Remote sensing image copyright protection method based on detail features

Abstract

The application discloses a remote sensing image copyright protection method based on detail features, and specifically comprises the following steps: step 1, a convolutional neural network ContourResUNet is designed and trained, and a binary detail feature map of a pre-protected remote sensing image is output; step 2, a binary watermark image is obtained; step 3, a zero watermark image of the protected remote sensing image is generated; step 4, a remote sensing image to be verified for copyright is input into the trained ContourResUNet network, and a binary detail feature map of the remote sensing image to be verified for copyright is output; step 5, the obtained binary detail feature map is logically unbonded with the zero watermark image, and a watermark image is obtained; and step 6, similarity value calculation is performed on the obtained watermark image and an original watermark image, and the copyright ownership of the remote sensing image is determined. The method solves the problems of false detection and invalidation of copyright watermark information caused by poor image feature recognition ability in the process of remote sensing image copyright.

Description

Technical Field

[0001] This invention belongs to the field of image information security and copyright protection technology, specifically relating to a method for copyright protection of remote sensing images based on detailed features. Background Technology

[0002] Remote sensing images can display various features of the Earth's surface, such as water bodies, land, farmland, towns, roads, and buildings, providing crucial data and scientific basis for military and civilian research on the Earth's surface. However, with the rapid development of multimedia and network technologies, any user can easily access remote sensing images, making them vulnerable to infringement. Protecting the intellectual property rights of remote sensing images and ensuring the legitimate rights and interests of research institutions and businesses has become essential.

[0003] As an effective copyright protection solution, the zero-watermark method does not embed watermark information into the carrier image; instead, it binds copyright information to robust features of the carrier image. This technique maintains the integrity of remote sensing image data, making the zero-watermark method particularly suitable for copyright protection of remote sensing images.

[0004] However, remote sensing images from the same modality often look very similar. The limited discriminative power of remote sensing image features makes it easy for watermark false positives to occur during copyright extraction. This poses a significant challenge to the robustness, effectiveness, and efficiency of existing zero-watermarking algorithms in protecting image copyright. Summary of the Invention

[0005] The purpose of this invention is to provide a method for copyright protection of remote sensing images based on detailed features. This method uses the binding of detailed features of remote sensing images with copyright information to construct a zero watermark, which solves the problem of false detection and failure of copyright watermark information caused by poor image feature recognition ability in the process of copyright verification of remote sensing images, thereby significantly improving the effectiveness and robustness of the algorithm.

[0006] The technical solution adopted in this invention is a remote sensing image copyright protection method based on detailed features, specifically as follows: Step 1: Design and train the ContourResUNet convolutional neural network to output a binary detail feature map of the remote sensing image to be protected. Step 2: Load the timestamp onto the original watermark image and binarize it to obtain a binary watermark image; Step 3: Generate a watermark-free image of the protected remote sensing image; Step 4: Input the remote sensing image of the copyright to be verified into the trained ContourResUNet network, and output the binary detail feature map of the remote sensing image of the copyright to be verified. Step 5: Logically unbind the obtained binary detail feature map from the zero-watermark image to obtain the watermark image; Step 6: Calculate the similarity value between the obtained watermark image and the original watermark image. ) Calculate and determine the copyright ownership of remote sensing images.

[0007] The invention is further characterized in that: Step 1 is as follows: Step 1.1: Obtain the ground truth labels learned by the ContourResUNet convolutional neural network; Step 1.1 specifically involves: Step 1.1.1: Read the remote sensing image to be protected, convert the remote sensing image to grayscale, and adjust the image size. size; Step 1.1.2: Daubechies-4 (db4) Discrete Wavelet Transform (DWT) is used to perform a two-level decomposition on the remote sensing image processed in Step 1.1.1, resulting in a low-frequency approximate sub-band and three high-frequency detail sub-bands in three directions. The discrete wavelet transform is shown in Equation (1): (1) In the formula, Represents remote sensing images; , ; Represents low-frequency approximate subbands, Represents horizontal high-frequency detail subbands, Represents vertical high-frequency detail subbands, Represents the diagonal high-frequency detail subband. ; Step 1.1.3: Perform detail feature fusion on the three high-frequency detail subbands obtained in step 1.1.2 through two-level discrete wavelet transform, as shown in equation (2), to obtain the detail feature map. : (2) In the formula, This represents the detail feature map obtained after fusing high-frequency detail subbands; H 2 represents the horizontal high-frequency detail subband obtained from the two-level wavelet decomposition. V 2 represents the vertical high-frequency detail subband obtained from the two-level wavelet decomposition. D 2 represents the diagonal high-frequency detail subband obtained from the two-level wavelet decomposition; because the three high-frequency detail subbands obtained from the first-level decomposition (horizontal high-frequency detail subbands) Vertical high-frequency detail subband and diagonal high-frequency detail subband Since the image contains more noise, we select three high-frequency detail sub-bands from the two-level wavelet decomposition for feature fusion to obtain the detail feature map of the image.

[0008] Step 1.1.4: Using the inter-class variance maximization criterion (Otsu's method), calculate the detail feature map obtained in step 1.1.3. Optimal segmentation threshold ; Step 1.1.5, based on the threshold obtained in step 1.1.4 For the detailed feature map obtained in step 1.1.3 Binarization is performed according to equation (3) to obtain a binary detail feature map, as shown in equation (3) below: (3) In the formula, Represents a binary detail feature map; where, This represents the background area, which corresponds to a flat area in the remote sensing image that has no obvious structure or abrupt changes. This indicates a region of significant detail features, corresponding to areas in a remote sensing image where there are significant gradient changes, i.e., regions of texture detail in the image.

[0009] Step 1.1.6: Use the binary detail feature map obtained in step 1.1.5 as the ground truth label for learning the convolutional neural network ContourResUNet; Step 1.2: Construct the ContourResUNet convolutional neural network structure; In step 1.2, the constructed convolutional neural network ContourResUNet consists of an encoder module, a bottleneck module, a decoder module, and an output layer; the output layer consists of a 1×1 convolutional layer (Conv) and a sigmoid activation function; Step 1.2 specifically involves: Step 1.2.1: Build the encoder module and extract shallow features from the detail feature map; In step 1.2.1, the encoder module consists of a first-level convolutional coding unit (Encoder1), a second-level convolutional coding unit (Encoder2), a third-level convolutional coding unit (Encoder3), and a fourth-level convolutional coding unit (Encoder4) connected in series. Each convolutional coding unit employs a dual convolutional structure, specifically consisting of two 3×3 convolutional (Conv) operations. Each convolutional layer is followed by batch normalization and ReLU activation functions to enhance feature representation and improve network training stability. To progressively expand the receptive field and extract higher-level spatial semantic information, adjacent convolutional coding units undergo spatial downsampling through 2×2 max-pooling, halving the feature map size at each level. During the encoding process, the number of output channels of the four-level convolutional coding units are set to 64, 128, 256 and 512 respectively, and the corresponding feature map spatial resolutions are N / 8×N / 8, N / 16×N / 16, N / 32×N / 32 and N / 64×N / 64 respectively, thus forming a multi-scale feature hierarchy structure from shallow to deep; Step 1.2.2: Build the bottleneck module between the encoder module and the decoder module, and extract deep features from the detail feature map; In step 1.2.2, the bottleneck module adopts a dual 3×3 convolution (Conv) structure, and uses batch normalization and ReLU activation function to perform deep fusion of deep features; the N / 64×N / 64×512 feature map output by the fourth level convolutional coding unit (Encoder4) of the encoder enters the bottleneck module, and through two layers of 3×3 convolution (Conv) operation, its output channel number is increased to 1024, resulting in a high-dimensional semantic feature map with size N / 64×N / 64×1024; This module can effectively capture the global structural information and complex texture relationships of the detailed feature maps of remote sensing images, providing sufficient semantic support for feature reconstruction in the subsequent decoding stage.

[0010] Step 1.2.3: Construct the decoder module and output the binary detail feature map; In step 1.2.3, the decoder module and the encoder module have four symmetrical convolutional layers. The decoder module consists of a first-level convolutional decoding unit (Decoder1), a second-level convolutional decoding unit (Decoder2), a third-level convolutional decoding unit (Decoder3), and a fourth-level convolutional decoding unit (Decoder4) connected in series. Each level of convolutional decoding unit adopts a double convolutional structure, specifically including two layers of 3x3 convolution (Conv) operations. After each convolutional layer, batch normalization and ReLU activation functions are sequentially connected, and combined with 2×2 deconvolution (UpConv) and channel concatenation skip connection (Concat) to achieve deep fusion of multi-scale features. This design aims to reconstruct high-resolution feature maps with rich details, ensure the integrity of detailed information and spatial accuracy, and provide robust and discriminative feature representations for zero-watermark generation and copyright verification.

