Robust watermarking method for medical images based on residual network and deep separable convolution
By combining residual network and depth separation convolution technology in medical images, designing RDSCnet network, extracting and embeding feature vectors and watermarks, the problem of poor anti-geometric attack performance of existing medical image watermark technology is solved, and high robustness is achieved and fast embedding effects protect the security and privacy of medical images.
Patent Information
- Application Number
- CN202111243397.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-25
- Publication Date
- 2024-11-01
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing medical image digital watermarking technology has poor performance against geometric attacks, and its robust performance needs to be improved, making it difficult to effectively protect the security and privacy information of medical images.
Using a robust watermark method for medical images based on residual network and depth separable convolution, the RDSCnet network is designed by combining residual network structure and depth separable convolution module, and using medical image data set to train the network to extract feature vectors and chaos settings. The messy watermark performs XOR operation, embeds the watermark into medical images, implements zero watermark embedding technology, and improves the resistance to geometric attacks.
It improves the anti-geometric attack and anti-conventional attack performance of medical image watermarks, has fast computing speed, is robust and invisible, and protects the patient's privacy information and medical image data security.
Smart Images

Figure CN114092302B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image watermarking, and in particular to a medical image robust watermarking method based on residual network and deep separable convolution. Background Art
[0002] The development of medicine has gradually shifted from traditional medicine to telemedicine, which has led to the transmission and sharing of a large number of medical images on the Internet, which involves the security of medical images. In order to ensure that medical images are not tampered with or misused during the transmission and sharing process, it is necessary to perform technical processing on the original medical images to ensure the security of the original medical images.
[0003] Digital watermarking technology was originally used for copyright protection of digital media. Now, by using the invisibility and robustness of zero watermark, patients' personal information can be hidden in medical images, which can protect patients' privacy and prevent patients' medical image data from being tampered with, so as to ensure its safe transmission on the Internet. Therefore, digital watermarking technology can be well applied in the medical field. However, digital watermarking technology for medical images is still not mature enough, has poor anti-geometric attack performance, and its robustness needs to be improved. Summary of the invention
[0004] The present invention provides a medical image robust watermarking method based on residual network and deep separable convolution, which is used to solve the technical problems that the current digital watermarking technology for medical images is still not mature enough, has poor anti-geometric attack performance, and has robust performance that needs to be improved.
[0005] In view of this, the present invention provides a medical image robust watermarking method based on residual network and deep separable convolution, comprising:
[0006] The residual network structure is combined with the depthwise separable convolutional module to obtain the RDSCnet network.
[0007] Initialize the parameters of the RDSCnet network, and use the constructed medical image dataset to train the RDSCnet network to obtain a trained RDSCnet network, wherein the medical image dataset includes original training images, original verification images, and test images;
[0008] The target medical image is input into the trained RDSCnet network, and the features output by the trained RDSCnet network are compressed, binarized and vectorized in sequence to obtain the feature vector of the target medical image;
[0009] Perform chaotic scrambling encryption on the original target watermark to obtain an encrypted chaotic scrambling watermark;
[0010] Performing an XOR operation on the obtained chaotic scrambled watermark and the characteristic vector of the target image to embed the original target watermark information into the target medical image, and obtaining a binary logic key sequence at the same time;
[0011] The medical image to be watermarked is input into the trained RDSCnet network to extract the feature vector of the medical image to be watermarked;
[0012] Retrieving a binary logic key sequence corresponding to the medical image to be used to extract watermark information, performing an XOR operation on a feature vector of the medical image to be used to extract watermark information and each row of the corresponding binary logic key sequence, and extracting an encrypted watermark;
[0013] Decrypt the encrypted watermark to obtain the restored watermark information.
[0014] Optionally, the basic construction unit of the RDSCnet network is a depth-wise separable convolution module, while the initial convolution layer and the terminal convolution layer of the RDSCnet network are conventional convolutions.
