Medical image zero watermark generation method and system based on noise reduction and convolution auto-encoder

Abstract

The invention specifically discloses a medical image zero-watermark generation method and system based on noise reduction and a convolutional auto-encoder (DCAE), and the method comprises the following steps: carrying out the binarization of a watermark image through employing a mean Hash algorithm (aHash), encrypting the binarized watermark image through employing a six-dimensional Lorenz-96 chaotic sequence, and storing a secret key; after an original medical image is collected, robust features of the image are obtained by using a pre-trained DCAE, and image features are binarized by using an aHash algorithm; and finally, carrying out XOR operation on the binarized features and the encrypted watermarks to generate zero watermarks. By adopting the technical scheme, the strong robustness feature capable of reconstructing the de-noised image can be automatically extracted from the image containing the noise in an end-to-end manner, and the technical problems that an existing medical image zero-watermark scheme is relatively weak in complex and combined attack resistance, and the generalization performance still needs to be improved in a real application scene are solved. And under the condition that the visual value of the medical image is not influenced, the requirements of privacy protection and information traceability of the medical image are effectively met.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, and relates to a method and system for generating zero watermarks for medical images based on noise reduction and convolutional autoencoders. Background Technology

[0002] Medical images are crucial in the medical process. However, while the internet and digital technologies are transforming the healthcare industry, medical images face the risks of leakage, tampering, and illegal use, directly threatening patient privacy and the security of medical images. Watermarking technology is a powerful tool for protecting digital content ownership and privacy, but embedded watermarking is generally unsuitable for high-fidelity images because the image content is altered after the watermark information is embedded.

[0003] To address this challenge, the concept of zero-watermarking was devised. Zero-watermarking is more suitable for medical images because it constructs a third-party watermark by extracting key features from the original data without modifying the original data. However, existing frequency domain and change domain-based zero-watermarking schemes typically require the manual design of complex mathematical function-based feature extractors and have low robustness to geometric attacks because they only extract shallow image features.

[0004] Neural networks have a multi-layered structure, enabling information to be passed layer by layer from lower to higher layers, thus automatically capturing complex structures and key information in images. Therefore, some researchers have conducted research on zero-watermarking schemes for image feature extraction using classification neural networks. However, classification neural networks are black-box models, resulting in poor interpretability of extracted medical image features. Most extracted features require further feature fusion, and current zero-watermarking schemes based on classification neural networks lack robustness against complex geometric attacks.

[0005] Compared to natural images such as portraits and landscapes, medical images often exhibit more complex and subtle tissue structures, and their signal-to-noise ratio is relatively low. Therefore, extracting robust features from medical images is more difficult, and zero-watermarking schemes designed for natural images show significantly reduced robustness when applied to medical image datasets. Furthermore, current zero-watermarking algorithms for medical images suffer from weak resistance to complex and combinatorial attacks, and their generalization performance in real-world applications needs improvement, thus failing to adequately protect medical images. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a method and system for generating zero-watermarked medical images based on denoising and convolutional autoencoders. This technical solution eliminates the need for manually designing complex mathematical functions and feature fusion operations, enabling end-to-end automated extraction of robust features from noisy images to reconstruct denoised images. It solves the technical problems of existing zero-watermarking schemes having weak resistance to complex and combinatorial attacks, and requiring further improvement in generalization performance in real-world applications.

[0007] To achieve the above objectives, the basic solution of this invention is: a method for generating zero-watermarks for medical images based on noise reduction and convolutional autoencoders, comprising the following steps:

[0008] A combination of a denoising autoencoder and a convolutional autoencoder is designed to obtain a DCAE network based on denoising and convolutional autoencoders.

[0009] Initialize the parameters of the DCAE network, and train the DCAE network using a pre-built medical image dataset to obtain a pre-trained DCAE network;

[0010] Collect or select any image to generate the original watermark image. Use the mean hash algorithm to binarize the original target watermark image to obtain the binarized target watermark image. ;

[0011] The original position index of the target watermark image is encrypted using the chaotic sequence sorted by the six-dimensional Lorenz-96 generator, resulting in the encrypted binary target watermark image. And save the key;

[0012] The target medical image is input into the pre-trained DCAE, and DCAE automatically extracts and outputs the feature matrix of the target medical image;

[0013] The mean hash algorithm is used to transform the feature matrix of the target medical image into a binary feature matrix, and then vectorization is performed to obtain the binary feature vector of the target medical image.

[0014] The encrypted binary target watermark image Perform an XOR operation with the binarized feature vector of the target medical image to generate a zero watermark, and save the generated zero watermark image to a qualified copyright authentication center.

[0015] The medical image from which the watermark is to be extracted is input into the pre-trained DCAE. DCAE automatically extracts and outputs the feature matrix of the medical image from which the watermark is to be extracted.

[0016] The features of the medical image from which the watermark is to be extracted are transformed into a binary feature matrix using the mean hash algorithm, and then vectorized to obtain the binary feature vector of the medical image from which the watermark is to be extracted.