[0011] During the decoding process: In the initial decoding stage, the first-level convolutional decoding unit (Decoder1) first performs a 2×2 deconvolution upsampling (UpConv) operation on the N / 64×N / 64×1024 feature map output by the bottleneck module, increasing the feature map size to N / 32×N / 32 and reducing the number of channels to 512. This is then concatenated with the N / 32×N / 32×512 feature map output by the fourth-level convolutional encoding unit (Encoder4) of the encoder along the channel dimension to obtain fused features, which are then integrated by a dual convolutional module. Subsequently, in the processing stages of the second-level convolutional decoding unit (Decoder2) and the third-level convolutional decoding unit (Decoder3), the decoder module gradually restores the spatial resolution through multiple levels of 2×2 deconvolution upsampling operations. Specifically, in the processing stage of the second-level convolutional decoding unit (Decoder2), the feature map is changed from N / 32×N / 32×512... The feature map is upsampled to N / 16×N / 16×256 and concatenated with the corresponding features of the third-level convolutional coding unit (Encoder3) of the encoder module. In the third-level convolutional decoding unit (Decoder3) processing stage, the feature map is further upsampled from N / 16×N / 16×256 to N / 8×N / 8×128 and fused with the output features of the second-level convolutional coding unit (Encoder2) of the encoder module. In the fourth-level convolutional decoding unit (Decoder4) processing stage, the feature map is further upsampled to N / 4×N / 4×64 and fused with the output features of the first-level convolutional coding unit (Encoder1) of the encoder module. In each decoding stage, the fused features are channel compressed and refined through a dual 3×3 convolutional structure, thereby restoring spatial resolution while suppressing redundant information and enhancing effective structural expression.

[0012] The final feature map output by the decoder module is compressed to 1 channel by a 1×1 convolutional layer in the output layer, generating a map of size [size missing]. Binary detail feature map; Step 1.3: Construct the dataset T for the ContourResUNet network; Step 1.3 specifically involves: Step 1.3.1: Select an image from the remote sensing image dataset RSOD-Datas as the remote sensing image to be protected; Step 1.3.2: Resize the remote sensing image selected in Step 1.3.1. Standard-sized images; Step 1.3.3 involves performing various levels of conventional image processing operations on the image obtained in Step 1.3.2, specifically including salt-and-pepper noise reduction, Gaussian noise reduction, compression, sharpening, smoothing, and quantization. It also includes various levels of geometric attack processing, including rotation, translation, and cropping. All images obtained from these conventional image processing operations and combinations thereof total [number missing]. The dataset T that constitutes the ContourResUNet network; Step 1.3.4, based on the results obtained in step 1.3.3 Zhang image, One set of images is used as the training set for the ContourResUNet network, and the remaining images are used as the test set for the ContourResUNet network. Step 1.4: Design the loss function for the ContourResUNet network; Step 1.4 specifically involves: Step 1.4.1: Obtain the ground truth labels of the ContourResUNet network based on Step 1.1. ; Step 1.4.2, calculate the loss function of the ContourResUNet network using equation (4): (4) In the formula, and The binary detail feature map obtained in step 1.1.5 and the image predicted by the ContourResUNet network at location ( x , y The pixel value of ); Since the dataset contains carrier images that have suffered almost all types of attacks, the trained ContourResUNet network can extract images from the attacked carrier images that are highly similar to the binary detail feature images obtained from the original carrier images using wavelet transform, thus ensuring the strong robustness of the proposed method.

[0013] Step 1.5, train the network; Step 1.5 specifically involves: Step 1.5.1, Optimizer and learning rate configuration; Specifically: Training uses the Adam optimizer, with an initial learning rate set to 5×10. - ³; Simultaneously, the ReduceLROnPlateau learning rate scheduling strategy is introduced. If the average loss does not decrease over 5 consecutive epochs, the learning rate is multiplied by a coefficient of 0.5 for decay, with a lower limit of 1×10⁻⁶. -6 To promote smooth convergence in the later stages of training; Step 1.5.2: Calculate the loss function and gradient processing; Specifically, in each batch, the input training image is forward propagated to obtain predicted binary detail features, and then the loss function is calculated; during backpropagation, the gradient is clipped to prevent gradient explosion and the network parameters are updated accordingly.

[0014] Step 1.5.3, Training process control; Specifically: after each epoch, the average loss is recorded and the best performing network weights are automatically saved; the training termination condition is set as follows: the loss change is less than 0.001, or the maximum number of training epochs of 150 is reached; Step 1.5.4: Training ends, and the ContourResUNet network outputs a binary detail feature map of the protected remote sensing image.

[0015] Step 2 is as follows: Step 2.1, adjust the size of the original watermark image to... Size, and load the zero-watermark image into the watermark image at the expected time of registration with the copyright center; Step 2.2: If the original watermark image is a color image, use formula (5) to convert the color watermark image into a grayscale image; (5) In the formula, Represents a watermarked image Pixel value at that location, Indicates the red channel of the color watermark image. Pixel value at that location, Indicates the green channel of the color watermark image Pixel value at that location, Indicates the blue channel of the color watermark image. Pixel value at; Step 2.3: Obtain the threshold by calculating the inter-class variance of the original watermark image. Each pixel value of the original watermark image is binarized using formula (6) to obtain a binary watermark image; (6) In the formula, This is a binary watermarked image after binarization.

[0016] Step 3 specifically involves: Step 3.1: Input the remote sensing image to be protected into the ContourResUNet network trained in Step 1 to obtain the binary detail feature map of the image to be protected. Step 3.2: XOR the binary detail feature map obtained in Step 3.1 with the binary watermark image obtained in Step 2 to obtain the initial zero-watermark image. ; Step 3.3, process the initial zero-watermark image obtained in step 3.2. Perform a Fibonacci transform to obtain a watermark-free image.

[0017] Step 3.3 specifically involves: Step 3.3.1, Set the Fibonacci scrambling matrix M as shown in equation (7): (7) Step 3.3.2, take the initial zero-watermark image obtained in step 3.2. Each pixel coordinate in Perform transformation processing using equation (8) and iterate. Next, get ; (8); Step 3.3.3: Based on the coordinate relationship obtained in step 3.3.2, the zero-watermark image is finally obtained by equation (9): (9) In the formula, The image is the result of the coordinate transformation in step 3.3.2 from the initial zero-watermark image. This is the final watermark-free image.

[0018] Step 5 specifically involves: Step 5.1: Perform an inverse Fibonacci transform on the zero-watermarked image registered with the software copyright center to obtain the initial zero-watermarked image. ; Step 5.2, take the initial zero-watermark image obtained in step 5.1. XOR the binary detail feature map obtained in step 4 to extract the binary watermark. ; Step 5.1 specifically involves: Step 5.1.1, set the Fibonacci inverse scrambling matrix. As shown in equation (10): (10) Step 5.1.2, for each pixel coordinate in the zero-watermark image obtained from the software copyright center ( Perform the transformation process of equation (11) and iterate. Next, get ; (11) Step 5.1.3: Based on the coordinate relationship obtained in step 5.1.2, recover the initial zero-watermark image using equation (12). : (12).

[0019] Step 6 specifically involves: Step 6.1, take the watermark image obtained in step 5. Compared with the original watermark image when generating the zero-watermark image Similarity is calculated, and a similarity value is obtained according to formula (13). The expression is as follows: (13) In the formula, To obtain the watermarked image, The original watermark image used when generating the zero-watermark image; Step 6.2: Verify copyright ownership based on the similarity value obtained in Step 6.1. Value > 0.5 indicates that the remote sensing image was generated from the watermarked image used when the zero-watermark image was created. Owned by the owner.

