Robust watermark embedding and extraction method for medical images based on InceptionV3 deep learning network

Abstract

The invention discloses a medical image robust watermark embedding and extraction method based on the InceptionV3 deep learning network. The medical image is convoluted using the InceptionV3 network, and the obtained fully connected layer coefficients are subjected to global discrete cosine transform (DCT transform). The low-frequency coefficients are selected from the coefficient matrix obtained by the DCT transform as the visual feature vector of the medical image and binarized to obtain the feature vector of the target medical image. Unlike the traditional image feature extraction method, the feature does not need to be manually designed when extracting features. The image feature extraction is realized by deep convolution learning. The algorithm has stronger feature learning and expression capabilities. The InceptionV3 network is combined with the DCT transform, which improves the anti-geometric attack capability of the digital watermark and retains the advantages of the DCT in terms of strong anti-conventional attack capability and good robustness. The invention solves the technical problems that the existing medical image processing method based on digital watermark technology has poor anti-geometric attack performance and the robustness needs to be improved.

Description

Technical Field

[0001] The present invention relates to the field of image watermarking technology, and in particular to a medical image robust watermark embedding and extraction method based on the InceptionV3 deep learning network. Background Art

[0002] With the continuous innovation and development of information technology and computer network technology, telemedicine technology and medical imaging information technology have been greatly promoted and improved. Among them, digital medical images have become the main basis for doctors' diagnosis and play an important role in clinical medicine and scientific research. With the help of telemedicine technology, relevant personnel transmit medical imaging information through the network to seek high-quality and comprehensive diagnostic opinions from experts. However, there may be security issues such as information leakage, illegal tampering and theft during the transmission of medical images in the public network. In order to protect medical images, technical processing of medical images is required to ensure that medical images are not leaked, tampered with or stolen during the transmission and sharing process.

[0003] Digital watermarking technology is a common information hiding technology with the characteristics of invisibility and robustness. It embeds specific digital signals into multimedia data, which can ensure that the multimedia data can still reliably extract watermark information and clarify copyright after experiencing various attacks. It is widely used in multimedia data protection. However, medical images are different from ordinary multimedia images. Medical images are an important basis for experts to obtain patient pathological information and make medical diagnoses. Therefore, when performing digital watermark copyright protection, no changes are allowed to medical images. Zero watermark technology can extract image features and generate zero watermark signals without making any changes to medical images, so as to be well applied in the field of medical image transmission. At present, digital watermarking technology for medical images is still in the development stage, and its anti-geometric attack performance is poor, and its robustness performance needs to be improved. Summary of the invention

[0004] The present invention provides a medical image robust watermark embedding method and extraction method based on the InceptionV3 deep learning network, which are used to solve the technical problems that the existing medical image processing methods based on digital watermark technology have poor anti-geometric attack performance and the robust performance needs to be improved.

[0005] In view of this, the first aspect of the present invention provides a medical image robust watermark embedding method based on the InceptionV3 deep learning network, comprising:

[0006] Select the medical image as the input image of the InceptionV3 network, use the InceptionV3 network to perform convolution operation on the medical image, take the output of the fully connected layer of the InceptionV3 network, and obtain the fully connected layer coefficient;

[0007] Perform global discrete cosine transform (DCT transform) on the fully connected layer coefficients, then select low-frequency coefficients from the coefficient matrix obtained by DCT transform as the visual feature vector of the medical image, perform binarization on the visual feature vector, and obtain the feature vector of the target medical image;

[0008] Perform chaotic scrambling encryption on the original target watermark to obtain a chaotic scrambled encrypted watermark;

[0009] The feature vector of the target medical image is dimensionally upgraded to obtain a feature matrix of the target medical image;

[0010] According to the obtained target medical image feature matrix and the chaotic scrambled encryption watermark, an XOR operation is performed bit by bit to embed the original target watermark information into the target medical image, and a binary logic key sequence is obtained at the same time.

[0011] Optionally, the original target watermark is encrypted by chaotic scrambling to obtain a chaotic scrambled encrypted watermark, including:

[0012] Generate chaotic sequence through Logistic Map;

[0013] Binarize the generated chaotic sequence to obtain a binary encryption matrix;

[0014] The binary encryption matrix and the converted original target watermark are operated bit by bit to obtain a chaotically scrambled encryption watermark.

[0015] Optionally, before selecting the medical image as the input image of the InceptionV3 network, the following steps are also included:

[0016] The image size of medical images was resized to 299×299.