[0015] Optionally, before training the RDSCnet network using the constructed medical image dataset, the following steps are also included:
[0016] Perform data augmentation on the original training images and the original validation images.
[0017] Optionally, data augmentation processing is performed on the original training image and the original verification image, including:
[0018] The original training images and the original verification images are subjected to Gaussian noise processing, JPEG compression processing, median filtering processing, rotation processing, scale enlargement and reduction processing, translation processing and shearing processing.
[0019] Optionally, when constructing the medical image dataset, labels are added to images of the medical image dataset, wherein adding labels to images of the medical image dataset includes:
[0020] Compress the image to 12×12 pixels using bicubic interpolation;
[0021] The min-max normalization method is used to make all element values of the 12×12 numerical matrix distributed between 0 and 1, and the 12×12 numerical matrix is used as the label of the image.
[0022] Optionally, the original target watermark is encrypted by chaotic scrambling to obtain an encrypted chaotic scrambling watermark, including:
[0023] Generate 36×36=1296 chaotic sequences through Logistic Map;
[0024] Convert the generated chaotic sequence into a binary chaotic sequence;
[0025] The original target watermark of 36×36 size is XORed with the generated binary chaotic sequence to obtain the encrypted chaotic scrambled watermark.
[0026] Optionally, it also includes:
[0027] The binary logic key sequence is stored in a third party, wherein the binary logic key sequence is bound to the corresponding medical image embedded with the original target watermark information.
[0028] Optionally, the features output by the trained RDSCnet network are compressed, binarized, and vectorized in sequence to obtain a feature vector of the target medical image, including:
[0029] The 12×12 feature matrix output by the trained RDSCnet network is subjected to discrete cosine transform, and the value of the 6×6 area in the upper left corner of the transform coefficient matrix is taken as the image eigenvalue. Then, the 36 eigenvalues are discriminated, and the eigenvalues greater than 0 are judged as 1, otherwise they are judged as 0. The discrimination results of the 36 eigenvalues are converted from matrix form to vector form, and the eigenvalues in vector form are used as the eigenvectors of the target medical image.
[0030] Optionally, the encrypted watermark is decrypted to obtain restored watermark information, including:
[0031] Generate 36×36=1296 chaotic sequences through Logistic Map;
[0032] Convert the generated chaotic sequence into a binary chaotic sequence;
[0033] The encrypted watermark is XORed with the generated binary chaotic sequence to obtain the restored watermark information.
[0034] Optionally, it also includes:
[0035] The original target watermark is obtained, and the normalized correlation coefficient of the original target watermark and the restored watermark is calculated to determine the ownership of the target medical image and the embedded watermark information.
[0036] It can be seen from the above technical solutions that the embodiments of the present invention have the following advantages:
[0037] The present invention provides a medical image robust watermarking method based on residual network and deep separable convolution, which combines the residual network structure with the deep separable convolution module to independently design the RDSCnet network, and uses the constructed medical image data set to train the RDSCnet network, so that the RDSCnet network has better anti-geometric attack performance and anti-conventional attack performance, and has a faster computing speed, and uses the trained RDSCnet network to extract the feature vector of the medical image and the watermark encrypted by chaotic scrambling to perform XOR operation encryption processing, and embed the watermark into the medical image. The embedding of the watermark does not change the content of the original encrypted volume data, which is a zero watermark embedding technology with invisibility. Therefore, the combination of the RDSCnet network and the zero watermark technology improves the anti-geometric performance of the watermark embedded in the medical image, has strong robustness and invisibility, can simultaneously protect the privacy information of the patient and the data security of the medical image, and solves the technical problems that the existing digital watermarking technology for medical images is still not mature enough, has poor anti-geometric attack performance, and has robust performance that needs to be improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] 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 related drawings can be obtained based on these drawings without paying creative work.
