frequency features extracted by the encoder are further combined and optimized, and on the other hand, the model learns to adaptively focus on the more critical features as clues. After the multi-domain representation unified mapping module, the image domain and spatial domain features can be preliminarily aligned and added together, but how to give full play to their complementary role is still quite challenging. Considering the complexity of remote sensing images themselves, the diversity and concealment of forgery traces, the underlying visual and frequency features extracted by the encoder are rich but not always effective, and there may even be certain contradictions between them. In this regard, in addition to integrating and optimizing these basic features, another important purpose of the hybrid domain feature optimization module is to achieve adaptive attention to the key features. First, the backbone of the hybrid domain feature optimization model is constructed with multi-layer residual convolution. On this basis, a cross-space and cross-channel dual attention mechanism is introduced to selectively emphasize important features in different dimensions to improve the accuracy of authenticity identification. Based on the average pooling and maximum pooling operations, the spatial information of the preliminary fusion features is aggregated to obtain a spatial context descriptor, that is, two different spatial context descriptors are generated, namely, an average pooling feature descriptor and a maximum pooling feature descriptor. Based on the shared network, the spatial context descriptors are used to generate a channel attention map, and the two spatial context descriptors are transmitted to the shared network constructed by the multi-layer perceptron to generate a channel attention map; based on the channel attention map, the preliminary fusion features are subjected to channel attention processing to obtain preliminary fusion features after channel attention processing, wherein the process of channel attention processing is:

[0115] ,

[0116] in, is the result of channel attention processing, is the channel attention mechanism operation, f is the initial fusion feature, is a pixel-level multiplication operation. The channel attention mechanism operation is:

[0117] ,

[0118] in, is the channel attention mechanism operation, is the sigmoid function, is a multi-layer perceptron, is the average pooling operation, f is the initial fusion feature, is the maximum pooling operation, and are the learnable weight parameters of the multilayer perceptron, is the average pooling feature, is the maximum pooling feature. Based on the pooling operation, the channel information of the preliminary fusion features after the channel attention processing is summarized, that is, the channel information of the features is summarized through the maximum pooling operation and the average pooling operation to obtain a two-dimensional graph, that is, a maximum pooling two-dimensional graph and an average pooling two-dimensional graph; based on the convolution layer, the two-dimensional graph is concatenated and convolved to obtain a two-dimensional spatial attention graph; based on the two-dimensional spatial attention graph, the preliminary fusion features after the channel attention processing are subjected to spatial attention processing to obtain the target fusion feature, wherein the process of spatial attention processing is:

[0119] ,

[0120] in, is the result of spatial attention processing, is the result of channel attention processing, is the spatial attention mechanism operation. The spatial attention mechanism operation is:

[0121] ,

[0122] in, For the spatial attention mechanism operation, is the sigmoid function, is a convolution operation with a filter size of 7×7, is the average pooling operation, f is the initial fusion feature, is the maximum pooling operation, is the average pooling feature, is the maximum pooling feature.

[0123] S15: Performing authenticity identification processing on the remote sensing image based on the target fusion features to obtain an authenticity identification result.

[0124] In the specific implementation process of the present invention, the target fusion features are input into a simple linear classifier to perform authenticity identification processing on the remote sensing image to obtain an authenticity identification result.

[0125] In the embodiment of the present invention, taking into account the differences between the image domain and the frequency domain and the key significance of high-frequency features for the authenticity identification of images, the remote sensing image is subjected to frequency domain feature extraction processing based on the frequency domain encoder to more accurately capture the forgery traces of the false image in the frequency domain. The image domain features and the frequency domain features are initially fused based on the multi-domain representation unified mapping module, and then the initial fusion features are optimized based on the hybrid domain feature optimization module. This can achieve deep fusion of features to address the problem of hindering the synergistic complementation of differentiated image domain and spatial features due to the lack of an effective feature fusion mechanism, achieve stronger generalization and robustness by capturing richer and more comprehensive features, achieve more advanced remote sensing image authenticity identification accuracy, and capture generalization features about different false images.