[0017] Optionally, the target medical image feature matrix size is 32×32.

[0018] Optionally, it also includes:

[0019] The binary logic key sequence is stored in a third party, wherein the binary logic key sequence is bound to the corresponding medical image embedded with the original target watermark information.

[0020] The second aspect of the present invention provides a medical image robust watermark extraction method based on the InceptionV3 deep learning network, comprising:

[0021] Input the medical image to be watermarked into the InceptionV3 network, use the InceptionV3 network to perform convolution operation on the medical image to be watermarked, take the output of the fully connected layer of the InceptionV3 network, and obtain the fully connected layer coefficient;

[0022] The fully connected layer coefficients are subjected to global discrete cosine transform (DCT transform), and then the low-frequency coefficients are selected from the coefficient matrix obtained by the DCT transform as the visual feature vector of the medical image to be extracted watermark information, and the visual feature vector is binarized to obtain the feature vector of the medical image to be extracted watermark information;

[0023] The feature vector of the medical image to be watermarked is dimensionally increased to obtain a 32×32 feature matrix of the medical image to be watermarked;

[0024] Retrieving the medical image feature matrix of the watermark information to be extracted and the logic key obtained by a third party, performing an XOR operation bit by bit, and extracting the encrypted watermark;

[0025] Decrypt the encrypted watermark to obtain the restored watermark information.

[0026] Optionally, the encrypted watermark is decrypted to obtain restored watermark information, including:

[0027] Generate chaotic sequence through Logistic Map;

[0028] Binarize the generated chaotic sequence to obtain a binary chaotic matrix;

[0029] The binary chaotic matrix and the encrypted watermark are XORed bit by bit to obtain the restored watermark information.

[0030] Optionally, it also includes:

[0031] The original target watermark is obtained, and the normalized correlation coefficient of the original target watermark and the restored watermark is calculated to determine the ownership of the target medical image and the embedded watermark information.

[0032] It can be seen from the above technical solutions that the embodiments of the present invention have the following advantages:

[0033] The present invention provides a medical image robust watermark embedding method based on the InceptionV3 deep learning network, uses the InceptionV3 network to perform convolution operation on the medical image, takes the output of the fully connected layer of the InceptionV3 network, obtains the fully connected layer coefficient, performs global discrete cosine transform (DCT transform) on the fully connected layer coefficient, then selects low-frequency coefficients from the coefficient matrix obtained by the DCT transform as the visual feature vector of the medical image, performs binarization processing on the visual feature vector, obtains the feature vector of the target medical image, and is different from the traditional image feature extraction method. When extracting features, it is not necessary to manually design features, and image feature extraction is realized by deep convolution learning. The algorithm has stronger feature learning and expression capabilities, and the InceptionV3 network is combined with the DCT transform, which improves the anti-geometric attack capability of the digital watermark, and retains the advantages of the DCT anti-conventional attack capability and good robustness. The existing medical image processing method based on digital watermarking technology has poor anti-geometric attack performance and the robustness needs to be improved. Technical problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0035] Figure 1 A schematic diagram of a flow chart of a medical image robust watermark embedding method based on the InceptionV3 deep learning network provided in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the hierarchical structure of the InceptionV3 network provided in an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of a medical image provided in an embodiment of the present invention;

[0038] Figure 4 A schematic diagram of an original target watermark provided in an embodiment of the present invention;

[0039] Figure 5 A schematic diagram of a chaotically scrambled encrypted watermark provided in an embodiment of the present invention;

[0040] Figure 6 A schematic diagram of a watermark extracted without interference provided in an embodiment of the present invention;

[0041] Figure 7A schematic diagram of a medical image with a Gaussian noise interference intensity of 16% provided in an embodiment of the present invention;

[0042] Figure 8 A schematic diagram of a watermark extracted when the Gaussian noise interference intensity is 16% provided in an embodiment of the present invention;

[0043] Fig. 9 A schematic diagram of a medical image with a JPEG compression quality of 21% provided in an embodiment of the present invention;

[0044] Fig.10 A schematic diagram of a watermark extracted during JPEG compression with a compression quality of 21% provided in an embodiment of the present invention;

[0045] Fig.11 This is a schematic diagram of a medical image with a window size of 3x3 and 15 filtering times after median filtering provided in an embodiment of the present invention;