[0039] Figure 1 A medical image robust watermarking method based on residual network and deep separable convolution provided in an embodiment of the present invention;
[0040] Figure 2 A residual learning unit of the residual network provided in an embodiment of the present invention;
[0041] Figure 3 Schematic diagram of 36 chest X-ray (CXR) images provided in an embodiment of the present invention;
[0042] Figure 4 A diagram of the RDSCnet network structure provided in an embodiment of the present invention;
[0043] Figure 5 Schematic diagram of 9 medical images in the selected test set provided in an embodiment of the present invention;
[0044] Figure 6 A schematic diagram of a selected original watermark provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0045] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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.
[0046] For easier understanding, see Figures 1 to 6 The present invention provides a medical image robust watermarking method based on residual network and deep separable convolution, including:
[0047] Step 101, the residual network structure is combined with the depthwise separable convolution module to obtain the RDSCnet network.
[0048] The residual network has a unique network connection method, which uses a layer shortcut to connect the front layer to the back layer, solving the degradation problem of deep networks, making the network deeper and achieving better performance. Figure 2 As shown in , every few layers, the feature map of the previous layer and the feature map of the next layer are added point by point, and the new feature map is input into the next layer. Expressed in formula, let the output of the Lth layer be X L , F represents the composite operation of each layer, including convolution, batch normalization, activation function, etc., then X L =F L (X L-1 )+X L-1 The operation based on Shortcut can realize feature reuse and facilitate the back propagation of gradients during training.
[0049] Compared with conventional convolution operations, depthwise separable convolution has lower parameter counts and computational costs, and the network's computational speed is faster, but the network's ability to extract features is slightly reduced. Suppose there is a 6×6 pixel, 3-channel image, the convolution kernel size is 3×3, the output Feature Map is 4, and the padding method is "same padding". If conventional convolution operations are used, the total parameters of the convolution layer are 3×3×3×4+4(bias)=112. Under the same conditions, when using depthwise separable convolution, the number of filters in the convolution layer is 3 (the number of filters is the same as the number of input channels), and each filter has only one convolution kernel, so the parameters of the convolution kernel are 3×3×3+3(bias)=30, but the number of output Feature Maps is 3 at this time. In order to increase the number of output channels to 4, the common method is to use point convolution operations later, that is, to use a convolution kernel of size 1×1. Point convolution operations are the same as conventional convolution operations. Therefore, the number of filters in the point convolution layer is 4, and each filter has 3 1×1 convolution kernels. Then the parameters of the point convolution layer are 1×1×3×4=16, and the number of parameters of the entire depthwise separable convolution operation is 30+16=46. It can be seen that the number of parameters of the depthwise separable convolution operation is much less than that of the conventional convolution.
[0050] In an embodiment of the present invention, a unique network connection method of a residual network structure is combined with a deep separable convolution module with a relatively low number of parameters and computing cost to obtain an RDSCnet network. The RDSCnet network is constructed deeper, has excellent performance, and has a faster computing speed.
[0051] Step 102, initialize the parameters of the RDSCnet network, use the constructed medical image data set to train the RDSCnet network, and obtain a trained RDSCnet network, wherein the medical image data set includes original training images, original verification images, and test images.
[0052] Medical image data is collected and divided into an original training image dataset, an original verification image dataset and a test image dataset in proportion to construct a medical image dataset. Specifically, the ratio of the original training image dataset, the original verification image dataset and the test image dataset is 3:1:1. For example, 1000 chest X-ray images may be selected, such as Figure 3 As shown ( Figure 3Only 36 images are selected for display). In order to improve the randomness of the sample data, the 1000 images are mixed together and then completely shuffled. The 1000 shuffled images are divided into three parts (600, 200 and 200 images) in a ratio of 3:1:1, which are used as the original training image dataset, the original verification image dataset and the test image dataset respectively.