[0020] The beneficial effects of this invention are: This invention presents a remote sensing image copyright protection method based on detailed features. Since the watermark copyright information in this method does not require modification of the remote sensing image data, it can maintain the integrity of the remote sensing image data, ensuring that the copyright-protected remote sensing image still has scientific research value in both military and civilian fields. Secondly, although the detailed features of remote sensing images themselves are not robust, the ContourResUNet network designed in this invention can obtain highly similar binary detailed feature images under various attacks. This makes the robustness index of the watermark copyright information obtained by this invention significantly higher than existing watermarking methods, demonstrating the strong robustness of this invention against various attacks. Furthermore, as a fundamental feature of an image, binary detailed features are effective identification identifiers for remote sensing images. This characteristic allows this invention to completely eliminate the phenomenon of detecting false copyright information from unauthorized remote sensing images. Moreover, the timestamp added to the watermark image can effectively prevent the protected remote sensing image from being infringed upon or re-bound with copyright information, thus preventing the failure of its legitimate copyright verification. Therefore, this invention can completely eliminate copyright false alarms. Overall, this invention provides a reliable copyright authentication scheme for the protection of remote sensing images. It has achieved very satisfactory results in terms of the transparency, robustness, and false alarm rate of watermark copyright information, and has high application and promotion value. Attached Figure Description

[0021] Figure 1This is an overall flowchart of the method of the present invention; Figure 2 This is a structural diagram of the ContourResUNet network in the method of this invention; Figure 3 These are remotely sensed images that have not been attacked; Figure 4 It utilizes the ContourResUNet network from Figure 3 Binary detail feature map extracted from remote sensing image; Figure 5 It is the watermark image used in the method of this invention; Figure 6 yes Figure 5 The watermarked image after adding a timestamp and binarizing it; Figure 7 It is by Figure 4 and Figure 6 The generated zero-watermark image; Figure 8 These are schematic diagrams of different attack types. In the diagram, a–f are schematic diagrams after JPEG compression, salt and pepper noise, Gaussian noise, rotation, scaling, and translation, respectively. Figure 9 Is Figure 8 The image shows a comparison of the copyright logo image recovered using the method of this invention under different attack conditions. In the figure, a–f correspond to schematic diagrams after six different attack types. Figure 10 It is by Figure 3 and Figure 5 A visualization of the process from generating a zero watermark to extracting the watermark; Figure 11 It is a series of visual results where a protected remote sensing image is protected by another watermarked image (i.e., illegal protection); Figure 12 This is the process of extracting illegal copyright marks from unprotected images, verifying the feasibility of this method in eliminating false watermarks; Figure 13 This is the set of remote sensing images used to verify and eliminate false watermarks, where a–e represent 5 remote sensing images respectively; Figure 14 This is the zero-watermark group used to verify and eliminate false watermarks, where 1–5 represent 5 zero-watermark images respectively. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0023] This invention provides a method for copyright protection of remote sensing images based on detailed features, including two processes: zero-watermark generation and watermark extraction (e.g., ...). Figure 1 (as shown) Zero-watermark generation process: First, design a neural network called ContourResUNet; the design includes network learning labels, network structure, dataset generation, loss function and network training; second, input the remote sensing image to be protected into the trained ContourResUNet network, and logically bind its output binary detail feature image with the loaded timestamp binary watermark image (copyright information) to generate a zero-watermark image. Copyright verification process: The remote sensing image whose copyright needs to be verified is input into a trained ContourResUNet network. The output of the trained ContourResUNet network is logically decoupled from the zero-watermark image registered in the copyright protection center to restore the watermark image. Finally, the NC value of the restored watermark image and the original watermark image is calculated. Based on the NC value, the copyright ownership of the image is verified.

[0024] Specifically, the following steps are included: Step 1: Design and train the ContourResUNe convolutional neural network to output a binary detail feature map of the image to be protected. Step 1 is as follows: Step 1.1: Obtain the ground truth labels learned by the ContourResUNet convolutional neural network; Step 1.1 specifically involves: Step 1.1.1: Read the remote sensing image to be protected, convert the remote sensing image to grayscale, and adjust the image size. size, 512; Step 1.1.2: Daubechies-4 (db4) Discrete Wavelet Transform (DWT) is used to perform a two-level decomposition on the remote sensing image processed in Step 1.1.1, resulting in a low-frequency approximate sub-band and three high-frequency detail sub-bands in three directions. The discrete wavelet transform is shown in Equation (1): (1) In the formula, Represents remote sensing images; , ; Represents low-frequency approximate subbands, Represents horizontal high-frequency detail subbands, Represents vertical high-frequency detail subbands, Represents the diagonal high-frequency detail subband. ; Step 1.1.3: Perform detail feature fusion on the three high-frequency detail subbands obtained in step 1.1.2 through two-level discrete wavelet transform, as shown in equation (2), to obtain the detail feature map. : (2) In the formula, This represents the detail feature map obtained after fusing high-frequency detail subbands; H 2 represents the horizontal high-frequency detail subband obtained from the two-level wavelet decomposition. V 2 represents the vertical high-frequency detail subband obtained from the two-level wavelet decomposition. D 2 represents the diagonal high-frequency detail subband obtained from the two-level wavelet decomposition; because the three high-frequency detail subbands obtained from the first-level decomposition (horizontal high-frequency detail subbands) Vertical high-frequency detail subband and diagonal high-frequency detail subband Since the image contains more noise, we select three high-frequency detail sub-bands from the two-level wavelet decomposition for feature fusion to obtain the detail feature map of the image.

[0025] Step 1.1.4: Using the inter-class variance maximization criterion (Otsu's method), calculate the detail feature map obtained in step 1.1.3. Optimal segmentation threshold , ; Step 1.1.5, based on the threshold obtained in step 1.1.4 For the detailed feature map obtained in step 1.1.3 Binarization is performed according to equation (3) to obtain a binary detail feature map, as shown in equation (3) below: (3) In the formula, Represents a binary detail feature map; where, This represents the background area, which corresponds to a flat area in the remote sensing image that has no obvious structure or abrupt changes. This indicates a region of significant detail features, corresponding to areas in a remote sensing image where there are significant gradient changes, i.e., regions of texture detail in the image.

[0026] Step 1.1.6: Use the binary detail feature map obtained in step 1.1.5 as the ground truth label for learning the convolutional neural network ContourResUNet; Step 1.2, construct the ContourResUNet convolutional neural network structure, as follows: Figure 2 As shown; In step 1.2, the constructed convolutional neural network ContourResUNet consists of an encoder module, a bottleneck module, a decoder module, and an output layer; the output layer consists of a 1×1 convolutional layer (Conv) and a sigmoid activation function; Step 1.2 specifically involves: Step 1.2.1: Build the encoder module and extract shallow features from the detail feature map; In step 1.2.1, the encoder module consists of a first-level convolutional coding unit (Encoder1), a second-level convolutional coding unit (Encoder2), a third-level convolutional coding unit (Encoder3), and a fourth-level convolutional coding unit (Encoder4) connected in series. Each convolutional coding unit employs a dual convolutional structure, specifically consisting of two 3×3 convolutional (Conv) operations. Each convolutional layer is followed by batch normalization and ReLU activation functions to enhance feature representation and improve network training stability. To progressively expand the receptive field and extract higher-level spatial semantic information, adjacent convolutional coding units undergo spatial downsampling through 2×2 max-pooling, halving the feature map size at each level. During the encoding process, the number of output channels of the four-level convolutional coding units are set to 64, 128, 256 and 512 respectively, and the corresponding feature map spatial resolutions are 64×64, 32×32, 16×16 and 8×8 respectively, thus forming a multi-scale feature hierarchy structure from shallow to deep. Step 1.2.2: Build the bottleneck module between the encoder module and the decoder module, and extract deep features from the detail feature map; In step 1.2.2, the bottleneck module adopts a dual 3×3 convolution (Conv) structure, and uses batch normalization and ReLU activation function to perform deep fusion of deep features; the 8×8×512 feature map output by the fourth level convolutional coding unit (Encoder4) of the encoder enters the bottleneck module, and through two layers of 3×3 convolution (Conv) operation, its output channel number is increased to 1024, resulting in a high-dimensional semantic feature map with a size of 8×8×1024; This module can effectively capture the global structural information and complex texture relationships of the detailed feature maps of remote sensing images, providing sufficient semantic support for feature reconstruction in the subsequent decoding stage.

[0027] Step 1.2.3: Construct the decoder module and output the binary detail feature map; In step 1.2.3, the decoder module and the encoder module have four symmetrical convolutional layers. The decoder module consists of a first-level convolutional decoding unit (Decoder1), a second-level convolutional decoding unit (Decoder2), a third-level convolutional decoding unit (Decoder3), and a fourth-level convolutional decoding unit (Decoder4) connected in series. Each level of convolutional decoding unit adopts a double convolutional structure, specifically including two layers of 3x3 convolution (Conv) operations. After each convolutional layer, batch normalization and ReLU activation functions are sequentially connected, and combined with 2×2 deconvolution (UpConv) and channel concatenation skip connection (Concat) to achieve deep fusion of multi-scale features. This design aims to reconstruct high-resolution feature maps with rich details, ensure the integrity of detailed information and spatial accuracy, and provide robust and discriminative feature representations for zero-watermark generation and copyright verification.

