A medical image robust zero-watermarking method based on residual-densenet

Abstract

The application relates to the technical field of image processing, and discloses a medical image robust zero watermark method based on a Residual-DenseNet, which comprises the following steps: constructing a Residual-DenseNet network model by using a residual structure and a DenseNet-121 network; training the Residual-DenseNet network model on a medical picture data sample set; extracting a preliminary feature vector of a medical image by using the trained Residual-DenseNet network model; performing mean binaryzation on the preliminary feature vector of the medical image to generate a feature vector matrix of the medical image; and performing exclusive or operation bit by bit on the feature vector matrix of the medical image and chaotic encryption watermark, so that the watermark information is embedded into the medical image. The method has strong anti-geometric attack and conventional attack capabilities, and can protect medical image data and patient privacy information.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a robust zero-watermarking method for medical images based on Residual-DenseNet. Background Technology

[0002] With the advent of the digital age, various advanced digital products and services have entered people's daily lives. To improve hospital management and provide better services to patients, many medical institutions tend to replace traditional medical systems with new digital medical systems. Various new digital medical imaging devices can conveniently and quickly generate and store large amounts of medical images. This means that a large amount of electronic medical data is being disseminated on the internet. However, internet security still presents many challenges. Ensuring the security of medical data shared via the internet is a key issue that needs to be addressed.

[0003] Digital watermarking is a key information security technology. It can create watermarks to conceal important information such as patient privacy, thereby ensuring the security of this information. However, conventional digital watermarking techniques cannot be directly applied to medical images because embedding watermarks directly into them would affect the image's visual quality.

[0004] In response, many researchers have specifically studied digital watermarking techniques applicable to medical images, referred to as Medical Image Watermarking (MIW). From the perspective of the watermark embedding domain, existing research can be broadly categorized into spatial domain MIW algorithms and transform domain MIW algorithms. However, spatial domain MIW algorithms are not robust to watermark attacks, while transform domain MIW algorithms are not robust to geometric attacks. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a robust zero-watermarking method for medical images based on Residual-DenseNet, which has strong resistance to geometric attacks and conventional attacks, and can protect medical image data and patient privacy information. The specific solution is as follows:

[0006] A robust zero-watermarking method for medical images based on Residual-DenseNet includes:

[0007] Construct a Residual-DenseNet network model using residual structure and DenseNet-121 network;

[0008] The Residual-DenseNet network model was trained on a medical image data sample set.

[0009] The trained Residual-DenseNet network model is used to extract preliminary feature vectors from medical images;

[0010] The initial feature vector of the medical image is subjected to mean binarization to generate the feature vector matrix of the medical image;

[0011] The feature vector matrix of the medical image and the chaotic encrypted watermark are XORed bit by bit to embed the watermark information into the medical image.

[0012] Preferably, in the above-described robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention, the process of constructing the Residual-DenseNet network model includes:

[0013] Build and pre-train the DenseNet-121 network;

[0014] A residual structure is added to the last dense block of the pre-trained DenseNet-121 network to construct the backbone network of the Residual-DenseNet network model; the backbone network is used to extract feature maps.

[0015] Preferably, in the above-described robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention, the process of constructing the Residual-DenseNet network model further includes:

[0016] The feature output module of the Residual-DenseNet network model is constructed using a two-dimensional convolutional layer, a global average pooling layer, and a one-dimensional convolutional layer. The feature output module is used to process the feature maps extracted by the backbone network to obtain a preliminary feature vector of length 64.

[0017] Preferably, in the above-described robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention, the process of training the Residual-DenseNet network model includes:

[0018] The training of the Residual-DenseNet network model is constrained by an objective function consisting of a smooth L1 loss and a counterexample loss; the smooth L1 loss is used to measure the difference between the predicted preliminary feature vector and the true preliminary feature vector; the counterexample loss is used to measure the similarity between the predicted preliminary feature vector and the true preliminary feature vector of the counterexample image.