[0053] After designing the RDSCnet network, the RDSCnet network is trained, and the learning rate of the RDSCnet network is set to 0.001, the MiniBatchSize is set to 30, and the Epochs is set to 3. Then, the original training image dataset in the medical image dataset is used to train the network parameters of the RDSCnet network, and the trained RDSCnet network is verified with the original verification image dataset. The test image dataset is used to test the performance of the entire watermark algorithm. In this way, the trained RDSCnet network has good anti-geometric attack performance and anti-conventional attack performance.
[0054] Step 103: input the target medical image into the trained RDSCnet network, and perform compression processing, binarization processing, and vectorization processing on the features output by the trained RDSCnet network in sequence to obtain a feature vector of the target medical image.
[0055] After obtaining the trained RDSCnet network, the trained RDSCnet network can be used to extract the eigenvector of the medical image. The target medical image is used as the input image of the trained RDSCnet network, and the network is forward propagated. The network output result is a 12×12 feature map. The feature map is subjected to discrete cosine transform, and the value of the 6×6 area in the upper left corner of the transform coefficient matrix is taken as the image eigenvalue. Then, the 36 eigenvalues are discriminated, and the eigenvalues greater than 0 are judged as 1, otherwise they are judged as 0. The discrimination results of the 36 eigenvalues are converted from matrix form to vector form, and the eigenvalues in vector form are used as the eigenvectors of the target medical image, recorded as V(i,j).
[0056] Step 104: Perform chaotic scrambling encryption on the original target watermark to obtain an encrypted chaotic scrambling watermark.
[0057] Select the original target watermark to be added, perform chaotic scrambling encryption on the original target watermark, and obtain the encrypted chaotic scrambling watermark. Specifically, the chaotic scrambling watermark process includes:
[0058] A 36×36=1296-bit chaotic sequence is generated by Logistic Map, where the initial value x0 of the chaotic coefficient is set to 0.2, the growth parameter is set to 4, and the number of iterations is 1295.
[0059] Change the generated chaotic sequence into a 0,1 binary chaotic sequence X(j);
[0060] The original target watermark (binary watermark image W(i,j)) of size 36×36 is XORed with the generated binary chaotic sequence X(j) to obtain the encrypted chaotic scrambled watermark BW(i,j).
[0061] Logistic Map is one of the chaotic maps. It is a simple dynamic nonlinear regression with chaotic behavior. Its mathematical definition can be expressed as follows:
[0062] X k+1 =μ·x k ·(1-x k )
[0063] Among them, x k ∈(0,1), 0<μ≤4. When 3.5699456<μ≤4, the Logistic map enters a chaotic state, and the Logistic chaotic sequence at this time can be used as an ideal key sequence.
[0064] Step 105: perform an XOR operation on the obtained chaotic scrambled watermark and the characteristic vector of the target image to embed the original target watermark information into the target medical image, and obtain a binary logic key sequence at the same time.
[0065] By performing bit-by-bit XOR operation on the feature vector V(i, j) of the medical image and each row of the encrypted watermark BW(i, j), the watermark BW(i, j) can be embedded into the target medical image, and the binary logic key sequence Key(i, j) can be obtained at the same time;
[0066]
[0067] Save Key(i,j); store the binary logic key sequence in a third party, where the binary logic key sequence is bound to the corresponding medical image embedded with the original target watermark information for subsequent watermark extraction. By applying to a third party using Key(i,j) as a key, the ownership and use rights of the original medical image can be obtained, thereby achieving the purpose of protecting the medical image.
[0068] Step 106: Input the medical image from which the watermark information is to be extracted into the trained RDSCnet network to extract the feature vector of the medical image from which the watermark information is to be extracted.
[0069] The method for obtaining the characteristic vector V'(i, j) of the medical image from which watermark information is to be extracted is the same as the method for obtaining the characteristic vector V(i, j) of the medical image in step 103, and will not be described in detail herein.
[0070] Step 107: retrieve the binary logic key sequence corresponding to the medical image whose watermark information is to be extracted, perform an XOR operation on the feature vector of the medical image whose watermark information is to be extracted and each row of the corresponding binary logic key sequence to extract the encrypted watermark.