[0017] The binarized feature vector of the medical image from which the watermark is to be extracted is XORed with the zero watermark stored in the copyright authentication center to obtain the extracted binarized encrypted watermark image. ;

[0018] Using the inverse permutation of the original position indices after sorting the chaotic sequence generated by six-dimensional Lorenz-96, the extracted binarized encrypted watermark image is decrypted. The decrypted binarized watermark image is obtained. .

[0019] The working principle and beneficial effects of this basic scheme are as follows: This technical scheme designs a DCAE network that combines a denoising autoencoder and a convolutional autoencoder. The DCAE network is trained using a constructed medical image dataset. Without requiring the manual design of complex mathematical functions and feature fusion operations, the DCAE network can automatically extract robust features from noisy images to reconstruct denoised images end-to-end. The features extracted from the target medical image by the trained DCAE network are then XORed with the target watermark image encrypted with six-dimensional Lorenz-96 to generate a zero watermark. This solves the technical problems of existing medical image zero-watermarking schemes having weak resistance to complex and combinatorial attacks, and their generalization performance in real-world applications still needs improvement. Without affecting the visual value of medical images, it effectively meets the needs of medical image privacy protection and information traceability.

[0020] The DCAE network is designed by combining a denoising autoencoder and a convolutional autoencoder. This network has a powerful automated image feature extraction capability and can extract robust image features from noisy medical images from end to end, thereby improving the robustness and generalization of the zero-watermarking scheme.

[0021] As an unsupervised generative network, an autoencoder compresses an image into a low-dimensional latent space and reconstructs it through an encoder-decoder structure. Its bottleneck layer design forces the network to filter redundant information and automatically extract key features. Combined with structures such as convolution, it can achieve hierarchical feature learning (from local edges to global semantics). At the same time, variations such as denoising and sparsity can enhance the robustness of features, thus possessing powerful image feature extraction capabilities and generalization.

[0022] Furthermore, the DCAE network is trained using the constructed medical image dataset to obtain a pre-trained DCAE network:

[0023] The medical image dataset consists of two sub-datasets: the attacked medical image dataset as input data for DCAE and the original medical image dataset as the reconstruction target for DCAE.

[0024] The attacked medical image dataset contains 62,600 medical images that have been subjected to attacks of varying intensities of Gaussian noise, JPEG compression, median filtering, rotation, scaling, translation, and cropping.

[0025] The original medical image dataset contains 62,600 original medical images that correspond to those in the attacked medical image dataset.

[0026] Furthermore, the binary target watermark image is encrypted using the original position index of the sorted chaotic sequence generated by six-dimensional Lorenz-96, resulting in the encrypted binary target watermark image. The method for saving the key is as follows:

[0027] Using a 6-dimensional Lorenz-96 hyperchaotic system as the chaotic sequence generator, its dynamic equation is:

[0028] ,

[0029] Where F is the external forcing term, which controls the chaotic behavior of the system. When F≥5, the system begins to exhibit chaotic characteristics. When F=8, the chaotic behavior is significant and the attractor structure is complex. xi is the initial value, representing the value of the i-th grid point.

[0030] F The initial value of xi is stored as a key. ;

[0031] Numerical integration of the 6-dimensional Lorenz-96 system was performed using the fourth-order Runge-Kutta method (RK4) to generate a chaotic sequence C of length equal to the total number of pixels in the target watermark image.

[0032] Sort the chaotic sequence C in ascending order and record the position of each element in the original sequence during the sorting process to obtain the original position index sorted_indices;

[0033] The binarized target watermark image W is rearranged according to the original position index sorted_indices. Specifically, the encrypted binarized target watermark image... The i-th pixel is taken from the sorted_indices[i]-th pixel of the binarized target watermark image W. This process uses the randomness of the chaotic sequence to scramble the order of the watermark pixels and achieve encryption.

[0034] The binarized target watermark image is encrypted to obtain an encrypted binarized watermark image, and the key is saved for later use.

[0035] Furthermore, the target medical image is input into the pre-trained DCAE, and DCAE automatically extracts and outputs the feature matrix of the target medical image, as follows:

[0036] ,

[0037] Where I represents the target medical image, and F represents the feature matrix of the target medical image;

[0038] The medical image from which the watermark is to be extracted is input into the pre-trained DCAE. DCAE automatically extracts and outputs the feature matrix of the medical image from which the watermark is to be extracted, as follows:

[0039] ,

[0040] in, This refers to a medical image from which the watermark is to be extracted. This represents the feature matrix of the medical image from which the watermark is to be extracted.

[0041] The DCAE network excels in feature extraction, demonstrating significant advantages in robustness, generalization, and interpretability. Compared to traditional methods, DCAE eliminates the tedious process of manually designing feature extractors and avoids subsequent feature fusion. It accurately captures key features from the input image's intrinsic characteristics in an end-to-end automated manner, cleverly filtering out noise-related interference, thus greatly improving the robustness and generalization of the solution. Furthermore, the key features extracted by DCAE contain rich image information and have the ability to reconstruct the complete image. This reversible mapping from features to image provides a clear logical chain for the solution, significantly enhancing its interpretability.