[0019] Preferably, in the above-described robust zero-watermarking method for medical images based on Residual-DenseNet provided in this embodiment of the invention, the objective function is:

[0020] L=L1+αL C

[0021]

[0022]

[0023] Where L1 is the smoothing L1 loss, L C The counterexample loss is α, where L is L. C The weight, y(i) and y(i) represent the i-th component of the predicted preliminary feature vector and the true preliminary feature vector of the current predicted sample image, respectively. It is the preliminary feature vector of the currently predicted sample image, y C(i) is the true preliminary feature vector of the i-th counterexample image of the currently predicted sample image, N is the number of original sample images, and CS(x1, x2) is the cosine similarity between the two vectors.

[0024] Preferably, the robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention further includes:

[0025] While embedding watermark information into the medical image, a binary logic key sequence is generated;

[0026] The trained Residual-DenseNet network is used to extract preliminary feature vectors from the medical image under test;

[0027] The preliminary feature vector of the medical image to be tested is binarized by mean to generate the feature vector matrix of the medical image to be tested;

[0028] The feature vector matrix of the medical image to be tested and the binary logic key sequence are XORed bit by bit to extract the encrypted watermark.

[0029] Preferably, in the above-described robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention, after extracting the encrypted watermark, the method further includes:

[0030] A chaotic sequence is generated by a Logistic mapping, and the chaotic sequence is binarized to generate a binary sequence.

[0031] An encryption matrix is ​​constructed using the binary sequence, and the encryption matrix and the extracted encrypted watermark are XORed bit by bit to obtain the decrypted watermark.

[0032] Preferably, in the above-described robust zero-watermarking method for medical images based on Residual-DenseNet provided in this embodiment of the invention, before performing a bitwise XOR operation on the feature vector matrix of the medical image and the chaotic encrypted watermark, the method further includes:

[0033] Logistic chaotic encryption is applied to the original watermark to obtain a chaotic encrypted watermark.

[0034] Preferably, in the above-described robust zero-watermarking method for medical images based on Residual-DenseNet provided in this embodiment of the invention, the original watermark is subjected to Logistic chaotic encryption to obtain a chaotic encrypted watermark, including:

[0035] A chaotic sequence is generated by a Logistic mapping, and the chaotic sequence is binarized to generate a binary sequence.

[0036] An encryption matrix is ​​constructed using the binary sequence, and the encryption matrix and the original watermark are XORed bit by bit to obtain a chaotic encrypted watermark.

[0037] Preferably, in the above-described robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention, before constructing the Residual-DenseNet network model, the method further includes:

[0038] The original sample image is processed using 2D-DCT, and a portion of the coefficients are selected from the transform coefficients to form a preliminary feature vector of length 64.

[0039] The constructed preliminary feature vectors are used as labels for the original sample image and the corresponding attacked sample image;

[0040] The medical image data sample set is obtained based on the original sample image, the corresponding attacked sample image, and the label.

[0041] As can be seen from the above technical solution, the robust zero-watermarking method for medical images based on Residual-DenseNet provided by the present invention includes: constructing a Residual-DenseNet network model using a residual structure and a DenseNet-121 network; training the Residual-DenseNet network model on a medical image data sample set; extracting preliminary feature vectors of medical images using the trained Residual-DenseNet network model; performing mean binarization on the preliminary feature vectors of medical images to generate a feature vector matrix of medical images; and performing a bitwise XOR operation between the feature vector matrix of medical images and a chaotic encrypted watermark to embed the watermark information into the medical images.