[0071] Obtain the binary logic key sequence Key(i,j) corresponding to the medical image whose watermark information is to be extracted from a third party, perform an XOR operation on the feature vector V'(i,j) of the medical image whose watermark information is to be extracted and each row of the binary logic key sequence Key(i,j), and extract BW'(i,j):
[0072]
[0073] When extracting the watermark, only the key Key(i,j) is needed, and the original image is not required. It is a zero watermark extraction method.
[0074] Step 108: decrypt the encrypted watermark to obtain restored watermark information.
[0075] The same 1296-bit chaotic sequence as in step 104 is generated through Logistic Map, and the generated chaotic sequence is converted into a binary chaotic series X(j). The encrypted watermark BW'(i, j) is XORed with the generated binary chaotic series X(j) to obtain the restored watermark W'(i, j):
[0076]
[0077] The robust watermarking method for medical images based on residual network and deep separable convolution provided by the embodiment of the present invention combines the residual network structure with the deep separable convolution module to independently design the RDSCnet network, and uses the constructed medical image data set to train the RDSCnet network, so that the RDSCnet network has better resistance to geometric attacks and conventional attacks, and has a faster computing speed, and uses the trained RDSCnet network to extract the feature vector of the medical image and perform XOR encryption processing with the watermark after chaotic scrambling encryption, and embed the watermark into the medical image. The embedding of the watermark does not change the content of the original encrypted volume data. It is a zero watermark embedding technology with invisibility. Therefore, the combination of the RDSCnet network and the zero watermark technology improves the anti-geometric performance of the watermark embedded in the medical image, has strong robustness and invisibility, and can simultaneously protect the patient's privacy information and the data security of the medical image, and solves the technical problems that the existing digital watermarking technology for medical images is still not mature enough, has poor anti-geometric attack performance, and has robust performance that needs to be improved.
[0078] In one embodiment, in step 101, the basic construction unit of the RDSCnet network is a depthwise separable convolution module, the detailed structure of which is shown in Table 1.
[0079] Table 1. Detailed structure of the depthwise separable convolution module
[0080]
[0081] Among them, h×w×k is the length×width×number of channels of the input, t is the multiple, s is the stride of the convolution kernel, k′ is the number of channels of the last layer output, conv2d is a conventional two-dimensional convolution operation, DSC_conv2d is a depth-separable convolution operation, and ReLU function is used as the activation function of the network.
[0082] The initial and terminal convolutional layers of the RDSCnet network use conventional convolution operations, with a convolution kernel size of 3×3, followed by 12 depth-separable convolutional modules as shown in Table 1, of which the 1st, 3rd, 5th, and 7th modules use the residual network structure. The designed network has 38 layers and 466293 parameters, and occupies only about 1.8M memory, which is much lower than the size of the current classic networks. The parameter comparison is shown in Table 2. Table 3 shows the detailed information of each module of the RDSCnet network.
[0083] Table 2. Parameter comparison of RDSCnet network and some current classic networks
[0084]
[0085]
[0086] Table 3. Detailed configuration of RDSCnet network
[0087] enter operate t s c 224×224×1 Conv1 - 2 36 112×112×36 Block1 2 1 36 112×112×36 Block2 2 2 36 56×56×36 Block3 3 1 36 56×56×36 Block4 4 2 36 28×28×36 Block5 5 1 36 28×28×36 Block6 5 2 108 14×14×108 Block7 10 / 3 1 108 14×14×108 Block8 20 / 3 1 144 12×12×114 Block9 5 / 2 1 72 12×12×72 Block10 2 1 36 12×12×36 Block11 2 1 6 12×12×6 Block12 3 1 3 12×12×3 Conv2 - 1 1 12×12×1 - - - -
[0088] Here, t and s are the same as those in Table 1, and c is the number of output channels. Conv1 and Conv2 are conventional convolution operations, and all 12 blocks belong to the depthwise separable convolution module introduced in Table 1. Except for the 3×3 convolution kernel in Block 8, which uses the "validpadding" filling method, all other convolution kernels in the network use the "same padding" filling method. The structure diagram of the RDSCnet network is shown in the figure. Figure 4As shown in the figure, each cuboid is a convolutional layer. Except for the Conv1 layer and the third convolutional layer in each Block, there is no activation function behind them. All other convolutional layers are followed by a ReLu activation function. Each Block represents a depth-separable convolution operation, and the network performs a total of 12 depth-separable convolution operations. The convolutional layers represented by the cuboids on the left and right ends of each Block use a 1×1 convolution kernel, and the convolutional layers represented by the cuboids in the middle of each Block use a 3×3 convolution kernel. Block1, Block3, Block5, and Block7 use a residual network structure. Batch Normalization is used to adjust the data after each convolution.