[0042] Furthermore, using the mean hash algorithm, the feature matrix of the target medical image is transformed into a binary feature matrix, as follows:

[0043] ,

[0044] in, This represents the pixel value at position (x, y) of the feature matrix of the target medical image. The pixel value at position (x,y) represents the binarized feature matrix of the target medical image.

[0045] Using the mean hash algorithm, the feature matrix of the medical image from which the watermark is to be extracted is transformed into a binary feature matrix, as follows:

[0046] ,

[0047] Among them, among them, This represents the pixel value at position (x, y) of the feature matrix of the medical image from which the watermark is to be extracted. The pixel value at position (x,y) represents the binarized feature matrix of the medical image from which the watermark is to be extracted.

[0048] Generating binary sequences using the mean hash algorithm is simple.

[0049] Furthermore, the encrypted binarized target watermark image is XORed with the binarized feature vector of the target medical image to generate a zero watermark, as follows:

[0050] ,

[0051] Where T represents the generated zero-watermark image, ZOR represents the XOR operation function, and V represents the binarized feature vector of the target medical image. This represents the encrypted, binarized target watermark image.

[0052] The combination of DCAE network and zero-watermarking technology improves the ability of medical image zero-watermarking schemes to resist complex and combined attacks, as well as their generalization performance in real-world application scenarios.

[0053] Furthermore, the binarized feature vector of the medical image from which the watermark is to be extracted is XORed with the zero watermark stored in the copyright authentication center to obtain the extracted binarized encrypted watermark image. ,for:

[0054] ,

[0055] in, Let T represent the binarized feature vector of the medical image from which the watermark is to be extracted, T represent the generated zero-watermark image, and ZOR represent the XOR operation function. This represents the extracted, binarized encrypted watermark image.

[0056] Use the key to obtain the original index position of the chaotic sequence, and decrypt the extracted watermark image. The watermarked image was obtained after the decryption. This provides guidelines for the secure protection of copyright and privacy information in medical images.

[0057] Furthermore, by obtaining the inverse permutation index of the sorted position index of the six-dimensional Lorenz-96 chaotic sequence, the extracted binary watermark image is decrypted, resulting in the decrypted binary watermark image, specifically:

[0058] First, the original position indices `sorted_indices` of the chaotic sequence generated by the six-dimensional Lorenz-96 algorithm are re-sorted to obtain its inverted indices `decrypted_indices`. Then, the extracted binarized encrypted watermark image is... Rearrange according to decrypted_indices to obtain the decrypted binarized watermark image. Specifically, the decrypted binarized watermark image. The i-th pixel is taken from the extracted binarized encrypted watermark image. The decrypted_indices[i]th pixel.

[0059] Furthermore, the binarized target watermark image is calculated. Compared with the decrypted binarized watermark image The normalized coefficient NC value between them is as follows:

[0060] ,

[0061] Where N represents the length and width of the watermark image. and The values ​​at positions (i, j) represent the pixel values ​​of the binarized target watermark image and the extracted and decrypted binarized watermark image, respectively. The NC value is an indicator of the similarity between the binarized target watermark image and the extracted and decrypted binarized watermark image. Generally, when the calculated NC value is greater than 0.8, the two watermark images have a strong similarity. This study further demonstrates that watermark information can be successfully extracted from medical images from which watermarks are to be extracted, and their ownership can be determined.

[0062] This invention also provides a medical image zero-watermark generation system based on noise reduction and convolutional autoencoder, including a data acquisition unit and a processing unit. The data acquisition unit is used to acquire a target watermark image, a target medical image, and a medical image from which the watermark is to be extracted. ;

[0063] The input terminal of the processing unit is connected to the output terminal of the data acquisition unit. The processing unit executes the method described in this invention to achieve zero-watermark generation of medical images.

[0064] This system utilizes data acquisition and processing units to extract key attributes from medical images and ignore noise-related information, thereby improving the robustness of the solution and enhancing the interpretability and credibility of the zero-watermark solution.

[0065] Furthermore, the processing unit is equipped with a noise-reducing convolutional autoencoder (DCAE), which includes an encoder module and a decoder module. In the encoder module, an original medical image of size 128×128×1 is compressed into image features of size 8×8×1 by alternately executing 6 convolutional layers and 4 max pooling layers.

[0066] In the decoder module, an image feature of size 8×8×1 is decoded into an output image of size 128×128×1 by alternately performing 6 deconvolution layers and 3 nearest neighbor upsampling operations.

[0067] The Deep Convolutional Autoencoder (DCAE) combines the advantages of denoising autoencoders and convolutional autoencoders, extracting features that exhibit superior performance, demonstrating stronger robustness, generalization, and interpretability. This network is trained unsupervised, requiring no image labels, no manually designed feature extractors, and no additional feature fusion operations. It can automatically mine and extract key features from the input image in an end-to-end manner, effectively ignoring noise-related redundant information. This characteristic significantly improves the robustness and generalization of the entire solution. Furthermore, the key features extracted by DCAE possess powerful image reconstruction capabilities, enabling the reconstruction of the original image based on these features. This provides an intuitive basis for the solution, further enhancing its interpretability. Attached Figure Description

[0068] Figure 1 This is a flowchart illustrating the medical image zero-watermark generation method based on noise reduction and convolutional autoencoder of the present invention.