[0028] During the decoding process: In the initial decoding stage, the first-level convolutional decoding unit (Decoder1) first performs a 2×2 deconvolution upsampling (UpConv) operation on the 8×8×1024 feature map output by the bottleneck module, increasing the feature map size to 16×16 and reducing the number of channels to 512. This is then concatenated with the 16×16×512 feature map output by the fourth-level convolutional encoding unit (Encoder4) of the encoder along the channel dimension to obtain fused features, which are then integrated by the dual convolutional module. Subsequently, in the processing stages of the second-level convolutional decoding unit (Decoder2) and the third-level convolutional decoding unit (Decoder3), the decoder module gradually restores the spatial resolution through multiple levels of 2×2 deconvolution upsampling operations. Specifically, in the processing stage of the second-level convolutional decoding unit (Decoder2), the feature map is changed from 16×16×512... The feature map is upsampled to 32×32×256 and concatenated with the corresponding features of the third-level convolutional coding unit (Encoder3) of the encoder module. In the third-level convolutional decoding unit (Decoder3) processing stage, the feature map is further upsampled from 32×32×256 to 64×64×128 and fused with the output features of the second-level convolutional coding unit (Encoder2) of the encoder module. In the fourth-level convolutional decoding unit (Decoder4) processing stage, the feature map is further upsampled to 128×128×64 and fused with the output features of the first-level convolutional coding unit (Encoder1) of the encoder module. In each decoding stage, the fused features are channel compressed and refined through a dual 3×3 convolutional structure, thereby restoring spatial resolution while suppressing redundant information and enhancing effective structural expression.

[0029] The final feature map output by the decoder module is compressed to 1 channel by a 1×1 convolutional layer in the output layer, generating a map of size [size missing]. Binary detail feature map; Step 1.3: Construct the dataset T for the ContourResUNet network; Step 1.3 specifically involves: Step 1.3.1: Select an image from the remote sensing image dataset RSOD-Datas as the remote sensing image to be protected, such as... Figure 3 As shown, it is used as a carrier for zero watermarking; Step 1.3.2: Resize the remote sensing image selected in Step 1.3.1. Standard-sized images; Step 1.3.3 involves performing various levels of conventional image processing operations on the image obtained in Step 1.3.2, specifically including salt-and-pepper noise reduction, Gaussian noise reduction, compression, sharpening, smoothing, and quantization. It also includes various levels of geometric attack processing, including rotation, translation, and cropping. All images obtained from these conventional image processing operations and combinations thereof total [number missing]. The dataset T that constitutes the ContourResUNet network, ; Step 1.3.4, based on the results obtained in step 1.3.3 Zhang image, One set of images is used as the training set for the ContourResUNet network, and the remaining images are used as the test set for the ContourResUNet network. Step 1.4: Design the loss function for the ContourResUNet network; Step 1.4 specifically involves: Step 1.4.1: Obtain the ground truth labels of the ContourResUNet network based on Step 1.1. Learning truth labels, such as Figure 4 As shown; Step 1.4.2, calculate the loss function of the ContourResUNet network using equation (4): (4) In the formula, and The binary detail feature map obtained from step 1.6 and the image predicted by the ContourResUNet network at location ( ) are respectively represented. x , y The pixel value of ); Since the dataset contains carrier images that have suffered almost all types of attacks, the trained ContourResUNet network can extract images from the attacked carrier images that are highly similar to the binary detail feature images obtained from the original carrier images using wavelet transform, thus ensuring the strong robustness of the proposed method.

[0030] Step 1.5, train the network; Step 1.5 specifically involves: Step 1.5.1, Optimizer and learning rate configuration; Specifically: Training uses the Adam optimizer, with an initial learning rate set to 5×10. - ³; Simultaneously, the ReduceLROnPlateau learning rate scheduling strategy is introduced. If the average loss does not decrease over 5 consecutive epochs, the learning rate is multiplied by a coefficient of 0.5 for decay, with a lower limit of 1×10⁻⁶. -6 To promote smooth convergence in the later stages of training; Step 1.5.2: Calculate the loss function and gradient processing; Specifically, in each batch, the input training image is forward propagated to obtain predicted binary detail features, and then the loss function is calculated; during backpropagation, the gradient is clipped to prevent gradient explosion and the network parameters are updated accordingly.

[0031] Step 1.5.3, Training process control; Specifically: after each epoch, the average loss is recorded and the best performing network weights are automatically saved; the training termination condition is set as follows: the loss change is less than 0.001, or the maximum number of training epochs of 150 is reached; Step 1.5.4: Training ends, and the ContourResUNet network outputs a binary detail feature map of the protected remote sensing image.

[0032] Step 2, extract the original watermark image, such as... Figure 5 As shown, the timestamp is loaded and binarized to obtain a binary watermark image, as shown. Figure 6 As shown; Step 2 is as follows: Step 2.1, adjust the size of the original watermark image to... Size, and load the zero-watermark image into the watermark image at the expected time of registration with the copyright center; Step 2.2: If the original watermark image is a color image, use formula (5) to convert the color watermark image into a grayscale image; (5) In the formula, Represents a watermarked image Pixel value at that location, Indicates the red channel of the color watermark image. Pixel value at that location, Indicates the green channel of the color watermark image Pixel value at that location, Indicates the blue channel of the color watermark image. Pixel value at; Step 2.3: Obtain the threshold by calculating the inter-class variance of the original watermark image. Each pixel value of the original watermark image is binarized using formula (6) to obtain a binary watermark image; (6) In the formula, For a binarized binary watermark image, the threshold is... ; Step 3: Logically bind the binary detail feature image output in Step 1.5.4 with the binary watermark image obtained in Step 2.3 to generate a zero-watermark image of the protected remote sensing image, such as... Figure 7 As shown; Step 3 specifically involves: Step 3.1: Input the remote sensing image to be protected into the ContourResUNet network trained in Step 1 to obtain the binary detail feature map of the image to be protected. Step 3.2: XOR the binary detail feature map obtained in Step 3.1 with the binary watermark image obtained in Step 2 to obtain the initial zero-watermark image. ; Step 3.3, process the initial zero-watermark image obtained in step 3.2. Perform a Fibonacci transform to obtain a watermark-free image; Step 3.3 specifically involves: Step 3.3.1, set the Fibonacci scrambling matrix M as shown in equation (7): (7) Step 3.3.2, take the initial zero-watermark image obtained in step 3.2. Each pixel coordinate in Perform transformation processing using equation (8) and iterate. Next, get , ; (8); Step 3.3.3: Based on the coordinate relationship obtained in step 3.3.2, the zero-watermark image is finally obtained by equation (9): (9) In the formula, The image is the result of the coordinate transformation in step 3.3.2 from the initial zero-watermark image. For the final watermark-free image; Step 4: Input the remote sensing image of the copyright to be verified into the trained ContourResUNet network, and output the binary detail feature map of the remote sensing image of the copyright to be verified. Step 5: Logically unbind the binary detail feature map obtained in Step 4 from the zero-watermark image to obtain the watermark image; Step 5 specifically involves: Step 5.1: Perform an inverse Fibonacci transform on the zero-watermarked image registered with the software copyright center to obtain the initial zero-watermarked image. ; Step 5.1.1, set the Fibonacci inverse scrambling matrix. As shown in equation (10): (10) Step 5.1.2, for each pixel coordinate in the zero-watermark image obtained from the software copyright center ( Perform the transformation process of equation (11) and iterate. Next, get , ; (11) Step 5.1.3: Based on the coordinate relationship obtained in step 5.1.2, recover the initial zero-watermark image using equation (12). : (12); Step 5.2, take the initial zero-watermark image obtained in step 5.1. XOR the binary detail feature map obtained in step 4 to extract the binary watermark. ; Step 6: Obtain the watermark image from Step 5. Similarity value to the original watermark image ( The calculation is used to determine the copyright ownership of remote sensing images; Step 6 specifically involves: Step 6.1, take the watermark image obtained in step 5. Compared with the original watermark image when generating the zero-watermark image Similarity is calculated, and a similarity value is obtained according to formula (13). The expression is as follows: (13) In the formula, To obtain the watermarked image, The original watermark image used when generating the zero-watermark image; Step 6.2: Verify copyright ownership based on the similarity value obtained in Step 6.1. Value > 0.5 indicates that the remote sensing image was generated from the watermarked image used when the zero-watermark image was created. Owned by the owner.