[0126] Embodiment 2

[0127] See also Figure 2 , Figure 2 : is a schematic diagram of the structure of a remote sensing image authenticity identification system in an embodiment of the present invention, the system comprises:

[0128] Image domain feature extraction module 21: used to obtain remote sensing images, perform image domain feature extraction processing on the remote sensing images based on an image domain encoder, and obtain image domain features;

[0129] In the specific implementation process of the present invention, the expression of the image domain encoder is:

[0130] ,

[0131] Among them, f is the output of the image domain encoder, x is the input remote sensing image, is a convolution operation, BN is a batch normalization operation, ReLU is an activation function, and Maxpool is a maximum pooling operation.

[0132] Specifically, remote sensing images are acquired based on earth observation equipment, and image domain feature extraction is performed on the remote sensing images based on an image domain encoder to obtain image domain features. In order to extract the forgery pattern in the image domain, a convolutional layer with good spatial feature extraction capability is introduced to learn and characterize the geometric, structural features and local texture features of the remote sensing images. It is worth noting that the features extracted by the shallow convolutional network are more generalized in the authenticity identification of images of different data. The main reason behind this is that high-level semantic features are usually not affected by local features because they are too abstract, and are shared in true and false images, which is not conducive to the characterization of subtle traces of forgery in the image. Inspired by this, a simple single-layer convolution is used to construct an image domain encoder to capture finer image domain features and enhance the separability of true and false images. The expression of the image domain encoder is:

[0133] ,

[0134] Among them, f is the output of the image domain encoder, x is the input remote sensing image, is a convolution operation, BN is a batch normalization operation, ReLU is an activation function, and Maxpool is a maximum pooling operation.

[0135] The frequency domain feature extraction module 22 is used to perform frequency domain feature extraction processing on the remote sensing image based on a frequency domain encoder to obtain frequency domain features;

[0136] In the specific implementation process of the present invention, the frequency domain encoder includes an image frequency domain representation module, a high-frequency feature extraction module and a frequency convolution layer.

[0137] Furthermore, the image frequency domain representation module is used to perform medium and low frequency information filtering processing on the remote sensing image, wherein the process of medium and low frequency information filtering processing is:

[0138] ,

[0139] in, is the remote sensing image after filtering the medium and low frequency information, iFFT() is the inverse fast Fourier transform, B() is the high pass filter, FFT() is the fast Fourier transform, x is the input remote sensing image,

[0140] The operation rule of the high-pass filter is:

[0141] ,

[0142] Among them, B() is a high-pass filter, It is the element in the i-th row and j-th column of the input feature x in the remote sensing image.

[0143] Furthermore, the high-frequency feature extraction module is used to perform high-frequency enhancement on the remote sensing image after the medium and low frequency information filtering process to obtain the remote sensing image after high-frequency enhancement, wherein the process of high-frequency enhancement is:

[0144] ,

[0145] in, is the remote sensing image after high frequency enhancement. To perform the inverse fast Fourier transform operation across space, B() is a high-pass filter, To perform a fast Fourier transform operation across space, is the inverse fast Fourier transform operation across the channel dimension, is the fast Fourier transform operation across channel dimensions, f is the input feature, For high frequency enhancement in the spatial dimension, It is the high frequency enhancement in the channel dimension.

[0146] Furthermore, the frequency convolution layer is used to perform feature analysis on the high-frequency enhanced remote sensing image to obtain frequency domain features, wherein the feature analysis process is:

[0147] ,

[0148] ,

[0149] ,

[0150] in, is the frequency domain feature, is the fast Fourier transform, f is the input feature, is the convolutional layer applied to the phase spectrum, is the convolutional layer applied to the amplitude spectrum, is the frequency domain feature after spectrum decomposition, is the activation function, is the frequency domain feature, is the inverse fast Fourier transform operation.