[0089] In one embodiment, in step 102, before using the constructed medical image dataset to train the RDSCnet network, the original training images and the original verification images need to be data enhanced. Specifically, a total of 1000 chest X-ray images are selected and divided into an original training image dataset, an original verification image dataset, and a test image dataset in a ratio of 3:1:1. Then, the original training image dataset and the original verification image dataset have a total of 800 medical images, and data enhancement is performed on these 800 medical images. The data enhancement uses Gaussian noise processing, JPEG compression processing, median filtering processing, rotation processing, scale enlargement and reduction processing, translation processing, and shearing processing, as shown in Table 4.
[0090] Table 4. Specific operations of data augmentation used when constructing the dataset
[0091]
[0092] In this way, one image can be augmented with data to obtain 59 new images, so the training set has (59+1)×600=36,000 images, the validation set has (59+1)×200=12,000 images, and the test set still has 200 images. These 36,000+12,000+200=48,200 images are all adjusted to 224×224 pixel size as a medical image dataset.
[0093] In one embodiment, when constructing a medical image data set, it includes adding labels to images of the medical image data set, wherein adding labels to images of the medical image data set includes: using bicubic interpolation to compress the image to a size of 12×12 pixels, using min-max normalization to make all element values of the 12×12 size numerical matrix distributed between 0 and 1, so as to facilitate network training, and the 12×12 size numerical matrix with element values distributed between 0 and 1 is used as the label of the image. The network training method is the same as the training method in step 102, and will not be repeated here. During training, the optimizer uses a stochastic gradient descent (SGDM) optimizer with momentum, and the loss function is a mean square error function. When the loss function value reaches oscillation and is difficult to decrease, the training is stopped. After the network is trained, the network is saved, and there is no need to train it again in the future.
[0094] In the test set, a 224×224×1 medical image is selected as the input image of the trained RDSCnet network, and the network is forward propagated. The network output result is a 12×12 numerical matrix. DCT transform (discrete cosine transform) is performed on this numerical matrix, and the values of the 6×6 area in the upper left corner of the DCT coefficient matrix are binarized (values greater than 0 are recorded as 1, and values less than or equal to 0 are recorded as 0). The obtained binary values are vectorized, and this 6×6=36-bit vector is used as the feature vector of the medical image.
[0095] A commonly used one-dimensional DCT transform formula is:
[0096]
[0097]
[0098] Among them, f(i) is the original signal, F(u) is the coefficient after DCT transformation, N is the number of points of the original signal f(i), and c(u) can be considered as a compensation coefficient, which can make the DCT transformation matrix an orthogonal matrix.
[0099] The two-dimensional DCT transform is to perform another DCT transform on the basis of the one-dimensional transform. The formula is as follows:
[0100]
[0101]
[0102] In addition, in one embodiment, step 108 further includes:
[0103] Step 109: Obtain the original target watermark, calculate the normalized correlation coefficient between the original target watermark and the restored watermark, and determine the ownership of the target medical image and the embedded watermark information.