[0033] Example 1 The effectiveness and feasibility of this watermarking method against attacks were determined through specific experiments. The NC (Normalized Correlation Value) was used to measure the similarity between the original and extracted watermark images. The ContourResUNet network was trained using a 3090 GPU. Software environment configuration: Python 3.9.13, PyTorch 1.8.1; Experimental parameter settings: Adam optimizer was used, with an initial learning rate of 5e-3, a minimum learning rate of 1e-6, and weight decay of 5e-4. The batch size was 16, and training lasted for 150 epochs.

[0034] To verify the robustness of this method, the remote sensing images used in this method are as follows: Figure 3 The following attacks were evaluated: JPEG compression, salt-and-pepper noise, Gaussian noise, rotation, scaling, and translation attacks. The attack diagrams are shown below. Figure 8 As shown in Table 1, the experimental results of this method are compared with those of the image watermarking algorithm based on improved singular value decomposition and Haar wavelet transform (ISVD-HWT), the medical image zero-watermarking algorithm based on Fourier parameterization (DO-FFT), and the non-zero watermarking method based on dense residual network (WDRN). The specific results show that this invention has strong robustness, significantly better than the traditional watermarking method (ISVD-HWT) and the zero-watermarking method (DO-FFT). Although the robustness is comparable to that of the neural network-based method (WDRN), this invention is a zero-watermarking method, making the watermark completely invisible, and its transparency is superior to the WDRN-based method. The copyright verification effect of remote sensing images under attacks such as JPEG compression, salt-and-pepper noise, Gaussian noise, rotation, scaling, and translation is shown in the figure. Figure 9 As shown, Figure 9 The results further validate the strong robustness of the present invention. Thus, the method of the present invention solves the problem that remote sensing images subjected to conventional image processing and geometric attacks cannot provide image features consistent with those generated during zero-watermarking, in the process of copyright verification of remote sensing images.

[0035] Table 1. NC values ​​of various watermarking methods under different attack levels.

[0036] The feasibility of watermarking methods in eliminating false positives refers to the ability of the watermarked image obtained through its copyright verification process to provide reliable copyright ownership information for the protected image and completely eliminate false copyright claims. For zero-watermarking methods, false positives typically occur in two situations: 1) extracting illegal copyright marks from a protected image; 2) extracting copyright marks from an unprotected image.

[0037] Table 2 and Figures 10-12 The feasibility of the present invention in eliminating watermark false alarms is demonstrated from both numerical results and visual quality perspectives.

[0038] Table 2 shows the NC coefficients of the watermark copyright identifiers extracted from legal and illegal zero-watermark images. The first column of Table 2 represents five remote sensing images, such as... Figure 13 Figures (a)-(e) show, in the first row, zero-watermark images generated by pairing these five remote sensing images with the five watermarked images respectively. Figure 14 The zero-watermark images are (1)-(5). The NC coefficient in the i-th row and j-th column represents the similarity between the watermark image extracted from the i-th remote sensing image and the j-th original watermark image. Here, the watermark image extracted from the i-th remote sensing image is obtained by performing an XOR operation between the binary detail feature image of the remote sensing image and the five zero-watermark images shown in the first row.

[0039] Table 2. NC coefficients of legal and illegal watermarked images

[0040] In Table 2, all diagonal elements are equal to 1, while off-diagonal elements are significantly lower. The NC coefficients on the diagonal are all 1, indicating that the watermark image extracted from the original zero-watermark image of the remote sensing image is completely identical to the original watermark image, providing reliable and trustworthy copyright protection information for the remote sensing image. The low NC values ​​off-diagonally indicate that the watermark extracted from the forged zero-watermark image is of poor quality, with a visual effect far inferior to a genuine watermark. This proves that the proposed method can provide reliable and trustworthy copyright information for protected images.

[0041] Figures 10-12 This document demonstrates a visual comparison of the extraction effects of the method of this invention on legal and illegal watermarked images. Image a in the leftmost column is a remote sensing image. Image b in the second column from the left shows the binary detail feature image used to construct the zero-watermark image. Image c in the third column from the left shows the original watermarked image. Image d in the fourth column from the left shows the zero-watermarked images registered with the Intellectual Property (IPR) center for each protected remote sensing image. Images e in the fifth column from the left and f in the sixth column from the left show the watermarked image and the restored binary detail feature image extracted from the corresponding zero-watermarked image, respectively.

[0042] Figure 10 As can be seen, the restored watermarked image has good image quality and the timestamp is clearly visible, indicating that the present invention can provide reliable copyright ownership information for protected images.

[0043] Figure 11 This demonstrates a series of visual results showing that a protected remote sensing image is protected by another watermarked image (i.e., illegally protected). Despite the fact that the remote sensing image... Figure 11 The watermark image extracted from image a is Figure 11 The image in the image is still clear and of good quality, but the timestamp on this label is later than [previous image]. Figure 10 The timestamp on the identifier obtained from image e in the image suggests that the copyright of this remote sensing image belongs to the watermarked image with the earliest timestamp. Figure 10 The watermarked image shown in Figure e (This invention stipulates that the timestamp on the watermarked image should be later than the registration time of the zero-watermarked image in the IPR center). This demonstrates that this invention can avoid false alarms caused by extracting illegal watermarked images from protected remote sensing images.

[0044] Figure 12 This demonstrates the results of extracting a watermarked image from an unprotected remote sensing image. Because this remote sensing image is unprotected, the IPR center cannot possess its own zero-watermarked image and must instead use zero-watermarked images of other images registered with the center (assuming here...). Figure 10 The image shown in Figure d is the zero-watermarked image of the remote sensing image. It can be seen that the restored watermarked image... Figure 12 The quality of the image in the middle is very poor, and the timestamp information is almost unrecognizable. Furthermore, the binary detail feature image recovered by XORing the corresponding original watermark image with the corresponding zero-watermark image (…) Figure 12 As can be seen from the middle f image, this binary detail feature image corresponds to... Figure 10 (a) is a carrier remote sensing image rather than itself ( Figure 12 (The carrier remote sensing image corresponding to Figure a). Based on these two points, it can be determined that the watermark copyright mark obtained from this remote sensing image is fake, which shows that the present invention can completely avoid the phenomenon of extracting fake watermark images from unprotected remote sensing images.

[0045] In summary, the method of this invention utilizes the binding of remote sensing image detail features with copyright information to construct zero watermarks, solving the problem of false detection and failure of copyright watermark information caused by poor image feature recognition ability in the process of copyrighting remote sensing images, thereby significantly improving the effectiveness and robustness of the algorithm.

[0046] Example 2 The method for protecting copyright of remote sensing images based on detailed features is as follows: Step 1: Design and train the ContourResUNet convolutional neural network to output a binary detail feature map of the remote sensing image to be protected. Step 2: Obtain the binary watermark image; Step 3: Generate a watermark-free image of the protected remote sensing image; Step 4: Input the remote sensing image of the copyright to be verified into the trained ContourResUNet network, and output the binary detail feature map of the remote sensing image of the copyright to be verified. Step 5: Logically unbind the obtained binary detail feature map from the zero-watermark image to obtain the watermark image; Step 6: Calculate the similarity value between the obtained watermarked image and the original watermarked image to determine the copyright ownership of the remote sensing image.