[0046] Fig.12 This is a schematic diagram of a watermark extracted after a window size of 3x3 and median filtering for 15 times provided in an embodiment of the present invention;

[0047] Fig.13 This is a schematic diagram of a medical image with a window size of 5x5 and a filtering number of 15 times after median filtering provided in an embodiment of the present invention;

[0048] Fig.14 This is a schematic diagram of a watermark extracted after a window size of 5x5 and 15 median filtering according to an embodiment of the present invention;

[0049] Fig.15 This is a schematic diagram of a medical image rotated 40° clockwise provided in an embodiment of the present invention;

[0050] Fig.16 A schematic diagram of a watermark extracted when rotating 40° clockwise provided in an embodiment of the present invention;

[0051] Fig.17 A schematic diagram of a medical image rotated 40° counterclockwise provided in an embodiment of the present invention;

[0052] Fig.18 A schematic diagram of a watermark extracted when rotating 40° counterclockwise provided in an embodiment of the present invention;

[0053] Fig.19 A schematic diagram of a medical image zoomed 1.2 times provided in an embodiment of the present invention;

[0054] Fig. 20 A schematic diagram of a watermark extracted when zooming in 1.2 times provided in an embodiment of the present invention;

[0055] Fig.21 A schematic diagram of a medical image with a horizontal left shift of 28% provided in an embodiment of the present invention;

[0056] Fig. 22 A schematic diagram of a watermark extracted when the horizontal shift is 28% to the left provided in an embodiment of the present invention;

[0057] Fig.23 A schematic diagram of a medical image vertically shifted downward by 35% provided in an embodiment of the present invention;

[0058] Fig.24 A schematic diagram of a watermark extracted when vertically shifted downward by 35% provided in an embodiment of the present invention;

[0059] Fig.25 A schematic diagram of a medical image sheared 32% along the X-axis provided in an embodiment of the present invention;

[0060] Fig.26 A schematic diagram of a watermark extracted when shearing 32% along the X-axis provided in an embodiment of the present invention;

[0061] Fig. 27 A schematic diagram of a medical image sheared 19% along the Y axis provided in an embodiment of the present invention;

[0062] Fig.28 A schematic diagram of a watermark extracted when shearing along the Y axis by 19% provided in an embodiment of the present invention;

[0063] Fig.29 A flowchart of a medical image robust watermark extraction method based on the InceptionV3 deep learning network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] For easier understanding, see Figure 1 The present invention provides an embodiment of a medical image robust watermark embedding method based on the InceptionV3 deep learning network, comprising:

[0066] Step 101: Select a medical image as an input image of the InceptionV3 network, use the InceptionV3 network to perform a convolution operation on the medical image, take the output of the fully connected layer of the InceptionV3 network, and obtain the fully connected layer coefficient.

[0067] The InceptionV3 network reduces the number of parameters, improves the operation speed and alleviates the overfitting phenomenon by splitting the two-dimensional convolution layer into two one-dimensional convolution layers. The InceptionV3 network is selected as the experimental network structure. The hierarchical structure of the InceptionV3 network is as follows: Figure 2 As shown, at the same time, the Inception V3 network adds a layer of nonlinear extended model expression ability. This asymmetric convolution structure splitting is better than symmetrically splitting several identical small convolution kernels, and can extract more and richer spatial features, increasing the diversity of features. In the embodiment of the present invention, the Inception V3 network inputs an RGB image, and the final output is the probability of dividing the input image into 1000 image categories. A 512×512 medical image is selected as the original medical image. Because the Inception V3 network requires the image input size to be 299×299, we need to set the image size to 299×299 before inputting the 512×512 original medical image, and use the InceptionV3 network to perform convolution operations on the medical image, take the output of the fully connected layer of the InceptionV3 network, and obtain the fully connected layer coefficient;

[0068] Step 102: Perform a global discrete cosine transform (DCT transform) on the fully connected layer coefficients, then select low-frequency coefficients from the coefficient matrix obtained by the DCT transform as the visual feature vector of the medical image, binarize the visual feature vector, and obtain the feature vector of the target medical image.

[0069] Different images have large visual differences, but the image features will not change much after a small transformation. Finding the visual feature vector that reflects the geometric characteristics of the medical image and associating the watermark to be embedded with the image feature vector can better solve the problem of watermark robustness.