[0151] Specifically, high-frequency information is crucial for detecting counterfeit images obtained by most generation methods, but in the original image domain, different frequency information is coupled with each other, resulting in the high-frequency components that can provide key information for authenticity detection being masked and difficult to fully mine. In this regard, a frequency domain encoder is designed. First, frequency decoupling and medium and low frequency filtering are performed at different levels such as images and features to fully extract the high-frequency information of the image. Then a special frequency feature convolution module is designed to directly perform deep feature mining on the phase spectrum and amplitude spectrum of high-frequency information. Frequency domain feature extraction processing is performed on the remote sensing image based on the frequency domain encoder. The frequency domain encoder includes an image frequency domain representation module, a high-frequency feature extraction module and a frequency convolution layer. The image frequency domain representation module is used to filter the medium and low frequency information of the remote sensing image to highlight the high-frequency information that is more critical to identifying traces of counterfeiting. The process of medium and low frequency information filtering processing is as follows:

[0152] ,

[0153] in, is the remote sensing image after filtering the medium and low frequency information, iFFT() is the inverse fast Fourier transform, B() is the high pass filter, FFT() is the fast Fourier transform, x is the input remote sensing image, the fast Fourier transform is the conversion of the remote sensing image from the spatial domain to the frequency domain, and the zero frequency component is moved to the center, the inverse fast Fourier transform is the conversion of the extracted high frequency information back to the spatial domain, and the high pass filter is used to extract high frequency information from the features.

[0154] The operation rule of the high-pass filter is:

[0155] ,

[0156] Among them, B() is a high-pass filter, It is the element in the i-th row and j-th column of the input feature x in the remote sensing image.

[0157] The high-frequency feature extraction module is used to perform high-frequency enhancement on the remote sensing image after filtering the medium and low frequency information, and obtain the remote sensing image after high-frequency enhancement. The frequency patterns of forged images by different methods are often different, which may cause the model to overfit only the frequency pattern of the current training data, thereby damaging the generalization ability. In this regard, the high-frequency feature extraction module forces the frequency encoder to give priority to the high-frequency information in the feature space to alleviate the tendency of the model to overfit the frequency pattern unique to the current data, that is, high-frequency enhancement is performed in the spatial dimension and the channel dimension respectively, wherein the process of high-frequency enhancement is:

[0158] ,

[0159] in, is the remote sensing image after high frequency enhancement. To perform the inverse fast Fourier transform operation across space, B() is a high-pass filter, To perform a fast Fourier transform operation across space, is the inverse fast Fourier transform operation across the channel dimension, is the fast Fourier transform operation across channel dimensions, f is the input feature, For high frequency enhancement in the spatial dimension, It is a high-frequency enhancement in the channel dimension. The high-frequency feature extraction module enables the high-frequency information of the intermediate features to be further purified and enhanced, thus better supporting the identification of forgery traces.

[0160] The frequency convolution layer is used to perform feature analysis on the high-frequency enhanced remote sensing image to obtain frequency domain features. The frequency convolution layer directly performs feature analysis on the phase and amplitude spectrum of the high-frequency component of the image obtained by the above module, thereby avoiding overfitting of insignificant low-frequency information in the data. The feature analysis process is as follows:

[0161] ,

[0162] ,

[0163] ,

[0164] in, is the frequency domain feature, is the fast Fourier transform, f is the input feature, is the convolutional layer applied to the phase spectrum, is the convolutional layer applied to the amplitude spectrum, is the frequency domain feature after spectrum decomposition, is the activation function, is the frequency domain feature, It is the inverse fast Fourier transform operation; the fast Fourier transform converts the input features into the frequency domain space, so that the obtained frequency domain features are complex tensors, which can be further decomposed into phase and amplitude spectra. Taking into account the fundamental difference between the two features in the physical sense, two convolutional layers with the same structure but non-shared parameters are set and They are applied to the phase spectrum and amplitude spectrum respectively to fully learn the frequency domain features. Finally, the learned frequency information is converted back to the standard feature space using the inverse fast Fourier transform to facilitate subsequent fusion and complementation with the image domain features.