[0042] The robust zero-watermarking method for medical images provided by this invention utilizes a trained Residual-DenseNet network to extract preliminary feature vectors from medical images. After mean binarization, the final feature vectors of the medical images are obtained. The extracted feature vectors have strong robustness. Applying them to the zero-watermarking algorithm enables the algorithm to gain strong resistance to geometric attacks and conventional attacks, especially geometric attacks such as rotation, translation, and shearing. Furthermore, it has strong invisibility, which can protect medical image data and patient privacy information. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0044] Figure 1 A flowchart of a robust zero-watermarking method for medical images based on Residual-DenseNet provided in an embodiment of the present invention;

[0045] Figure 2 This is a flowchart illustrating the feature vector extraction process of the Residual-DenseNet network model provided in an embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the Residual-DenseNet network model provided in an embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of the Feature Output module provided in an embodiment of the present invention;

[0048] Figure 5This is a schematic diagram of the structure of a Residual Block provided in an embodiment of the present invention;

[0049] Figure 6 This is a schematic diagram of the structure of a Dense Block provided in an embodiment of the present invention;

[0050] Figure 7 This is an overall architecture diagram of the robust zero-watermarking method for medical images provided in an embodiment of the present invention;

[0051] Figure 8 These are test images for testing the robustness of the algorithm provided in this embodiment of the invention.

[0052] Figure 9 The original watermark image provided in the embodiments of the present invention;

[0053] Figure 10 The image is a chaotically encrypted watermark image provided in an embodiment of the present invention;

[0054] Figure 11 This is an image with 30% Gaussian noise provided in an embodiment of the present invention;

[0055] Figure 12 This is an image provided by an embodiment of the present invention after being rotated 45° counterclockwise;

[0056] Figure 13 This is an image of the Y-axis after being cut by 8% according to an embodiment of the present invention;

[0057] Figure 14 Test images for comparative experiments provided in embodiments of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] This invention provides a robust zero-watermarking method for medical images based on Residual-DenseNet, such as... Figure 1 As shown, it includes the following steps:

[0060] S101. Construct a Residual-DenseNet network model using the residual structure and DenseNet-121 network.

[0061] S102. Train the Residual-DenseNet network model on the medical image data sample set.

[0062] S103. Use the trained Residual-DenseNet network model to extract preliminary feature vectors from medical images.

[0063] In practical applications, two medical images, one of the head and one of the abdomen (each 256 pixels × 256 pixels × 3 channels), can be selected as the original medical images, denoted as I. I(i, j) represents the pixel value of each point in the original medical image.

[0064] When performing step S103, as Figure 2 As shown, the original medical image can be used as the input of the Residual-DenseNet network model to extract the preliminary feature vector PFV(i) of the original medical image.

[0065] S104. Perform mean binarization on the preliminary feature vectors of the medical image to generate the feature vector matrix of the medical image.

[0066] Specifically, such as Figure 2 As shown, the initial feature vector PFV(i) of the medical image is binarized using mean binarization to generate a 64-bit binary sequence FV(i). FV(i) is the final feature vector of the original medical image.

[0067]

[0068] The feature vector FV(i) of the medical image is reshaped into a 2×32 matrix, and 16 such matrices are used to construct the feature vector matrix FV of the medical image. m (i, j).

[0069] S105. Perform a bitwise XOR operation on the feature vector matrix of the medical image and the chaotic encrypted watermark to embed the watermark information into the medical image.

[0070] Specifically, the feature vector matrix FV of the medical image m By performing a bitwise XOR operation between (i, j) and the chaotic encrypted watermark EW(i, j), the watermark can be embedded into the original medical image.

[0071] In specific implementation, before performing the bitwise XOR operation between the feature vector matrix of the medical image and the chaotic encrypted watermark in step S105, the method may further include: performing Logistic chaotic encryption on the original watermark to obtain the chaotic encrypted watermark EW(i,j). In practical applications, the patient's privacy information is made into a binary text image as the original watermark embedded in the medical image (here, 32 pixels × 32 pixels is used).

[0072] The original watermark can be denoted as W = {W(i,j) | W(i,j) = 0, 1; 1 ≤ i ≤ 32, 1 ≤ j ≤ 32}. This invention employs zero-watermark embedding technology, which effectively solves the defects caused by the modification of the original medical image data by traditional watermark embedding technology, thus ensuring the visual quality of medical images.