[0047] Example 3 The method for protecting copyright of remote sensing images based on detailed features is as follows: Step 1: Design and train the ContourResUNet convolutional neural network to output a binary detail feature map of the remote sensing image to be protected. Step 1 is as follows: Step 1.1: Obtain the ground truth labels learned by the ContourResUNet convolutional neural network; Step 1.1 specifically involves: Step 1.1.1: Read the remote sensing image to be protected, convert the remote sensing image to grayscale, and adjust the image size. size; Step 1.1.2: Using Daubechies-4 discrete wavelet transform, the remote sensing image processed in step 1.1.1 is decomposed into a two-level decomposition, resulting in a low-frequency approximate sub-band and three high-frequency detail sub-bands in three directions. The discrete wavelet transform is shown in equation (1): (1) In the formula, Represents remote sensing images; , ; Represents low-frequency approximate subbands, Represents horizontal high-frequency detail subbands, Represents vertical high-frequency detail subbands, Represents the diagonal high-frequency detail subband. ; Step 1.1.3: Perform detail feature fusion on the three high-frequency detail subbands obtained in step 1.1.2 through two-level discrete wavelet transform, as shown in equation (2), to obtain the detail feature map. : (2) In the formula, This represents the detail feature map obtained after fusing high-frequency detail subbands; H 2 represents the horizontal high-frequency detail subband obtained from the two-level wavelet decomposition. V 2 represents the vertical high-frequency detail subband obtained from the two-level wavelet decomposition. D 2 represents the diagonal high-frequency detail subband obtained from the two-level wavelet decomposition; Step 1.1.4: Using the inter-class variance maximization criterion, calculate the detail feature map obtained in step 1.1.3. Optimal segmentation threshold ; Step 1.1.5, based on the threshold obtained in step 1.1.4 For the detailed feature map obtained in step 1.1.3 Binarization is performed according to equation (3) to obtain a binary detail feature map, as shown in equation (3) below: (3) In the formula, Represents a binary detail feature map; where, This represents the background area, which corresponds to a flat area in the remote sensing image that has no obvious structure or abrupt changes. This indicates regions of significant detail features, corresponding to areas in remote sensing images where there are significant gradient changes, i.e., regions of texture detail in the image; Step 1.1.6: Use the binary detail feature map obtained in step 1.1.5 as the ground truth label for learning the convolutional neural network ContourResUNet; Step 1.2: Construct the ContourResUNet convolutional neural network structure; In step 1.2, the constructed convolutional neural network ContourResUNet consists of an encoder module, a bottleneck module, a decoder module, and an output layer; the output layer consists of a 1×1 convolutional layer and a sigmoid activation function. Step 1.2 specifically involves: Step 1.2.1: Build the encoder module and extract shallow features from the detail feature map; In step 1.2.1, the encoder module consists of a first-level convolutional coding unit (Encoder1), a second-level convolutional coding unit (Encoder2), a third-level convolutional coding unit (Encoder3), and a fourth-level convolutional coding unit (Encoder4) connected in series. Each convolutional coding unit employs a dual convolutional structure, specifically consisting of two 3×3 convolutional (Conv) operations. Each convolutional layer is followed by batch normalization and ReLU activation functions to enhance feature representation and improve network training stability. To progressively expand the receptive field and extract higher-level spatial semantic information, adjacent convolutional coding units undergo spatial downsampling through 2×2 max-pooling, halving the feature map size at each level. During the encoding process, the number of output channels of the four-level convolutional coding units are set to 64, 128, 256 and 512 respectively, and the corresponding feature map spatial resolutions are 64×64, 32×32, 16×16 and 8×8 respectively, thus forming a multi-scale feature hierarchy structure from shallow to deep. Step 1.2.2: Build the bottleneck module between the encoder module and the decoder module, and extract deep features from the detail feature map; In step 1.2.2, the bottleneck module adopts a dual 3×3 convolution (Conv) structure, and uses batch normalization and ReLU activation function to perform deep fusion of deep features; the 8×8×512 feature map output by the fourth level convolutional coding unit (Encoder4) of the encoder enters the bottleneck module, and through two layers of 3×3 convolution (Conv) operation, its output channel number is increased to 1024, resulting in a high-dimensional semantic feature map with a size of 8×8×1024; This module can effectively capture the global structural information and complex texture relationships of the detailed feature maps of remote sensing images, providing sufficient semantic support for feature reconstruction in the subsequent decoding stage.

[0048] Step 1.2.3: Construct the decoder module and output the binary detail feature map; In step 1.2.3, the decoder module and the encoder module have four symmetrical convolutional layers. The decoder module consists of a first-level convolutional decoding unit (Decoder1), a second-level convolutional decoding unit (Decoder2), a third-level convolutional decoding unit (Decoder3), and a fourth-level convolutional decoding unit (Decoder4) connected in series. Each level of convolutional decoding unit adopts a double convolutional structure, specifically including two layers of 3x3 convolution (Conv) operations. After each convolutional layer, batch normalization and ReLU activation functions are sequentially connected, and combined with 2×2 deconvolution (UpConv) and channel concatenation skip connection (Concat) to achieve deep fusion of multi-scale features. This design aims to reconstruct high-resolution feature maps with rich details, ensure the integrity of detailed information and spatial accuracy, and provide robust and discriminative feature representations for zero-watermark generation and copyright verification.

[0049] During the decoding process: In the initial decoding stage, the first-level convolutional decoding unit (Decoder1) first performs a 2×2 deconvolution upsampling (UpConv) operation on the 8×8×1024 feature map output by the bottleneck module, increasing the feature map size to 16×16 and reducing the number of channels to 512. This is then concatenated with the 16×16×512 feature map output by the fourth-level convolutional encoding unit (Encoder4) of the encoder along the channel dimension to obtain fused features, which are then integrated by the dual convolutional module. Subsequently, in the processing stages of the second-level convolutional decoding unit (Decoder2) and the third-level convolutional decoding unit (Decoder3), the decoder module gradually restores the spatial resolution through multiple levels of 2×2 deconvolution upsampling operations. Specifically, in the processing stage of the second-level convolutional decoding unit (Decoder2), the feature map is changed from 16×16×512... The feature map is upsampled to 32×32×256 and concatenated with the corresponding features of the third-level convolutional coding unit (Encoder3) of the encoder module. In the third-level convolutional decoding unit (Decoder3) processing stage, the feature map is further upsampled from 32×32×256 to 64×64×128 and fused with the output features of the second-level convolutional coding unit (Encoder2) of the encoder module. In the fourth-level convolutional decoding unit (Decoder4) processing stage, the feature map is further upsampled to 128×128×64 and fused with the output features of the first-level convolutional coding unit (Encoder1) of the encoder module. In each decoding stage, the fused features are channel compressed and refined through a dual 3×3 convolutional structure, thereby restoring spatial resolution while suppressing redundant information and enhancing effective structural expression.

[0050] The final feature map output by the decoder module is compressed to 1 channel by a 1×1 convolutional layer in the output layer, generating a map of size [size missing]. Binary detail feature map; Step 1.3: Construct the dataset T for the ContourResUNet network; Step 1.3 specifically involves: Step 1.3.1: Select an image from the remote sensing image dataset RSOD-Datas as the remote sensing image to be protected; Step 1.3.2: Resize the remote sensing image selected in Step 1.3.1. Standard-sized images; Step 1.3.3 involves performing various levels of conventional image processing operations on the image obtained in Step 1.3.2, specifically including salt-and-pepper noise reduction, Gaussian noise reduction, compression, sharpening, smoothing, and quantization. It also includes various levels of geometric attack processing, including rotation, translation, and cropping. All images obtained from these conventional image processing operations and combinations thereof total [number missing]. The dataset T that constitutes the ContourResUNet network; Step 1.3.4, based on the results obtained in step 1.3.3 Zhang image, One set of images is used as the training set for the ContourResUNet network, and the remaining images are used as the test set for the ContourResUNet network. Step 1.4: Design the loss function for the ContourResUNet network; Step 1.4 specifically involves: Step 1.4.1: Obtain the ground truth labels of the ContourResUNet network based on Step 1.1. ; Step 1.4.2, calculate the loss function of the ContourResUNet network using equation (4): (4) In the formula, and The binary detail feature map obtained in step 1.1.5 and the image predicted by the ContourResUNet network at location ( x , y The pixel value of ); Step 1.5, train the network; Step 1.5 specifically involves: Step 1.5.1, Optimizer and learning rate configuration; Specifically: Training uses the Adam optimizer, with an initial learning rate set to 5×10. - ³; Simultaneously, the ReduceLROnPlateau learning rate scheduling strategy is introduced. If the average loss does not decrease over 5 consecutive epochs, the learning rate is multiplied by a coefficient of 0.5 for decay, with a lower limit of 1×10⁻⁶. -6 To promote smooth convergence in the later stages of training; Step 1.5.2: Calculate the loss function and gradient processing; Specifically, in each batch, the input training image is forward propagated to obtain predicted binary detail features, and then the loss function is calculated; Step 1.5.3, Training process control; Specifically: after each epoch, the average loss is recorded and the best performing network weights are automatically saved; the training termination condition is set as follows: the loss change is less than 0.001, or the maximum number of training epochs of 150 is reached; Step 1.5.4: Training ends, and the ContourResUNet network outputs a binary detail feature map of the protected remote sensing image. Step 2: Obtain the binary watermark image; Step 3: Generate a watermark-free image of the protected remote sensing image; Step 4: Input the remote sensing image of the copyright to be verified into the trained ContourResUNet network, and output the binary detail feature map of the remote sensing image of the copyright to be verified. Step 5: Logically unbind the obtained binary detail feature map from the zero-watermark image to obtain the watermark image; Step 6: Calculate the similarity value between the obtained watermarked image and the original watermarked image to determine the copyright ownership of the remote sensing image.