[0165] A feature preliminary fusion module 23 is used to perform preliminary fusion of the image domain features and the frequency domain features based on a multi-domain representation unified mapping module to obtain preliminary fusion features;

[0166] In the specific implementation process of the present invention, the unified mapping module based on multi-domain representation performs preliminary fusion of the image domain features and the frequency domain features to obtain preliminary fusion features, including: setting shortcut connections across projection units, performing feature mapping on the image domain features and the frequency domain features based on shortcut connections across projection units and a number of feature projection units to obtain preliminary fusion features, wherein each feature projection unit is generated by stacking a number of 3*3 convolutional layers and 1*1 convolutional layers, and each convolutional layer includes a nonlinear activation function.

[0167] Specifically, in order to give full play to the complementary role of image domain and frequency domain features, a multi-domain representation unified mapping module is set up. It aims to overcome the significant differences between image domain and frequency domain features, and achieve the initial fusion of the two by uniformly mapping multi-domain features to a common feature subspace. The differences in content and physical meaning between image domain and frequency domain features are also the basis for their complementarity. If the two features are simply strictly aligned and fused, it will inevitably destroy their unique key information, and ultimately lose the meaning of fusion due to the damage to complementarity. Therefore, how to ensure that common ground is sought while reserving differences is the core issue for achieving effective fusion of multi-domain features. In this regard, the design process of the multi-domain representation unified mapping module mainly follows the following two principles. The first principle is a higher degree of nonlinearity to enhance the ability of feature transformation. In the multi-domain representation unified mapping module, a total of 3 feature projection units are set up, each of which is composed of several 3*3 and 1*1 convolutions stacked to achieve complex feature mapping. In addition, a nonlinear activation function ReLU is introduced between the convolutional layers to increase the overall nonlinear feature transformation capability. The second principle is to provide quick connections to ensure that key information is fully preserved: to avoid the loss of key information in complex feature transformations, quick connections across projection units are set up so that some underlying features have the opportunity to be fully preserved. Based on the above two principles, two branches with the same structure but non-shared parameters are set for image domain and frequency domain features.

[0168] Feature optimization module 24: used to optimize the preliminary fusion feature based on the hybrid domain feature optimization module to obtain the target fusion feature;

[0169] In the specific implementation process of the present invention, the hybrid domain feature optimization module is used to optimize the preliminary fusion features to obtain target fusion features, including: performing spatial information aggregation on the preliminary fusion features based on average pooling and maximum pooling operations to obtain a spatial context descriptor; generating a channel attention map based on a shared network using the spatial context descriptor; performing channel attention processing on the preliminary fusion features based on the channel attention map to obtain preliminary fusion features after channel attention processing, wherein the process of channel attention processing is:

[0170] ,

[0171] in, is the result of channel attention processing, is the channel attention mechanism operation, f is the initial fusion feature, It is a pixel-level multiplication operation;

[0172] Based on the pooling operation, the channel information of the preliminary fusion features after the channel attention processing is summarized to obtain a two-dimensional graph; based on the convolution layer, the two-dimensional graph is concatenated and convolved to obtain a two-dimensional spatial attention graph; based on the two-dimensional spatial attention graph, the preliminary fusion features after the channel attention processing are subjected to spatial attention processing to obtain target fusion features, wherein the process of spatial attention processing is:

[0173] ,

[0174] in, is the result of spatial attention processing, is the result of channel attention processing, Operates for the spatial attention mechanism.

[0175] Furthermore, the channel attention mechanism operates as follows:

[0176] ,

[0177] in, is the channel attention mechanism operation, is the sigmoid function, is a multi-layer perceptron, is the average pooling operation, f is the initial fusion feature, is the maximum pooling operation, and are the learnable weight parameters of the multilayer perceptron, is the average pooling feature, is the maximum pooling feature.

[0178] The spatial attention mechanism operates as follows:

[0179] ,

[0180] in, For the spatial attention mechanism operation, is the sigmoid function, is a convolution operation with a filter size of 7×7, is the average pooling operation, f is the initial fusion feature, is the maximum pooling operation, is the average pooling feature, is the maximum pooling feature.