[0069] Figure 2 This is a schematic diagram of the structure of the noise reduction and convolutional autoencoder (DCAE) of the medical image zero-watermark generation method based on noise reduction and convolutional autoencoder of the present invention.

[0070] Figure 3 These are six medical images representing different parts of the human body used in the evaluation experiment of the medical image zero-watermark generation method based on noise reduction and convolutional autoencoder of this invention.

[0071] Figure 4 This is the binarized target watermark image of the medical image zero-watermark generation method based on noise reduction and convolutional autoencoder of the present invention;

[0072] Figure 5 This is the encrypted binarized target watermark image of the medical image zero-watermark generation method based on noise reduction and convolutional autoencoder of the present invention.

[0073] Figure 6This invention relates to a method for generating zero watermarks for medical images based on noise reduction and convolutional autoencoders. The method consists of six medical images with watermarks to be extracted after undergoing a 25% rotation attack, and the watermarked images after extraction and decryption.

[0074] Figure 7 This invention relates to a method for generating zero-watermark medical images based on noise reduction and convolutional autoencoders. The method involves six medical images subjected to attacks of 0.125x and 8x magnification, with the watermarks to be extracted, along with the extracted and decrypted watermark images.

[0075] Figure 8 This invention relates to a method for generating zero-watermark medical images based on noise reduction and convolutional autoencoders. The method involves six medical images to be watermarked after undergoing a 30% translation attack, along with the extracted and decrypted watermarked images.

[0076] Figure 9 This invention relates to a method for generating zero-watermark medical images based on noise reduction and convolutional autoencoders. The method comprises six medical images, each subjected to a 25% cropping attack, with the watermark to be extracted, along with the extracted and decrypted watermarked images. Detailed Implementation

[0077] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0078] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0079] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0080] Traditional zero-watermarking schemes typically rely on manual feature extraction using conventional mathematical techniques. Most of these schemes require further feature fusion, have weak resistance to complex geometric attacks and combinatorial attacks, and their generalization performance in real-world applications still needs improvement, thus failing to adequately protect medical images.

[0081] Given the advantages of autoencoders in image feature extraction, this invention discloses a method for generating zero-watermarks for medical images based on denoising and convolutional autoencoders, aiming to enhance the robustness and generalization of zero-watermarking schemes for medical images. Figure 1 As shown, the method for generating zero-watermarks for medical images based on denoising and convolutional autoencoders (DCAEs) includes the following steps:

[0082] A design combining denoising autoencoders and convolutional autoencoders (specifically, using convolutional layers to replace the fully connected layers of traditional autoencoders and incorporating the denoising logic of denoising autoencoders; during training, using noisy data as input data to the network, for example, adding noise (such as Gaussian noise) to the input medical images during training, allowing the model to learn to recover the original image from the noisy input, thereby obtaining robustness to noise) yields a DCAE network based on denoising and convolutional autoencoders;

[0083] Initialize the parameters of the DCAE network, and train the DCAE network using a pre-built medical image dataset to obtain a pre-trained DCAE network;

[0084] Collect or select any image to generate the original watermark image (save the watermark information as an image format, which can be done in any way (system built-in function, office software, online tools, or programming language); the content in the watermark image is customizable). Binarize the original target watermark image using the mean hash algorithm (the following steps can be used: convert the original target watermark image to a grayscale image, calculate the average grayscale value of all pixels, iterate through the grayscale value of each pixel, and binarize the pixel according to the average grayscale value; mark a grayscale value greater than the average value as 1, otherwise mark it as 0), to obtain the binarized target watermark image. (like Figure 4 (as shown)

[0085] The original position index of the sorted chaotic sequence generated by six-dimensional Lorenz-96 is used to encrypt the binarized target watermark image, resulting in the encrypted binarized target watermark image. (like Figure 5 (as shown), and save the key;

[0086] The target medical image is input into the pre-trained DCAE, and DCAE automatically extracts and outputs the feature matrix of the target medical image;

[0087] The mean hash algorithm is used to transform the feature matrix of the target medical image into a binary feature matrix, and then vectorization is performed to obtain the binary feature vector of the target medical image.

[0088] The encrypted binary target watermark image Perform an XOR operation with the binarized feature vector of the target medical image to generate a zero watermark, and save the generated zero watermark image to a qualified copyright authentication center.

[0089] The medical image from which the watermark is to be extracted is input into the pre-trained DCAE. DCAE automatically extracts and outputs the feature matrix of the medical image from which the watermark is to be extracted.

[0090] The features of the medical image from which the watermark is to be extracted are transformed into a binary feature matrix using the mean hash algorithm, and then vectorized to obtain the binary feature vector of the medical image from which the watermark is to be extracted.