[0073] In the above-mentioned robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention, the preliminary feature vector of the medical image is extracted using the trained Residual-DenseNet network. After mean binarization, the final feature vector of the medical image is obtained. The extracted feature vector has strong robustness. Applying it to the zero-watermarking algorithm enables the algorithm to obtain strong resistance to geometric attacks and conventional attacks, especially geometric attacks such as rotation, translation and shearing. It also has strong invisibility, which can protect medical image data and patient privacy information.

[0074] In specific implementation, in the above-mentioned robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention, the process of constructing the Residual-DenseNet network model in step S101 may specifically include: establishing and pre-training a DenseNet-121 network; adding a residual structure at the last dense block of the pre-trained DenseNet-121 network to construct the backbone network of the Residual-DenseNet network model; the backbone network is used to extract feature maps.

[0075] like Figure 3 As shown, the Residual-DenseNet network model can be divided into two parts: the backbone network and the feature output module. This invention is based on the DenseNet-121 network pre-trained on ImageNet, and adds a residual structure at Dense Block 4 to obtain the backbone network of the Residual-DenseNet network model. Its input is a medical image, and its output is a feature map extracted from the medical image.

[0076] In specific implementation, the above-mentioned robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention may further include the following in step S101: constructing a feature output module of the Residual-DenseNet network model using a two-dimensional convolutional layer, a global average pooling layer, and a one-dimensional convolutional layer; the feature output module is used to process the feature map extracted by the backbone network to obtain a preliminary feature vector of length 64.

[0077] like Figure 4 As shown, the feature output module contains a two-dimensional convolutional layer (kernel size: 1×1), a global average pooling layer, and a one-dimensional convolutional layer. After processing by the feature output module, the feature maps extracted by the backbone network yield a feature vector of length 64. This is the initial feature vector PFV(i) of the medical image.

[0078] It should be added that, such as Figure 5 As shown, the core module of Residual Networks (ResNet) is the Residual Block, where shortcut connections play a major role. There are two forms: identity shortcuts and projection shortcuts. Residual learning is achieved through shortcut connections and element-wise addition. Furthermore, the most significant characteristic of Dense Convolutional Networks (DenseNet) is that the input to each layer comes from the outputs of all preceding layers. These outputs are combined by concatenating them along the channel dimension; this structure is called dense connectivity. Figure 6 As shown, the core module of DenseNet is the Dense Block, which extensively uses dense connections to combine feature maps from different layers. This core module greatly promotes feature reuse.

[0079] In a specific implementation, the robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention, in step S102 of training the Residual-DenseNet network model, may specifically include: using an objective function composed of smoothed L1 loss and counterexample loss to constrain the training of the Residual-DenseNet network model; the smoothed L1 loss is used to measure the preliminary predicted feature vector (predicted label). The difference between the predicted initial feature vector (the true label y) and the true preliminary feature vector (the true label y); the counterexample loss is used to measure the difference between the predicted initial feature vector (the predicted label y) and the true preliminary feature vector (the true label y). ) and the true preliminary feature vector of the counterexample image (the true label y of the counterexample image) c The similarity between the two. Counterexample images refer to other original sample images that do not include the original sample image corresponding to the currently predicted sample image.

[0080] In practical implementation, the objective function can be:

[0081] L=L1+αL c

[0082]

[0083] Where L1 is the smoothing L1 loss, L C It is the counterexample loss, α is L C The weight, y(i) and y(i) represent the i-th component of the predicted preliminary feature vector and the true preliminary feature vector of the current predicted sample image, respectively. It is the preliminary feature vector of the currently predicted sample image, y C(i) is the true preliminary feature vector of the i-th counterexample image of the currently predicted sample image, N is the number of original sample images, and CS(x1, x2) is the cosine similarity between the two vectors.

[0084] It should be noted that Residual-DenseNet is implemented using the PyTorch framework. This invention can use the Adam optimizer, where β1 is 0.9, β2 is 0.999, and the weight decay is 10. -5 The initial learning rate is 0.001, decreasing by 10% every 15 epochs. The entire training process can consist of 100 epochs. The objective function set in this invention constrains the network training process, enabling the trained Residual-DenseNet network model to extract feature vectors that simultaneously satisfy robustness and representativeness.