[0104] The normalized cross-correlation (NC) method is used to measure the quantitative similarity between the embedded original watermark and the extracted watermark, which is defined as:
[0105]
[0106] Among them, W (i,j) is the original watermark image, W ( ' i,j) is the restored watermark image. The normalized correlation coefficient is a method to measure the similarity between two images. By calculating the normalized correlation coefficient, the similarity of images can be more accurately and objectively evaluated with data.
[0107] In order to better illustrate the technical effect of the medical image robust watermarking method based on residual network and deep separable convolution provided in the embodiment of the present invention, a specific application example of the medical image robust watermarking method based on residual network and deep separable convolution is provided in the embodiment of the present invention:
[0108] The test objects of this application example are 9 medical images with a size of 224×224 pixels in the test set, such as Figure 5 As shown in , a meaningful binary image is selected as the original watermark, and the watermark size is 36×36, as shown in Figure 6 As shown, it is denoted as W={{w(i,j)|w(i,j)=0,1; 1≤i≤M1, 1≤j≤M2}, where M1 and M2 are preset boundary values.
[0109] First, the nine medical images are used as the input images of the trained RDSCnet network. After forward propagation, the output of the network is transformed into a two-dimensional DCT. Then, the value of the 6×6 area in the upper left corner of the DCT coefficient matrix is taken, and binarized and vectorized. The vector value of the 36-bit discrimination result is used as the feature vector of the image. The initial value of the chaotic coefficient is set to 0.2, the growth parameter is 4, and the number of iterations is 1295. Then the original watermark is chaotically encrypted. After the restored watermark W'(i,j) is detected by the watermark algorithm, the normalized correlation coefficient NC (Normalized Cross Correlation) is used to determine whether there is a watermark embedded. When its value is closer to 1, the similarity is higher, thereby judging the robustness of the algorithm.
[0110] In this application example, the robust watermarking method for medical images based on residual networks and deep separable convolutions was tested for its ability to resist conventional attacks and geometric attacks. The experimental data are shown in Table 5.
[0111] Table 5 NC values between watermarks in the 9 medical images tested in the test set
[0112]
[0113]
[0114]
[0115] As can be seen from Table 5, the watermark algorithm has strong resistance to Gaussian noise, JPEG compression, median filtering, rotation, scale, translation and shearing attacks, indicating that the specially trained RDSCnet network has strong resistance to conventional attacks and geometric attacks.
[0116] In addition, this application example also tests Figure 5 The NC values between the feature vectors of the 9 medical images after the feature vectors are extracted by the trained RDSCnet network. The experimental data are shown in Table 6.
[0117] Table 6 Figure 5 The NC values between the feature vectors of the 9 medical images
[0118]
[0119]
[0120] As can be seen from Table 6, the NC values between the feature vectors of these nine medical images are all lower than 0.6, indicating that the method of extracting image feature vectors through the trained RDSCnet deep network is feasible and has a good degree of differentiation for images. Of course, the method of the present invention can be applied not only to medical images in the data set, but also to some medical images outside the data set.
[0121] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A robust watermarking method for medical images based on residual networks and depthwise separable convolution, characterized by: include: The residual network structure is combined with the depthwise separable convolution module to obtain the RDSCnet network; Initializing the parameters of the RDSCnet network and training the RDSCnet network using the constructed medical image dataset to obtain a trained RDSCnet network, wherein the medical image dataset includes original training images, original verification images, and test images; The target medical image is input into the trained RDSCnet network, and the features output by the trained RDSCnet network are compressed, binarized, and vectorized in sequence to obtain the feature vector of the target medical image; Perform chaotic scrambling encryption on the original target watermark to obtain the encrypted chaotic scrambling watermark; Performing an XOR operation on the obtained chaotic scrambled watermark and the characteristic vector of the target image to embed the original target watermark information into the target medical image and obtain a binary logic key sequence at the same time; The medical image to be watermarked is input into the trained RDSCnet network to extract the feature vector of the medical image to be watermarked; Retrieving a binary logic key sequence corresponding to the medical image to be watermarked, performing an XOR operation on a feature vector of the medical image to be watermarked and each row of the corresponding binary logic key sequence to extract the encrypted watermark; Decrypt the encrypted watermark to obtain the restored watermark information.