[0091] The binarized feature vector of the medical image from which the watermark is to be extracted is XORed with the zero watermark stored in the copyright authentication center to obtain the extracted binarized encrypted watermark image. ;

[0092] Using the inverse permutation of the original position indices after sorting the chaotic sequence generated by six-dimensional Lorenz-96, the extracted binarized encrypted watermark image is decrypted. The decrypted binarized watermark image is obtained. .

[0093] By using a trained DCAE network to extract features from a target medical image and performing an XOR operation with the target watermark image encrypted with six-dimensional Lorenz-96 to generate a zero watermark, guidelines can be provided for the secure protection of copyright and privacy information in medical images.

[0094] and These represent the target medical image and the binarized target watermark image, respectively. and These represent the feature matrix extracted from the target medical image and the generated binarized feature vector, respectively.

[0095] The DCAE network is designed by combining a denoising autoencoder and a convolutional autoencoder. This network has a powerful automated image feature extraction capability and can extract robust image features from noisy medical images from end to end, thereby improving the robustness and generalization of the zero-watermarking scheme.

[0096] In a preferred embodiment of the present invention, the DCAE network is trained using a constructed medical image dataset to obtain a pre-trained DCAE network:

[0097] The medical image dataset consists of two sub-datasets: the attacked medical image dataset as input data for DCAE and the original medical image dataset as the reconstruction target for DCAE.

[0098] The attacked medical image dataset contains 62,600 medical images that have been subjected to attacks of varying intensities of Gaussian noise, JPEG compression, median filtering, rotation, scaling, translation, and cropping.

[0099] The original medical image dataset contains 62,600 original medical images corresponding to those in the attacked medical image dataset (training set, used to train the network). The "target medical images" are used as the carriers for embedding the "target watermark images." "Target medical images" and "target watermark images" are two different concepts.

[0100] In a preferred embodiment of the present invention, the method for encrypting the binarized target watermark image using the original position index of the sorted chaotic sequence generated by six-dimensional Lorenz-96, and obtaining the encrypted binarized target watermark image, and storing the key is as follows:

[0101] Using a 6-dimensional Lorenz-96 hyperchaotic system as the chaotic sequence generator, its dynamic equation is:

[0102] ,

[0103] Where F is the external forcing term, which controls the chaotic behavior of the system. When F≥5, the system begins to exhibit chaotic characteristics. When F=8, the chaotic behavior is significant and the attractor structure is complex. xi is the initial value, representing the value of the i-th grid point.

[0104] F The initial value of xi is stored as a key. ;

[0105] Numerical integration of the 6-dimensional Lorenz-96 system was performed using the fourth-order Runge-Kutta method (RK4) to generate a chaotic sequence C of length equal to the total number of pixels in the target watermark image.

[0106] Sort the chaotic sequence C in ascending order and record the position of each element in the original sequence during the sorting process to obtain the original position index sorted_indices;

[0107] The binarized target watermark image W is rearranged according to the original position index sorted_indices. Specifically, the encrypted binarized target watermark image... The i-th pixel is taken from the sorted_indices[i]-th pixel of the binarized target watermark image W. This process uses the randomness of the chaotic sequence to scramble the order of the watermark pixels and achieve encryption.

[0108] In a preferred embodiment of the present invention, the target medical image is input into a pre-trained DCAE, and the DCAE automatically extracts and outputs the feature matrix of the target medical image, which is:

[0109] ,

[0110] Where I represents the target medical image, and F represents the feature matrix of the target medical image;

[0111] The medical image from which the watermark is to be extracted is input into the pre-trained DCAE. DCAE automatically extracts and outputs the feature matrix of the medical image from which the watermark is to be extracted, as follows:

[0112] ,

[0113] in, This refers to a medical image from which the watermark is to be extracted. This represents the feature matrix of the medical image from which the watermark is to be extracted.

[0114] The features extracted by the DCAE network are more robust. Noise reduction is essentially a process of enhancing feature learning. Neural networks can extract key attributes from medical images and ignore noise-related information, thereby improving the robustness of the solution.

[0115] DCAE is trained using unsupervised learning. This allows it to automatically discover the inherent organization and layout of information using a large amount of unlabeled data, avoiding the tedious process of manual feature design, achieving an end-to-end feature extraction process without the need for further feature fusion.

[0116] In a preferred embodiment of the present invention, the features of the target medical image are transformed into a binary feature matrix using the mean hash algorithm, as follows:

[0117] ,

[0118] in, This represents the pixel value at position (x, y) of the feature matrix of the target medical image. The pixel value at position (x,y) represents the binarized feature matrix of the target medical image.

[0119] Using the mean hash algorithm, the feature matrix of the medical image from which the watermark is to be extracted is transformed into a binary feature matrix, as follows:

[0120] ,

[0121] Among them, among them, This represents the pixel value at position (x, y) of the feature matrix of the medical image from which the watermark is to be extracted. The pixel value at position (x,y) represents the binarized feature matrix of the medical image from which the watermark is to be extracted.