[0085] In specific implementation, in the above-described robust zero-watermarking method for medical images based on Residual-DenseNet provided in this embodiment of the invention, after executing step S105, as follows: Figure 7 As shown, it may also include: generating a binary logic key sequence Key(i, j) while embedding the watermark information into the medical image:

[0086]

[0087] The binary logic key sequence Key(i,j) is stored on a third-party platform. To obtain ownership and usage rights of the original medical images, it is necessary to apply for the binary logic key sequence Key(i,j) from the third-party platform. This achieves the purpose of protecting medical images, enhances security, and adapts to the practicality and standardization of modern network technology.

[0088] Next, the trained Residual-DenseNet network is used to extract the preliminary feature vector of the medical image I′(i,j) to be tested; the preliminary feature vector of the medical image I′(i,j) to be tested is binarized by mean to obtain the feature vector FV′(i) of the medical image I′(i,j) to be tested; the feature vector matrix FV′m(i,j) of the medical image I′(i,j) to be tested is constructed using the feature vector FV′(i) of the medical image I′(i,j) to be tested; the feature vector matrix FV′m(i,j) of the medical image I′(i,j) to be tested is then processed. m The encrypted watermark EW′(i,j) is extracted by performing a bitwise XOR operation between (i,j) and the binary logic key sequence Key(i,j).

[0089]

[0090] This algorithm is a zero-watermark extraction algorithm. When extracting the watermark, only the key Key(i,j) is needed, and the original image is not required.

[0091] In specific implementation, the robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention may further include the following steps after extracting the encrypted watermark: generating a chaotic sequence X(i) through Logistic mapping, performing a binarization operation on the chaotic sequence X(i) to generate a binary sequence; constructing an encryption matrix EM(i,j) using the binary sequence, and performing a bitwise XOR operation between the encryption matrix EM(i,j) and the extracted encrypted watermark EW′(i,j) to obtain the decrypted watermark W′(i,j).

[0092] Specifically, a chaotic sequence X(i) is generated using a Logistic mapping. In this experiment, the initial value of the chaotic coefficient x0 is set to 0.2, the growth parameter μ is set to 4, and the number of iterations is set to 1023. The chaotic sequence X(i) is binarized (threshold set to 0.5) and reshaped into a 32×32 matrix to obtain the binary encryption matrix EM(i,j). Then, the encryption matrix EM(i,j) and the extracted encrypted watermark EW′(i,j) are XORed bitwise to obtain the decrypted watermark W′(i,j).

[0093]

[0094] In specific implementation, in the above-mentioned robust zero-watermarking method for medical images based on Residual-DenseNet provided in the embodiments of the present invention, the original watermark is subjected to Logistic chaotic encryption to obtain a chaotic encrypted watermark. Specifically, it may include: generating a chaotic sequence X(i) through Logistic mapping, and performing a binarization operation on the chaotic sequence X(i) to generate a binary sequence; constructing an encryption matrix EM(i,j) using the binary sequence, and performing a bitwise XOR operation between the encryption matrix EM(i,j) and the original watermark W(i,j) to obtain a chaotic encrypted watermark EW(i,j).

[0095] Specifically, a chaotic sequence X(i) is generated using a Logistic mapping. In this experiment, the initial value of the chaotic coefficient x0 is set to 0.2, the growth parameter μ is set to 4, and the number of iterations is set to 1023. The chaotic sequence X(i) is binarized (threshold set to 0.5) and reshaped into a 32×32 matrix to obtain the binary encryption matrix EM(i). The encryption method is the same as described above. Then, the encryption matrix EM(i,j) and the binary watermark W(i,j) are XORed bit by bit to obtain the chaotic encryption watermark EW(i,j):

[0096]

[0097] Understandably, because this invention establishes a direct relationship between the watermark and the image's feature vector through an XOR operation, comparing the similarity between the original watermark and the extracted watermark, or comparing the similarity between the feature vector of the original image and the feature vector of the attacked image, can measure the algorithm's robustness. The similarity between two watermarked images can be measured using the following formula:

[0098]

[0099] Where W(i,j) represents the original watermark, and W′(i,j) represents the extracted watermark. It is the mean of W(i,j). It is the mean of W′(i,j).