[0070] According to the human visual characteristics (HVS), low-intermediate frequency signals have a greater impact on human vision and represent the main features of medical images. Therefore, the visual feature vector of the selected medical image is the low-intermediate frequency coefficient. The length of the feature vector is related to the size of the original medical image, the correlation between medical images, and the security of encryption. Generally speaking, the smaller the length, the greater the correlation.

[0071] Therefore, the fully connected layer coefficients are subjected to global discrete cosine transform (DCT transform). After the signal passes through 1D-DCT, its main energy will be concentrated in the low-frequency part. The fully connected layer coefficients are subjected to DCT transform to obtain a set of sequences. The low and medium frequency coefficients of the sequence are then processed using a hash function to obtain a binary sequence, which is used as the feature vector of the target medical image.

[0072] Step 103: Perform chaotic scrambling encryption on the original target watermark to obtain a chaotic scrambled encrypted watermark.

[0073] Select the original target watermark to be added, perform chaotic scrambling encryption on the original target watermark, and obtain a chaotic scrambled encrypted watermark. Specifically, the process includes:

[0074] A 32-bit chaotic sequence is generated by Logistic Map, where the initial value x0 of the chaotic coefficient is set to 0.2, the growth parameter is 4, and the number of iterations is 32.

[0075] The generated chaotic sequence is binarized, and the chaotic sequence values ​​greater than 0.5 are "1", and the rest are "0", thereby generating a binary encryption matrix;

[0076] The binary encryption matrix and the converted original target watermark (binary watermark W(i, j)) are operated bit by bit to obtain the chaotically scrambled encryption watermark BW(i, j).

[0077] Chaos is a seemingly irregular movement, which refers to a random process that occurs in a deterministic system. With initial values ​​and parameters, this chaotic system can be generated.

[0078] Logistic Map is one of the chaotic mapping methods. It is a simple dynamic nonlinear regression method with chaotic behavior. Its mathematical definition can be expressed as:

[0079] X k+1 =μ·x k .(1-x k )

[0080] Among them, x k ∈(0, 1), 0<μ≤4. When 3.5699456<μ≤4, the Logistic map enters a chaotic state, and the Logistic chaotic sequence can be used as an ideal key sequence.

[0081] Step 104: Up-dimension the feature vector of the target medical image to obtain a feature matrix of the target medical image.

[0082] The feature vector of the target medical image is upgraded to obtain a target medical image feature matrix of size 32×32.

[0083] Step 105: performing an XOR operation bit by bit based on the obtained target medical image feature matrix and the chaotically scrambled encrypted watermark to embed the original target watermark information into the target medical image, and simultaneously obtain a binary logic key sequence.

[0084] By performing bit-by-bit XOR operation on the target medical image feature matrix Vm(i, j) and the chaotic scrambled encryption watermark BW(i, j), the watermark BW(i, j) can be embedded into the medical image, and the binary logic key sequence Key(i, j) can be obtained at the same time.

[0085]

[0086] Keep Key(i, j), store the binary logic key sequence in a third party, and bind the binary logic key sequence to the corresponding medical image embedded with the original target watermark information for subsequent watermark extraction. By applying for Key(i, j) as a key to a third party, the ownership and use rights of the medical image can be obtained, thereby achieving the purpose of protecting the medical image.

[0087] The medical image robust watermark embedding method based on the InceptionV3 deep learning network provided by the embodiment of the present invention uses the InceptionV3 network to perform convolution operation on the medical image, takes the output of the fully connected layer of the InceptionV3 network, obtains the fully connected layer coefficients, performs global discrete cosine transform (DCT transform) on the fully connected layer coefficients, and then selects low-frequency coefficients from the coefficient matrix obtained by the DCT transform as the visual feature vector of the medical image, binarizes the visual feature vector, and obtains the feature vector of the target medical image. Unlike the traditional image feature extraction method, it does not require manual feature design when extracting features, and realizes image feature extraction through deep convolution learning. The algorithm has stronger feature learning and expression capabilities. The InceptionV3 network is combined with the DCT transform, which improves the digital watermark's ability to resist geometric attacks and retains the advantages of DCT's strong ability to resist conventional attacks and good robustness. The existing medical image processing method based on digital watermarking technology has poor performance in resisting geometric attacks and the robustness needs to be improved.

[0088] In one embodiment, in step 101, the InceptionV3 network hierarchy is as follows: Figure 2 Its network structure is shown in Table 1.