[0122] In a preferred embodiment of the present invention, the encrypted binarized target watermark image is XORed with the binarized feature vector of the target medical image to generate a zero watermark, as follows:

[0123] ,

[0124] Where T represents the generated zero-watermark image, ZOR represents the XOR operation function, and V represents the binarized feature vector of the target medical image. This represents the encrypted, binarized target watermark image.

[0125] In a preferred embodiment of the present invention, the binarized feature vector of the medical image from which the watermark is to be extracted is XORed with the zero watermark stored in the copyright authentication center to obtain the extracted binarized encrypted watermark image. ,for:

[0126] ,

[0127] in, Let T represent the binarized feature vector of the medical image from which the watermark is to be extracted, T represent the generated zero-watermark image, and ZOR represent the XOR operation function. This represents the extracted, binarized encrypted watermark image.

[0128] In a preferred embodiment of the present invention, the reverse permutation index of the sorting position index of the chaotic sequence generated by the six-dimensional Lorenz-96 is obtained, and the extracted binarized watermark image is decrypted to obtain the decrypted binarized watermark image, specifically as follows:

[0129] First, the original position indices `sorted_indices` of the chaotic sequence generated by the six-dimensional Lorenz-96 algorithm are re-sorted to obtain its inverted indices `decrypted_indices`. Then, the extracted binarized encrypted watermark image is... Rearrange according to decrypted_indices to obtain the decrypted binarized watermark image. The decrypted binarized watermark image The i-th pixel is taken from the extracted binarized encrypted watermark image. The decrypted_indices[i]th pixel.

[0130] In a preferred embodiment of the present invention, the binarized target watermark image is calculated. Compared with the decrypted binarized watermark image The normalized coefficient NC value between them is as follows:

[0131] ,

[0132] Where N represents the length and width of the watermark image. and The values ​​at positions (i, j) represent the pixel values ​​of the binarized target watermark image and the extracted and decrypted binarized watermark image, respectively. The NC value is an indicator of the similarity between the binarized target watermark image and the extracted and decrypted binarized watermark image. Generally, when the calculated NC value is greater than 0.8, the two watermark images have a strong similarity. This study further demonstrates that watermark information can be successfully extracted from medical images from which watermarks are to be extracted, enabling image traceability and determination of ownership.

[0133] Gaussian noise attacks, JPEG compression attacks, and median filtering attacks are common image attack methods. All three attacks affect the overall image quality, leading to the loss of image detail and a decrease in overall sharpness. In the experiment, to ensure a comprehensive evaluation of the reliability of the scheme, six images (representing different body parts) were randomly selected as target medical images, such as... Figure 3 As shown, the binarized target watermark image is as follows: Figure 4 As shown, the NC value generally shows a decreasing trend, decreasing with increasing attack intensity. Under common attacks, the average NC value in this experiment was 0.9707. Under the three attacks—Gaussian noise attack parameter of 25%, JPEG compression attack parameter of 5%, and median filter attack parameter of 3×3 and 10 times—the edges and details of the medical image from which the watermark is to be extracted become very blurred, but the relatively clear watermark image can still be extracted. The solution of this invention can effectively resist common attacks and has good robustness.

[0134] Rotation attacks, scaling attacks, translation attacks, and cropping attacks are four types of geometric attacks on images. Rotation attacks involve rotating an image around a central point; scaling attacks alter the number of pixels or the pixel spacing; translation attacks move an image a certain distance in a specific direction; and cropping attacks crop a portion of the original image. These attacks change the original presentation of the image, thus affecting the observer's visual perception and potentially making the image difficult to recognize or understand, especially for information that relies on a specific visual layout or content.

[0135] like Figure 6 , Figure 7 , Figure 8 , Figure 9 As shown, when the rotation angle is less than 10%, the NC value is still higher than 0.9119. When the rotation angle reaches 25%, part of the image is missing, but the NC value is still higher than 0.8603, and the visual recognition of the extracted watermark image is completed. When the image is first magnified by 2 times and then by 0.5 times, most of the NC values ​​reach 1. When the image is first magnified by 0.125 times and then by 8 times, the visual effect of the medical image is significantly enhanced, but the image details become blurred, but the NC value is still higher than 0.9628, and the extracted watermark image is relatively clear.

[0136] As the degree of translation in medical images increases, the NC value tends to decrease. For example... Figure 8 As shown, when the translation attack intensity reaches 25%, many important parts of the medical image from which the watermark is to be extracted are missing, but the NC value is still higher than 0.8068, which is within the resolution range of the human eye. As the shearing attack intensity increases, the NC value gradually decreases. Figure 9 As shown, when the cropping attack intensity is 25%, the medical image from which the watermark is to be extracted loses a significant amount of crucial information, but the NC value is still higher than 0.9246. The results indicate that the proposed scheme effectively resists geometric attacks and possesses strong robustness.