[0100] For ease of calculation, this invention uses a method of comparing the similarity between the feature vectors of the original image and the feature vectors of the image after the attack, and uses the Pearson correlation coefficient to measure their similarity.

[0101] The aforementioned robust zero-watermarking method for medical images can include medical image feature extraction, watermark encryption, watermark embedding, watermark extraction, and watermark decryption based on the Residual-DenseNet network model. It should be noted that the core of this invention is the extraction of medical image feature vectors, and the quality of these feature vectors directly determines the performance of the algorithm. This invention uses a trainable Residual-DenseNet network to extract robust feature vectors from medical images. Unlike traditional image feature extraction methods, it does not require manual feature design; instead, it achieves image feature extraction and abstraction through training, resulting in more powerful feature learning and representation capabilities. Furthermore, since deep neural network models require many parameters to be trained, a large amount of labeled data is needed as the training dataset.

[0102] To achieve better training results with a small training set and to address the weakness of traditional DCT-based zero-watermarking algorithms in resisting geometric attacks, the implementation may include the following steps before constructing the Residual-DenseNet network model: processing the original sample images using 2D-DCT, selecting a portion of the coefficients from the transform coefficients to construct a preliminary feature vector of length 64; using the constructed preliminary feature vector as the label for the original sample image and the corresponding attacked sample image; and obtaining a medical image data sample set based on the original sample image, the corresponding attacked sample image, and the labels. Abdominal and brain images can be selected as the original sample images. To train the network model's resistance to watermark attacks, data augmentation methods are used to obtain the corresponding attacked sample images. This invention can apply various watermark attacks to the original sample images, such as Gaussian noise, JPEG compression, rotation, scaling, and translation.

[0103] In other words, this invention can use a DenseNet-121 network model pre-trained on ImageNet as the initial model, and use the low-frequency coefficients extracted from the DCT of the original medical image as labels. Through training, the Residual-DenseNet network model becomes more robust than the feature vectors extracted by the DCT algorithm. From the final result, the trained Residual-DenseNet network model is approximately equivalent to an inverse transform module followed by a DCT module. The inverse transform module can approximately transform the input sample image into the original sample image, regardless of whether the input is the original sample image or an attacked sample image.

[0104] The following examples demonstrate the effectiveness of the Residual-DenseNet-based robust zero-watermarking method for medical images against conventional and geometric attacks, and compare it with other algorithms.

[0105] The experimental test consisted of three original medical images, each measuring 256×256×3 pixels. (See attached image.) Figure 8 Select a 32×32 binary image with special meaning as the original watermark, see... Figure 9 .

[0106] Preliminary feature vectors of the original medical images were extracted using a trained Residual-DenseNet network, and mean binarization was performed. The attacked images were also processed using the same method to extract preliminary feature vectors and perform mean binarization. Furthermore, the initial value of the chaos coefficient x0 was set to 0.2, the growth parameter μ was set to 4, and the number of iterations was set to 1023 to perform chaotic encryption on the original watermark. The encrypted watermark is shown below. Figure 10 After extracting the watermark W′(i,j), this invention determines the similarity between the two watermarks by calculating the correlation coefficient NC between the original watermark and the extracted watermark, thereby assessing the robustness of the algorithm. The closer the NC value is to 1, the higher the similarity. Furthermore, this invention uses PSNR to represent the degree of image distortion. The higher the PSNR value, the lower the image distortion.

[0107] First, we tested the ability of robust zero-watermarking methods for medical images to resist conventional attacks.