[0181] Specifically, the hybrid domain feature optimization module is constructed by introducing the attention mechanism on the basis of the standard residual convolutional network. On the one hand, the low-level image and frequency features extracted by the encoder are further combined and optimized, and on the other hand, the model learns to adaptively focus on the more critical features as clues. After the multi-domain representation unified mapping module, the image domain and spatial domain features can be preliminarily aligned and added together, but how to give full play to their complementary role is still quite challenging. Considering the complexity of remote sensing images themselves, the diversity and concealment of forgery traces, the underlying visual and frequency features extracted by the encoder are rich but not always effective, and there may even be certain contradictions between them. In this regard, in addition to integrating and optimizing these basic features, another important purpose of the hybrid domain feature optimization module is to achieve adaptive attention to the key features. First, the backbone of the hybrid domain feature optimization model is constructed with multi-layer residual convolution. On this basis, a cross-space and cross-channel dual attention mechanism is introduced to selectively emphasize important features in different dimensions to improve the accuracy of authenticity identification. Based on the average pooling and maximum pooling operations, the spatial information of the preliminary fusion features is aggregated to obtain a spatial context descriptor, that is, two different spatial context descriptors are generated, namely, an average pooling feature descriptor and a maximum pooling feature descriptor. Based on the shared network, the spatial context descriptors are used to generate a channel attention map, and the two spatial context descriptors are transmitted to the shared network constructed by the multi-layer perceptron to generate a channel attention map; based on the channel attention map, the preliminary fusion features are subjected to channel attention processing to obtain preliminary fusion features after channel attention processing, wherein the process of channel attention processing is:

[0182] ,

[0183] in, is the result of channel attention processing, is the channel attention mechanism operation, f is the initial fusion feature, is a pixel-level multiplication operation. The channel attention mechanism operation is:

[0184] ,

[0185] in, is the channel attention mechanism operation, is the sigmoid function, is a multi-layer perceptron, is the average pooling operation, f is the initial fusion feature, is the maximum pooling operation, and are the learnable weight parameters of the multilayer perceptron, is the average pooling feature, is the maximum pooling feature. Based on the pooling operation, the channel information of the preliminary fusion features after the channel attention processing is summarized, that is, the channel information of the features is summarized through the maximum pooling operation and the average pooling operation to obtain a two-dimensional graph, that is, a maximum pooling two-dimensional graph and an average pooling two-dimensional graph; based on the convolution layer, the two-dimensional graph is concatenated and convolved to obtain a two-dimensional spatial attention graph; based on the two-dimensional spatial attention graph, the preliminary fusion features after the channel attention processing are subjected to spatial attention processing to obtain the target fusion feature, wherein the process of spatial attention processing is:

[0186] ,

[0187] in, is the result of spatial attention processing, is the result of channel attention processing, is the spatial attention mechanism operation. The spatial attention mechanism operation is:

[0188] ,

[0189] in, For the spatial attention mechanism operation, is the sigmoid function, is a convolution operation with a filter size of 7×7, is the average pooling operation, f is the initial fusion feature, is the maximum pooling operation, is the average pooling feature, is the maximum pooling feature.

[0190] Image authenticity identification module 25: used for performing authenticity identification processing of remote sensing images based on the target fusion features to obtain authenticity identification results.

[0191] In the specific implementation process of the present invention, the target fusion features are input into a simple linear classifier to perform authenticity identification processing on the remote sensing image to obtain an authenticity identification result.

[0192] In the embodiment of the present invention, taking into account the differences between the image domain and the frequency domain and the key significance of high-frequency features for the authenticity identification of images, the remote sensing image is subjected to frequency domain feature extraction processing based on the frequency domain encoder to more accurately capture the forgery traces of the false image in the frequency domain. The image domain features and the frequency domain features are initially fused based on the multi-domain representation unified mapping module, and then the initial fusion features are optimized based on the hybrid domain feature optimization module. This can achieve deep fusion of features to address the problem of hindering the synergistic complementation of differentiated image domain and spatial features due to the lack of an effective feature fusion mechanism, achieve stronger generalization and robustness by capturing richer and more comprehensive features, achieve more advanced remote sensing image authenticity identification accuracy, and capture generalization features about different false images.