[0137] This invention also provides a medical image zero-watermark generation system based on noise reduction and convolutional autoencoder, including a data acquisition unit and a processing unit. The data acquisition unit is used to acquire the target watermark image, the target medical image, and the medical image from which the watermark is to be extracted. .

[0138] The input end of the processing unit is connected to the output end of the data acquisition unit. The processing unit executes the method described in this invention to achieve zero-watermark generation of medical images.

[0139] In a preferred embodiment of the present invention, such as Figure 2 As shown, the processing unit is equipped with a noise reduction and convolutional autoencoder (DCAE). The noise reduction and convolutional autoencoder (DCAE) includes an encoder module and a decoder module. In the encoder module, a medical image with a size of 128×128×1 as network input is compressed into image features of a size of 8×8×1 by alternately executing 6 convolutional layers and 4 max pooling layers.

[0140] In the decoder module, the image features compressed to a size of 8×8×1 are decoded into an output image of a size of 128×128×1 by alternately executing 6 deconvolution layers and 3 nearest neighbor upsampling operations.

[0141] The ultimate goal of network training is to make the input image infinitely close to the reconstructed target, thereby obtaining image features with strong robustness. The detailed training process is as follows: The medical image training dataset consists of two subsets: an attacked medical image dataset as input data for DCAE and the original medical image dataset as the reconstruction target for DCAE. The attacked medical image dataset contains 62,600 medical images subjected to attacks of varying intensities of Gaussian noise, JPEG compression, median filtering, rotation, scaling, translation, and cropping; the original medical image dataset contains the corresponding original medical images from the attacked medical image dataset. The network uses the mean squared error function as the loss function, stochastic gradient descent as the optimizer, a learning rate of 0.001, and 130 training epochs.

[0142] The DCAE network layers are shown in Table 1:

[0143] Table 1 Parameters of DCAE Network Layer

[0144]

[0145] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for generating zero-watermarks for medical images based on noise reduction and convolutional autoencoders, characterized in that, Includes the following steps: A combination of denoising autoencoder and convolutional autoencoder design yields a DCAE network based on denoising and convolutional autoencoder. Initialize the parameters of the DCAE network, and train the DCAE network using a pre-built medical image dataset to obtain a pre-trained DCAE network; Collect or select any image to generate the original watermark image. Use the mean hash algorithm to binarize the original target watermark image to obtain the binarized target watermark image. ; The original position index of the target watermark image is encrypted using the chaotic sequence sorted by the six-dimensional Lorenz-96 generator, resulting in the encrypted binary target watermark image. And save the key; The target medical image is input into the pre-trained DCAE, and DCAE automatically extracts and outputs the feature matrix of the target medical image; The mean hash algorithm is used to transform the feature matrix of the target medical image into a binary feature matrix, and then vectorization is performed to obtain the binary feature vector of the target medical image. The encrypted binary target watermark image Perform an XOR operation with the binarized feature vector of the target medical image to generate a zero watermark, and save the generated zero watermark image to a qualified copyright authentication center. The medical image from which the watermark is to be extracted is input into the pre-trained DCAE. DCAE automatically extracts and outputs the feature matrix of the medical image from which the watermark is to be extracted. The features of the medical image from which the watermark is to be extracted are transformed into a binary feature matrix using the mean hash algorithm, and then vectorized to obtain the binary feature vector of the medical image from which the watermark is to be extracted. The binarized feature vector of the medical image from which the watermark is to be extracted is XORed with the zero watermark stored in the copyright authentication center to obtain the extracted binarized encrypted watermark image. ; Using the inverse permutation of the original position indices after sorting the chaotic sequence generated by six-dimensional Lorenz-96, the extracted binarized encrypted watermark image is decrypted. The decrypted binarized watermark image is obtained. .

2. The method for generating zero-watermark for medical images based on noise reduction and convolutional autoencoders according to claim 1, characterized in that, The DCAE network was trained using the constructed medical image dataset to obtain a pre-trained DCAE network: The medical image dataset consists of two sub-datasets: the attacked medical image dataset as input data for DCAE and the original medical image dataset as the reconstruction target for DCAE. The attacked medical image dataset contains 62,600 medical images that have been subjected to attacks of varying intensities of Gaussian noise, JPEG compression, median filtering, rotation, scaling, translation, and cropping. The original medical image dataset contains 62,600 original medical images that correspond to those in the attacked medical image dataset.

3. The method for generating zero-watermark for medical images based on noise reduction and convolutional autoencoders according to claim 1, characterized in that, The method for encrypting the binary target watermark image using the original position index of the sorted chaotic sequence generated by six-dimensional Lorenz-96, and saving the key, is as follows: Using a 6-dimensional Lorenz-96 hyperchaotic system as the chaotic sequence generator, its dynamic equation is: ， Where F is the external forcing term, which controls the chaotic behavior of the system. When F≥5, the system begins to exhibit chaotic characteristics. When F=8, the chaotic behavior is significant and the attractor structure is complex. xi is the initial value, representing the value of the i-th grid point. F The initial value of xi is stored as a key. ; Numerical integration of the 6-dimensional Lorenz-96 system was performed using the fourth-order Runge-Kutta method (RK4) to generate a chaotic sequence C of length equal to the total number of pixels in the target watermark image. Sort the chaotic sequence C in ascending order and record the position of each element in the original sequence during the sorting process to obtain the original position index sorted_indices; The binarized target watermark image W is rearranged based on the original location index sorted_indices, resulting in the encrypted binarized target watermark image. The i-th pixel is taken from the sorted_indices[i]-th pixel of the binarized target watermark image W. The randomness of the chaotic sequence is used to shuffle the order of the watermark pixels to achieve encryption.