[0108] This invention tested the robustness of the provided medical image robust zero-watermarking method against conventional attacks, and the test results are shown in Table 1. The test results show that the NC values ​​are all higher than 0.7, indicating that the provided method has strong robustness against conventional attacks. That is, even under high-intensity conventional attacks, the watermarked image can still be extracted relatively accurately. For example, when the Gaussian noise variance reaches 0.3, the image is severely contaminated by noise, but the extracted watermark content can still be clearly identified, such as... Figure 11 As shown.

[0109] Table 1. Test results of conventional attacks

[0110]

[0111] Currently, the robustness of watermarking algorithms against geometric attacks is a key area of ​​research. Table 2 shows further experiments conducted to test the robustness of the provided zero-watermarking method for medical images against geometric attacks. As can be seen from Table 2, the provided method exhibits strong robustness against geometric attacks, especially rotational attacks and X-axis shearing. Figure 12 As shown, when the rotation angle is 45°, the NC value is still greater than 0.8; Figure 13 As shown, when the X-axis shearing ratio is 39%, a portion of the image content is missing and stretched to maintain the image size, yet the watermark image can still be extracted relatively well.

[0112] Table 2 Test Results of Geometric Attacks

[0113]

[0114]

[0115] To better illustrate the robustness of the proposed method, it is compared with some existing robust watermarking algorithms. Figure 14 This image, a built-in MATLAB image, is frequently used by other researchers for comparison and testing. Therefore, this invention selected this image as the base test image during the algorithm comparison process. The test results of this image under different attacks are shown in Table 3. From the algorithm comparison results, the provided algorithm performs better than the other two algorithms in resisting geometric attacks.

[0116] Table 3 shows the comparison results with other algorithms.

[0117]

[0118] It should be noted that the robust zero-watermarking method for medical images provided by this invention can first use the properties of Logistic Map to scramble and encrypt the watermark; then, train a Residual-DenseNet network on a training set composed of brain and abdominal medical images, and constrain the training process with a pre-designed objective function; then, use the trained Residual-DenseNet network model to extract the feature vector of the medical image; XOR the feature vector with the encrypted watermark to obtain a binary logic sequence, and store this binary logic sequence on a third-party platform; similarly, the Residual-DenseNet network model is used to extract the feature vector of the medical image to be tested, and XOR it with the binary logic sequence stored on the third-party platform to extract the watermark. In this way, deep learning and zero-watermarking technology are combined to achieve zero-watermarking of medical images against geometric attacks and conventional attacks.

[0119] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0120] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0121] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0122] In summary, the present invention provides a robust zero-watermarking method for medical images based on Residual-DenseNet, comprising: constructing a Residual-DenseNet network model using a residual structure and a DenseNet-121 network; training the Residual-DenseNet network model on a medical image data sample set; extracting preliminary feature vectors of the medical image using the trained Residual-DenseNet network model; performing mean binarization on the preliminary feature vectors of the medical image to generate a feature vector matrix of the medical image; and performing a bitwise XOR operation between the feature vector matrix of the medical image and the chaotic encrypted watermark to embed the watermark information into the medical image. The aforementioned robust zero-watermarking method for medical images utilizes a trained Residual-DenseNet network to extract preliminary feature vectors from medical images. After mean binarization, the final feature vectors of the medical images are obtained. The extracted feature vectors exhibit strong robustness, and applying them to the zero-watermarking algorithm enhances the algorithm's resistance to geometric and conventional attacks, especially geometric attacks such as rotation, translation, and shearing. Furthermore, it possesses strong invisibility, thus protecting medical image data and patient privacy information.