[0051] Example 4 The method for protecting copyright of remote sensing images based on detailed features is as follows: Step 1: Design and train the ContourResUNet convolutional neural network to output a binary detail feature map of the remote sensing image to be protected. Step 2: Obtain the binary watermark image; Step 2 is as follows: Step 2.1, adjust the size of the original watermark image to... Size, and load the zero-watermark image into the watermark image at the expected time of registration with the copyright center; Step 2.2: If the original watermark image is a color image, use formula (5) to convert the color watermark image into a grayscale image; (5) In the formula, Represents a watermarked image Pixel value at that location, Indicates the red channel of the color watermark image. Pixel value at that location, Indicates the green channel of the color watermark image Pixel value at that location, Indicates the blue channel of the color watermark image. Pixel value at; Step 2.3: Obtain the threshold by calculating the inter-class variance of the original watermark image. Each pixel value of the original watermark image is binarized using formula (6) to obtain a binary watermark image; (6) In the formula, This is a binary watermarked image after binarization. Step 3: Generate a watermark-free image of the protected remote sensing image; Step 4: Input the remote sensing image of the copyright to be verified into the trained ContourResUNet network, and output the binary detail feature map of the remote sensing image of the copyright to be verified. Step 5: Logically unbind the obtained binary detail feature map from the zero-watermark image to obtain the watermark image; Step 6: Calculate the similarity value between the obtained watermarked image and the original watermarked image to determine the copyright ownership of the remote sensing image.

[0052] Example 5 The method for protecting copyright of remote sensing images based on detailed features is as follows: Step 1: Design and train the ContourResUNet convolutional neural network to output a binary detail feature map of the remote sensing image to be protected. Step 2: Obtain the binary watermark image; Step 3: Generate a watermark-free image of the protected remote sensing image; Step 3 specifically involves: Step 3.1: Input the remote sensing image to be protected into the ContourResUNet network trained in Step 1 to obtain the binary detail feature map of the image to be protected. Step 3.2: XOR the binary detail feature map obtained in Step 3.1 with the binary watermark image obtained in Step 2 to obtain the initial zero-watermark image. ; Step 3.3, process the initial zero-watermark image obtained in step 3.2. Perform a Fibonacci transform to obtain a watermark-free image.

[0053] Step 4: Input the remote sensing image of the copyright to be verified into the trained ContourResUNet network, and output the binary detail feature map of the remote sensing image of the copyright to be verified. Step 5: Logically unbind the obtained binary detail feature map from the zero-watermark image to obtain the watermark image; Step 6: Calculate the similarity value between the obtained watermarked image and the original watermarked image to determine the copyright ownership of the remote sensing image.

[0054] Example 6 The method for protecting copyright of remote sensing images based on detailed features is as follows: Step 1: Design and train the ContourResUNet convolutional neural network to output a binary detail feature map of the remote sensing image to be protected. Step 2: Obtain the binary watermark image; Step 3: Generate a watermark-free image of the protected remote sensing image; Step 4: Input the remote sensing image of the copyright to be verified into the trained ContourResUNet network, and output the binary detail feature map of the remote sensing image of the copyright to be verified. Step 5: Logically unbind the obtained binary detail feature map from the zero-watermark image to obtain the watermark image; Step 5 specifically involves: Step 5.1: Perform an inverse Fibonacci transform on the zero-watermarked image registered with the software copyright center to obtain the initial zero-watermarked image. ; Step 5.2, take the initial zero-watermark image obtained in step 5.1. XOR the binary detail feature map obtained in step 4 to extract the binary watermark. .

[0055] Step 6: Calculate the similarity value between the obtained watermarked image and the original watermarked image to determine the copyright ownership of the remote sensing image.

Claims

1. A method for copyright protection of remote sensing images based on detailed features, characterized in that, Specifically: Step 1: Design and train the ContourResUNet convolutional neural network to output a binary detail feature map of the remote sensing image to be protected. Step 2: Obtain the binary watermark image; Step 3: Generate a watermark-free image of the protected remote sensing image; Step 4: Input the remote sensing image of the copyright to be verified into the trained ContourResUNet network, and output the binary detail feature map of the remote sensing image of the copyright to be verified. Step 5: Logically unbind the obtained binary detail feature map from the zero-watermark image to obtain the watermark image; Step 6: Calculate the similarity value between the obtained watermarked image and the original watermarked image to determine the copyright ownership of the remote sensing image.

2. The method for copyright protection of remote sensing images based on detailed features according to claim 1, characterized in that, Step 1 is as follows: Step 1.1: Obtain the ground truth labels learned by the ContourResUNet convolutional neural network; Step 1.1 specifically involves: Step 1.1.1: Read the remote sensing image to be protected, convert the remote sensing image to grayscale, and adjust the image size. size; Step 1.1.2: Using Daubechies-4 discrete wavelet transform, the remote sensing image processed in step 1.1.1 is decomposed into a two-level decomposition, resulting in a low-frequency approximate sub-band and three high-frequency detail sub-bands in three directions. The discrete wavelet transform is shown in equation (1): (1) In the formula, Represents remote sensing images; , ; Represents low-frequency approximate subbands, Represents horizontal high-frequency detail subbands, Represents vertical high-frequency detail subbands, Represents the diagonal high-frequency detail subband. ; Step 1.1.3: Perform detail feature fusion on the three high-frequency detail subbands obtained in step 1.1.2 through two-level discrete wavelet transform, as shown in equation (2), to obtain the detail feature map. : (2) In the formula, This represents the detail feature map obtained after fusing high-frequency detail subbands; H 2 represents the horizontal high-frequency detail subband obtained from the two-level wavelet decomposition. V 2 represents the vertical high-frequency detail subband obtained from the two-level wavelet decomposition. D 2 represents the diagonal high-frequency detail subband obtained from the two-level wavelet decomposition; Step 1.1.4: Using the inter-class variance maximization criterion, calculate the detail feature map obtained in step 1.1.