[0193] A person skilled in the art may understand that all or part of the steps in the various methods of the above embodiments may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the storage medium may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0194] In addition, the above is a detailed introduction to a remote sensing image authenticity identification method and system provided by an embodiment of the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for a person skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A remote sensing image authenticity identification method, characterized in that: The method comprises: Acquire a remote sensing image, and perform image domain feature extraction processing on the remote sensing image based on an image domain encoder to obtain image domain features; Performing frequency domain feature extraction processing on the remote sensing image based on a frequency domain encoder to obtain frequency domain features; Based on a multi-domain representation unified mapping module, the image domain features and the frequency domain features are preliminarily fused to obtain a preliminary fusion feature; Optimizing the preliminary fusion features based on a hybrid domain feature optimization module to obtain target fusion features; Performing authenticity identification processing on the remote sensing image based on the target fusion feature to obtain an authenticity identification result; Among them, the expression of the image domain encoder is: Among them, f is the output of the image domain encoder, x is the input remote sensing image, is a convolution operation, BN is a batch normalization operation, ReLU is an activation function, and Maxpool is a maximum pooling operation; The hybrid domain feature optimization module is used to optimize the preliminary fusion feature to obtain the target fusion feature, including: Aggregating spatial information of the preliminary fusion features based on average pooling and maximum pooling operations to obtain a spatial context descriptor; Generate a channel attention map using the spatial context descriptor based on a shared network; Based on the channel attention map, channel attention processing is performed on the preliminary fusion feature to obtain the preliminary fusion feature after the channel attention processing, wherein the process of the channel attention processing is: in, is the result of channel attention processing, is the channel attention mechanism operation, f is the initial fusion feature, It is a pixel-level multiplication operation; Based on the pooling operation, the channel information of the preliminary fusion features after the channel attention processing is summarized to obtain a two-dimensional graph; Based on the convolution layer, the two-dimensional graph is concatenated and convolved to obtain a two-dimensional spatial attention graph; Based on the two-dimensional spatial attention map, spatial attention processing is performed on the preliminary fusion features after channel attention processing to obtain target fusion features, wherein the process of spatial attention processing is: in, is the result of spatial attention processing, is the result of channel attention processing, Operates for the spatial attention mechanism.

2. The remote sensing image authenticity identification method according to claim 1, characterized in that: The frequency domain encoder includes an image frequency domain representation module, a high frequency feature extraction module and a frequency convolution layer.

3. The remote sensing image authenticity identification method according to claim 2, characterized in that: The image frequency domain representation module is used to perform medium and low frequency information filtering processing on the remote sensing image, wherein the process of medium and low frequency information filtering processing is: in, is the remote sensing image after filtering the medium and low frequency information, iFFT() is the inverse fast Fourier transform, B() is the high pass filter, FFT() is the fast Fourier transform, x is the input remote sensing image, The operation rule of the high-pass filter is: Among them, B() is a high-pass filter, It is the element in the i-th row and j-th column of the input feature x in the remote sensing image.

4. The remote sensing image authenticity identification method according to claim 2, characterized in that: The high-frequency feature extraction module is used to perform high-frequency enhancement on the remote sensing image after the medium and low frequency information filtering process to obtain the remote sensing image after high-frequency enhancement, wherein the process of high-frequency enhancement is: in, is the remote sensing image after high frequency enhancement. To perform the inverse fast Fourier transform operation across space, B() is a high-pass filter, To perform a fast Fourier transform operation across space, is the inverse fast Fourier transform operation across the channel dimension, is a fast Fourier transform operation across channel dimensions, f is the input feature, HFRF-S is the high-frequency enhancement in the spatial dimension, and HFRF-C is the high-frequency enhancement in the channel dimension.