2. The robust watermarking method for medical images based on residual network and depthwise separable convolution according to claim 1 is characterized in that The basic construction unit of the RDSCnet network is the depthwise separable convolution module, while the initial and terminal convolution layers of the RDSCnet network are conventional convolutions.
3. The robust watermarking method for medical images based on residual network and depthwise separable convolution according to claim 1 is characterized in that Before training the RDSCnet network using the constructed medical image dataset, the following steps are also included: Perform data augmentation on the original training images and the original validation images.
4. The robust watermarking method for medical images based on residual network and depthwise separable convolution according to claim 3 is characterized in that Perform data augmentation on the original training images and original validation images, including: The original training images and the original verification images are subjected to Gaussian noise processing, JPEG compression processing, median filtering processing, rotation processing, scale enlargement and reduction processing, translation processing and shearing processing.
5. The medical image robust watermarking method based on residual network and depthwise separable convolution according to claim 4 is characterized in that: When constructing a medical image dataset, labels are added to images in the medical image dataset. The steps of adding labels to images in the medical image dataset include: Compress the image to 12×12 pixels using bicubic interpolation; The min-max normalization method is used to make all element values of the 12×12 numerical matrix distributed between 0 and 1, and the 12×12 numerical matrix is used as the label of the image.
6. The robust watermarking method for medical images based on residual network and depthwise separable convolution according to claim 1, characterized in that: The original target watermark is encrypted with chaotic scrambling to obtain the encrypted chaotic scrambling watermark, including: Generate 36×36=1296-bit chaotic sequence through Logistic Map; Convert the generated chaotic sequence into a binary chaotic sequence; The original target watermark of size 36×36 is XORed with the generated binary chaotic sequence to obtain the encrypted chaotic scrambled watermark.
7. The robust watermarking method for medical images based on residual network and depthwise separable convolution according to any one of claims 1 to 6, characterized in that: Also includes: The binary logic key sequence is stored in a third party, wherein the binary logic key sequence is bound to the corresponding medical image embedded with the original target watermark information.
8. The medical image robust watermarking method based on residual network and depthwise separable convolution according to claim 1, characterized in that The features output by the trained RDSCnet network are compressed, binarized, and vectorized in sequence to obtain the feature vector of the target medical image, including: The 12×12 feature matrix output by the trained RDSCnet network is subjected to discrete cosine transform, and the value of the 6×6 area in the upper left corner of the transform coefficient matrix is taken as the image eigenvalue. Then, these 36 eigenvalues are discriminated, and the eigenvalues greater than 0 are judged as 1, otherwise they are judged as 0. The discrimination results of the 36 eigenvalues are converted from matrix form to vector form, and the eigenvalues in vector form are used as the eigenvectors of the target medical image.
9. The medical image robust watermarking method based on residual network and depthwise separable convolution according to claim 1, characterized in that: Decrypt the encrypted watermark to obtain the restored watermark information, including: Generate 36×36=1296-bit chaotic sequence through Logistic Map; Convert the generated chaotic sequence into a binary chaotic sequence; The encrypted watermark is XORed with the generated binary chaotic sequence to obtain the restored watermark information.
10. The medical image robust watermarking method based on residual network and depthwise separable convolution according to claim 1 or 9, characterized in that: Also includes: The original target watermark is obtained, and the normalized correlation coefficient between the original target watermark and the restored watermark is calculated to determine the ownership of the target medical image and the embedded watermark information.
Citation Information
Patent Citations
Medical image robust zero-watermark method based on Bandelet-DCT (Discrete Cosine Transform)
CN111968025A
Medical image robust watermarking method based on LBP-DCT
CN113160030A