3. Optimal segmentation threshold ; Step 1.1.5, based on the threshold obtained in step 1.1.4 For the detailed feature map obtained in step 1.1.3 Binarization is performed according to equation (3) to obtain a binary detail feature map, as shown in equation (3) below: (3) In the formula, Represents a binary detail feature map; where, This represents the background area, which corresponds to a flat area in the remote sensing image that has no obvious structure or abrupt changes. This indicates regions of significant detail features, corresponding to areas in remote sensing images where there are significant gradient changes, i.e., regions of texture detail in the image; Step 1.1.6: Use the binary detail feature map obtained in step 1.1.5 as the ground truth label for learning the convolutional neural network ContourResUNet; Step 1.2: Construct the ContourResUNet convolutional neural network structure; In step 1.2, the constructed convolutional neural network ContourResUNet consists of an encoder module, a bottleneck module, a decoder module, and an output layer; the output layer consists of a 1×1 convolutional layer and a sigmoid activation function. Step 1.2 specifically involves: Step 1.2.1: Build the encoder module and extract shallow features from the detail feature map; In step 1.2.1, the encoder module consists of a first-level convolutional coding unit, a second-level convolutional coding unit, a third-level convolutional coding unit, and a fourth-level convolutional coding unit connected in series. Each level of convolutional coding unit adopts a double convolutional structure, specifically including two layers of 3×3 convolutional operations; after each convolutional layer, batch normalization and ReLU activation functions are connected in sequence to enhance feature representation ability and improve network training stability; in order to gradually expand the receptive field and extract higher-level spatial semantic information, spatial downsampling is achieved between adjacent convolutional coding units through 2×2 max pooling operations, so that the feature map size is halved step by step; During the encoding process, the number of output channels of the four-level convolutional coding units are set to 64, 128, 256 and 512 respectively, and the corresponding feature map spatial resolutions are N / 8×N / 8, N / 16×N / 16, N / 32×N / 32 and N / 64×N / 64 respectively, thus forming a multi-scale feature hierarchy structure from shallow to deep; Step 1.2.2: Build the bottleneck module between the encoder module and the decoder module, and extract deep features from the detail feature map; In step 1.2.2, the bottleneck module adopts a dual 3×3 convolutional structure, combined with batch normalization and ReLU activation function, to perform deep fusion of deep features; the N / 64×N / 64×512 feature map output by the fourth-level convolutional coding unit of the encoder enters the bottleneck module, and through two layers of 3×3 convolutional operations, its output channel number is increased to 1024, resulting in a high-dimensional semantic feature map with a size of N / 64×N / 64×1024; Step 1.2.3: Construct the decoder module and output the binary detail feature map; In step 1.2.3, the decoder module consists of a first-level convolutional decoding unit, a second-level convolutional decoding unit, a third-level convolutional decoding unit, and a fourth-level convolutional decoding unit connected in series. Each level of convolutional decoding unit adopts a double convolutional structure, specifically including two layers of 3x3 convolutional operations. After each convolutional layer, batch normalization and ReLU activation functions are connected in sequence, and combined with 2×2 deconvolution and channel splicing skip connection, to achieve deep fusion of multi-scale features. During the decoding process: In the initial decoding stage, the first-level convolutional decoding unit first performs a 2×2 deconvolution upsampling operation on the N / 64×N / 64×1024 feature map output by the bottleneck module, increasing the feature map size to N / 32×N / 32 and reducing the number of channels to 512. This is then concatenated with the N / 32×N / 32×512 feature map output by the fourth-level convolutional encoding unit of the encoder along the channel dimension to obtain fused features, which are then integrated by the dual convolutional module. Subsequently, in the processing stages of the second and third-level convolutional decoding units, the decoder module gradually restores the spatial resolution through multiple levels of 2×2 deconvolution upsampling operations. Specifically, in the second-level convolutional decoding unit processing stage, the feature map size is increased from N / 32×N / 32×512... The feature map is upsampled to N / 16×N / 16×256 and concatenated with the features corresponding to the third-level convolutional coding unit of the encoder module. In the third-level convolutional decoding unit processing stage, the feature map is further upsampled from N / 16×N / 16×256 to N / 8×N / 8×128 and fused with the output features of the second-level convolutional coding unit of the encoder module. In the fourth-level convolutional decoding unit processing stage, the feature map is further upsampled to N / 4×N / 4×64 and fused with the output features of the first-level convolutional coding unit of the encoder module. The final feature map output by the decoder module is compressed to 1 channel by a 1×1 convolutional layer in the output layer, generating a map of size [size missing]. Binary detail feature map; Step 1.3: Construct the dataset T for the ContourResUNet network; Step 1.3 specifically involves: Step 1.3.1: Select an image from the remote sensing image dataset RSOD-Datas as the remote sensing image to be protected; Step 1.3.2: Resize the remote sensing image selected in Step 1.3.

1. Standard-sized images; Step 1.3.3 involves performing various levels of conventional image processing operations on the image obtained in Step 1.3.2, specifically including salt-and-pepper noise reduction, Gaussian noise reduction, compression, sharpening, smoothing, and quantization. It also includes various levels of geometric attack processing, including rotation, translation, and cropping. All images obtained from these conventional image processing operations and combinations thereof total [number missing]. The dataset T that constitutes the ContourResUNet network; Step 1.3.4, based on the results obtained in step 1.3.3 Zhang image, One set of images is used as the training set for the ContourResUNet network, and the remaining images are used as the test set for the ContourResUNet network. Step 1.4: Design the loss function for the ContourResUNet network; Step 1.4 specifically involves: Step 1.4.1: Obtain the ground truth labels of the ContourResUNet network based on Step 1.

1. ; Step 1.4.2, calculate the loss function of the ContourResUNet network using equation (4): （4） In the formula, and The binary detail feature map obtained in step 1.1.5 and the image predicted by the ContourResUNet network at location ( x , y The pixel value of ); Step 1.5, train the network; Step 1.5 specifically involves: Step 1.5.1, Optimizer and learning rate configuration; Specifically: Training uses the Adam optimizer, with an initial learning rate set to 5×10. - ³; Simultaneously, the ReduceLROnPlateau learning rate scheduling strategy is introduced. If the average loss does not decrease over 5 consecutive epochs, the learning rate is multiplied by a coefficient of 0.5 for decay, with a lower limit of 1×10⁻⁶. -6 To promote smooth convergence in the later stages of training; Step 1.5.2: Calculate the loss function and gradient processing; Specifically, in each batch, the input training image is forward propagated to obtain predicted binary detail features, and then the loss function is calculated; Step 1.5.3, Training process control; Specifically: after each epoch, the average loss is recorded and the best performing network weights are automatically saved; the training termination condition is set as follows: the loss change is less than 0.001, or the maximum number of training epochs of 150 is reached; Step 1.5.4: Training ends, and the ContourResUNet network outputs a binary detail feature map of the protected remote sensing image.

3. The method for copyright protection of remote sensing images based on detailed features according to claim 1, characterized in that, Step 2 is as follows: Step 2.1, adjust the size of the original watermark image to... Size, and load the zero-watermark image into the watermark image at the expected time of registration with the copyright center; Step 2.2: If the original watermark image is a color image, use formula (5) to convert the color watermark image into a grayscale image; (5) In the formula, Represents a watermarked image Pixel value at that location, Indicates the red channel of the color watermark image. Pixel value at that location, Indicates the green channel of the color watermark image Pixel value at that location, Indicates the blue channel of the color watermark image. Pixel value at; Step 2.3: Obtain the threshold by calculating the inter-class variance of the original watermark image. Each pixel value of the original watermark image is binarized using formula (6) to obtain a binary watermark image; (6) In the formula, This is a binary watermarked image after binarization.

4. The method for copyright protection of remote sensing images based on detailed features according to claim 1, characterized in that, Step 3 specifically involves: Step 3.1: Input the remote sensing image to be protected into the ContourResUNet network trained in Step 1 to obtain the binary detail feature map of the image to be protected. Step 3.2: XOR the binary detail feature map obtained in Step 3.1 with the binary watermark image obtained in Step 2 to obtain the initial zero-watermark image. ; Step 3.3, process the initial zero-watermark image obtained in step 3.

2. Perform a Fibonacci transform to obtain a watermark-free image.

5. The method for copyright protection of remote sensing images based on detailed features according to claim 4, characterized in that, Step 3.3 specifically involves: Step 3.3.1, Set the Fibonacci scrambling matrix M as shown in equation (7): (7) Step 3.3.2, take the initial zero-watermark image obtained in step 3.

2. Each pixel coordinate in Perform transformation processing using equation (8) and iterate. Next, get ; (8)； Step 3.3.3: Based on the coordinate relationship obtained in step 3.3.2, the zero-watermark image is finally obtained by equation (9): (9) In the formula, The image is the result of the coordinate transformation in step 3.3.2 from the initial zero-watermark image. This is the final watermark-free image.

6. The method for copyright protection of remote sensing images based on detailed features according to claim 5, characterized in that, Step 5 specifically involves: Step 5.1: Perform an inverse Fibonacci transform on the zero-watermarked image registered with the software copyright center to obtain the initial zero-watermarked image. ; Step 5.2, take the initial zero-watermark image obtained in step 5.

1. XOR the binary detail feature map obtained in step 4 to extract the binary watermark. .

7. The method for copyright protection of remote sensing images based on detailed features according to claim 1, characterized in that, Step 5.1 specifically involves: Step 5.1.1, set the Fibonacci inverse scrambling matrix. As shown in equation (10): (10) Step 5.1.2, for each pixel coordinate in the zero-watermark image obtained from the software copyright center ( Perform the transformation process of equation (11) and iterate. Next, get ; (11) Step 5.1.3: Based on the coordinate relationship obtained in step 5.1.2, recover the initial zero-watermark image using equation (12). : (12)。 8. The method for copyright protection of remote sensing images based on detailed features according to claim 1, characterized in that, Step 6 specifically involves: Step 6.1, take the watermark image obtained in step 5. Compared with the original watermark image when generating the zero-watermark image Similarity is calculated, and a similarity value is obtained according to formula (13). The expression is as follows: (13) In the formula, To obtain the watermarked image, The original watermark image used when generating the zero-watermark image; Step 6.2: Verify copyright ownership based on the similarity value obtained in Step 6.

1. Value > 0.5 indicates that the remote sensing image was generated from the watermarked image used when the zero-watermark image was created. Owned by the owner.

Images