5. The remote sensing image authenticity identification method according to claim 2, characterized in that: The frequency convolution layer is used to perform feature analysis on the high-frequency enhanced remote sensing image to obtain frequency domain features, wherein the feature analysis process is: in, is the frequency domain feature, FFT() is the fast Fourier transform, f is the input feature, is the convolutional layer applied to the phase spectrum, is the convolutional layer applied to the amplitude spectrum, is the frequency domain feature after spectrum decomposition, ReLU is the activation function, It is the frequency domain feature, and iFFT() is the inverse fast Fourier transform operation.

6. The remote sensing image authenticity identification method according to claim 1, characterized in that: The multi-domain representation-based unified mapping module performs preliminary fusion of the image domain features and the frequency domain features to obtain preliminary fusion features, including: A shortcut connection across projection units is set, and feature mapping is performed on the image domain features and the frequency domain features based on the shortcut connection across projection units and a plurality of feature projection units to obtain preliminary fusion features, wherein each feature projection unit is generated by stacking a plurality of 3*3 convolutional layers and 1*1 convolutional layers, and each convolutional layer includes a nonlinear activation function.

7. The remote sensing image authenticity identification method according to claim 1, characterized in that: The channel attention mechanism operates as follows: , in, is the channel attention mechanism operation, is the sigmoid function, MLP is the multi-layer perceptron, AvgPool is the average pooling operation, f is the initial fusion feature, MaxPool is the maximum pooling operation, and are the learnable weight parameters of the multilayer perceptron, is the average pooling feature, is the maximum pooling feature; The spatial attention mechanism operates as follows: , in, For the spatial attention mechanism operation, is the sigmoid function, is a convolution operation with a filter size of 7×7, AvgPool is an average pooling operation, f is a preliminary fusion feature, MaxPool is a maximum pooling operation, is the average pooling feature, is the maximum pooling feature.

8. A remote sensing image authenticity identification system, characterized in that: The system comprises: Image domain feature extraction module: used to obtain remote sensing images, perform image domain feature extraction processing on the remote sensing images based on the image domain encoder, and obtain image domain features; Frequency domain feature extraction module: used for performing frequency domain feature extraction processing on the remote sensing image based on a frequency domain encoder to obtain frequency domain features; A feature preliminary fusion module: used to perform preliminary fusion of the image domain features and the frequency domain features based on a multi-domain representation unified mapping module to obtain preliminary fusion features; Feature optimization module: used to optimize the preliminary fusion features based on the hybrid domain feature optimization module to obtain target fusion features; Image authenticity identification module: used to perform authenticity identification processing of remote sensing images based on the target fusion features to obtain authenticity identification results; Among them, the expression of the image domain encoder is: Among them, f is the output of the image domain encoder, x is the input remote sensing image, is a convolution operation, BN is a batch normalization operation, ReLU is an activation function, and Maxpool is a maximum pooling operation; The hybrid domain feature optimization module is used to optimize the preliminary fusion feature to obtain the target fusion feature, including: Aggregating spatial information of the preliminary fusion features based on average pooling and maximum pooling operations to obtain a spatial context descriptor; Generate a channel attention map using the spatial context descriptor based on a shared network; Based on the channel attention map, channel attention processing is performed on the preliminary fusion feature to obtain the preliminary fusion feature after the channel attention processing, wherein the process of the channel attention processing is: in, is the result of channel attention processing, is the channel attention mechanism operation, f is the initial fusion feature, It is a pixel-level multiplication operation; Based on the pooling operation, the channel information of the preliminary fusion features after the channel attention processing is summarized to obtain a two-dimensional graph; Based on the convolution layer, the two-dimensional graph is concatenated and convolved to obtain a two-dimensional spatial attention graph; Based on the two-dimensional spatial attention map, spatial attention processing is performed on the preliminary fusion features after channel attention processing to obtain target fusion features, wherein the process of spatial attention processing is: in, is the result of spatial attention processing, is the result of channel attention processing, Operates for the spatial attention mechanism.