4. The method for generating zero-watermark for medical images based on noise reduction and convolutional autoencoders according to claim 1, characterized in that, The target medical image is input into the pre-trained DCAE. The DCAE automatically extracts and outputs the feature matrix of the target medical image, which is: ， Where I represents the target medical image, and F represents the feature matrix of the target medical image; The medical image from which the watermark is to be extracted is input into the pre-trained DCAE. DCAE automatically extracts and outputs the feature matrix of the medical image from which the watermark is to be extracted, as follows: ， in, This refers to a medical image from which the watermark is to be extracted. This represents the feature matrix of the medical image from which the watermark is to be extracted.

5. The method for generating zero-watermark for medical images based on noise reduction and convolutional autoencoders according to claim 1, characterized in that, Using the mean hash algorithm, the feature matrix of the target medical image is transformed into a binary feature matrix, as follows: ， in, This represents the pixel value at position (x, y) of the feature matrix of the target medical image. The pixel value at position (x,y) represents the binarized feature matrix of the target medical image. Using the mean hash algorithm, the feature matrix of the medical image from which the watermark is to be extracted is transformed into a binary feature matrix, as follows: ， Among them, among them, This represents the pixel value at position (x, y) of the feature matrix of the medical image from which the watermark is to be extracted. The pixel value at position (x,y) represents the binarized feature matrix of the medical image from which the watermark is to be extracted.

6. The method for generating zero-watermark for medical images based on noise reduction and convolutional autoencoders according to claim 1, characterized in that, The encrypted, binarized target watermark image is XORed with the binarized feature vector of the target medical image to generate a zero watermark, as follows: ， Where T represents the generated zero-watermark image, ZOR represents the XOR operation function, and V represents the binarized feature vector of the target medical image. This represents the encrypted, binarized target watermark image.

7. The method for generating zero-watermark for medical images based on noise reduction and convolutional autoencoders according to claim 1, characterized in that, The binarized feature vector of the medical image from which the watermark is to be extracted is XORed with the zero watermark stored in the copyright authentication center to obtain the extracted binarized encrypted watermark image. ,for: ， in, Let T represent the binarized feature vector of the medical image from which the watermark is to be extracted, T represent the generated zero-watermark image, and ZOR represent the XOR operation function. This represents the extracted, binarized encrypted watermark image.

8. The method for generating zero-watermark for medical images based on noise reduction and convolutional autoencoder according to claim 1, characterized in that, Obtain the inverse permutation index of the sorted position index of the chaotic sequence generated by the six-dimensional Lorenz-96 algorithm, decrypt the extracted binarized watermark image, and obtain the decrypted binarized watermark image, specifically: First, the original position indices `sorted_indices` of the chaotic sequence generated by the six-dimensional Lorenz-96 are re-sorted to obtain its inverted index `decrypted_indices`. Then, the extracted binarized encrypted watermark image is... Rearrange according to decrypted_indices to obtain the decrypted binarized watermark image. ; Decrypted binarized watermark image The i-th pixel is taken from the extracted binarized encrypted watermark image. The decrypted_indices[i]th pixel.

9. The method for generating zero-watermark for medical images based on noise reduction and convolutional autoencoders according to claim 1, characterized in that, Calculate the binarized target watermark image and The normalized coefficient NC value between them is as follows: ， Where N represents the length and width of the watermark image. and represents the pixel values ​​at their respective (i, j) positions in the binarized target watermark image and the extracted and decrypted binarized watermark image; the NC value is an indicator that measures the similarity between the binarized target watermark image and the extracted and decrypted binarized watermark image.

10. A medical image zero-watermark generation system based on noise reduction and convolutional autoencoder, characterized in that, It includes a data acquisition unit and a processing unit. The data acquisition unit is used to acquire the target watermark image, the target medical image, and the medical image from which the watermark is to be extracted. The input terminal of the processing unit is connected to the output terminal of the data acquisition unit, and the processing unit executes the method described in any one of claims 1-9 to realize zero-watermark generation of medical images; The processing unit is equipped with a noise reduction and convolutional autoencoder (DCAE), which includes an encoder module and a decoder module. In the encoder module, an original medical image of size 128×128×1 is compressed into image features of size 8×8×1 by alternately executing 6 convolutional layers and 4 max pooling layers. In the decoder module, an image feature of size 8×8×1 is decoded into an output image of size 128×128×1 by alternately performing 6 deconvolution layers and 3 nearest neighbor upsampling operations.

Images