[0123] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0124] The robust zero-watermarking method for medical images based on Residual-DenseNet provided by this invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A robust zero-watermarking method for medical images based on Residual-DenseNet, characterized in that, include: Construct a Residual-DenseNet network model using residual structure and DenseNet-121 network; The process of constructing the Residual-DenseNet network model includes: establishing and pre-training a DenseNet-121 network; adding a residual structure at the last dense block of the pre-trained DenseNet-121 network to construct the backbone network of the Residual-DenseNet network model; the backbone network is used to extract feature maps; a feature output module of the Residual-DenseNet network model is constructed using a two-dimensional convolutional layer, a global average pooling layer, and a one-dimensional convolutional layer; the feature output module is used to process the feature maps extracted by the backbone network to obtain a preliminary feature vector of length 64. The Residual-DenseNet network model is trained on a medical image data sample set. The training process includes: constraining the training of the Residual-DenseNet network model using an objective function composed of a smoothed L1 loss and a counterexample loss; the smoothed L1 loss is used to measure the difference between the predicted initial feature vector and the true initial feature vector; the counterexample loss is used to measure the similarity between the predicted initial feature vector and the true initial feature vector of a counterexample image; the objective function is: ； ； ； in, It is the smoothed L1 loss, The loss is the aforementioned counterexample. yes The weight, and Let i and represent the i-th components of the predicted preliminary feature vector and the true preliminary feature vector of the current predicted sample image, respectively. It is the preliminary feature vector of the currently predicted sample image. It is the true preliminary feature vector of the i-th negative example image of the currently predicted sample image. N It is the number of original sample images. It is the cosine similarity between two vectors; The trained Residual-DenseNet network model is used to extract preliminary feature vectors from medical images; The preliminary feature vectors of the medical image are subjected to mean binarization using the following formula to generate the feature vector matrix of the medical image: ； in, This is the initial feature vector of the medical image. The feature vector of the medical image; the feature vector of the medical image Reconstructing it into a 2×32 matrix, the feature vector matrix of the medical image is constructed using 16 matrices. ; The feature vector matrix of the medical image and the chaotic encrypted watermark are XORed bit by bit to embed the watermark information into the medical image.

2. The robust zero-watermarking method for medical images based on Residual-DenseNet according to claim 1, characterized in that, Also includes: While embedding watermark information into the medical image, a binary logic key sequence is generated; The trained Residual-DenseNet network is used to extract preliminary feature vectors from the medical image under test; The preliminary feature vector of the medical image to be tested is binarized by mean to generate the feature vector matrix of the medical image to be tested; The feature vector matrix of the medical image to be tested and the binary logic key sequence are XORed bit by bit to extract the encrypted watermark.

3. The robust zero-watermarking method for medical images based on Residual-DenseNet according to claim 2, characterized in that, After extracting the encrypted watermark, the process also includes: A chaotic sequence is generated by a Logistic mapping, and the chaotic sequence is binarized to generate a binary sequence. An encryption matrix is ​​constructed using the binary sequence, and the encryption matrix and the extracted encrypted watermark are XORed bit by bit to obtain the decrypted watermark.

4. The robust zero-watermarking method for medical images based on Residual-DenseNet according to claim 1, characterized in that, Before performing a bitwise XOR operation on the feature vector matrix of the medical image and the chaotic encrypted watermark, the method further includes: Logistic chaotic encryption is applied to the original watermark to obtain a chaotic encrypted watermark.

5. The robust zero-watermarking method for medical images based on Residual-DenseNet according to claim 4, characterized in that, Logistic chaotic encryption is applied to the original watermark to obtain a chaotic encrypted watermark, including: A chaotic sequence is generated by a Logistic mapping, and the chaotic sequence is binarized to generate a binary sequence. An encryption matrix is ​​constructed using the binary sequence, and the encryption matrix and the original watermark are XORed bit by bit to obtain a chaotic encrypted watermark.

6. The robust zero-watermarking method for medical images based on Residual-DenseNet according to claim 1, characterized in that, Before building the Residual-DenseNet network model, the following is also included: The original sample image is processed using 2D-DCT, and a portion of the coefficients are selected from the transform coefficients to form a preliminary feature vector of length 64. The constructed preliminary feature vectors are used as labels for the original sample image and the corresponding attacked sample image; The medical image data sample set is obtained based on the original sample image, the corresponding attacked sample image, and the label.

Images