[0089] Table 1. InceptionV3 network structure

[0090] type kemel size / step size (or comment) Input size convolution 3x3 / 2 299x299x3 convolution 3x3 / 1 149x149x32 convolution 3x3 / 1 147x147x32 Pooling 3x3 / 2 147x147x64 convolution 1x1 / 1 73x73x64 convolution 3x3 / 1 73x73x80 Pooling 3x3 / 2 71x71x192 Inception module group 3 Inception Modules 35x35x192 Inception module group 4 Inception Modules 17x17x288 Inception module group 2 Inception Modules 8x8x768 Pooling 8x8 8x8x2048 Linear Logits 1x1x2048 Softmax Classification Output 1x1x1000

[0091] The object of the experimental test is a 512×512 brain medical image as the original medical image. The original medical image is as follows Figure 3 As shown, a 32×32 binary image is selected as the original watermark. The original watermark is as follows Figure 4 As shown, the initial value x0 of the chaotic coefficient is set to 0.2, the growth parameter is 4, the number of iterations is 32, and the original watermark is chaotically encrypted. The encrypted watermark is as follows Figure 5 As shown;

[0092] First, the original image is feature extracted using the Inception V3 network, and the output of the feature extraction network is transformed by DCT. Considering the robustness and the capacity of one-time watermark embedding, the first 32 coefficients are selected to form the feature vector, and the hash function is used to complete the binarization operation.

[0093] When confirming the attacked medical image at a later stage, the features of the attacked medical image are also extracted in the same way. After the extracted restored watermark W′(i, j), the similarity of the two watermarks is determined by calculating the correlation coefficient NC between the original watermark and the extracted restored watermark W′(i, j), thereby determining the robustness of the algorithm. The closer the NC value is to 1, the higher the similarity. Figure 6 This is the watermark extracted without interference. We can see that NC = 1.00, indicating that the embedded original watermark is accurately extracted. In addition, we use PSNR to indicate the degree of distortion of the image. The larger the PSNR value, the smaller the distortion of the image.

[0094] In one embodiment, the fully connected layer coefficients are subjected to global discrete cosine transform (DCT transform). After the signal passes through 1D-DCT, its main energy will be concentrated in the low-frequency part. The fully connected layer coefficients are subjected to DCT transform to obtain a set of sequences, and then the low- and medium-frequency coefficients of the set of sequences are processed using a hash function to obtain a binary sequence, which is used as the feature vector of the target medical image.

[0095] For a sequence of length N, the formula for 1D-DCT transform is as follows:

[0096]

[0097]

[0098] Among them, is the 1D-DCT transform coefficient of.

[0099] In order to better demonstrate the technical effect of the medical image robust watermark embedding method based on the InceptionV3 deep learning network provided in the embodiment of the present invention, a specific application example of the medical image robust watermark embedding method based on the InceptionV3 deep learning network is provided in the embodiment of the present invention:

[0100] This application example tests the ability of the medical image robust watermark embedding method based on the InceptionV3 deep learning network to resist conventional attacks and geometric attacks.

[0101] 1. Test of algorithm's ability to resist conventional attacks

[0102] (1) Adding Gaussian noise

[0103] Table 2 Watermark anti-Gaussian noise interference data

[0104] Noise intensity (%) 2 4 8 16 20 30 PSNR(dB) 19.12 16.28 13.51 10.83 10.08 8.81 NC 0.90 0.82 0.69 0.76 0.66 0.53

[0105] Table 2 is the experimental data of watermark anti-Gaussian noise interference. It can be seen from the table that when the Gaussian noise intensity is as high as 20%, the PSNR of the image after the attack drops to 10.08dB. At this time, the extracted watermark has a correlation coefficient of NC = 0.66, and the watermark can still be extracted more accurately. This shows that the invention can resist Gaussian noise.

[0106] like Figure 7 As shown in the figure, the medical image with Gaussian noise intensity of 16% is visually different from the original medical image.

[0107] like Figure 8 As shown, the watermark extracted when the Gaussian noise intensity is 16%, NC = 0.76.

[0108] (2) JPEG compression attack

[0109] Table 3 Experimental data of watermark anti-JPEG compression

[0110] Compression quality (%) 1 7 14 21 28 35 PSNR(dB) 24.88 28.62 31.00 32.37 33.27 33.96 NC 0.41 0.40 0.60 0.60 0.66 0.72

[0111] The image compression quality percentage is used as a parameter to perform JPEG compression on medical images; Table 3 shows the experimental data of watermark anti-JPEG compression. When the compression quality is 14%, the watermark can still be extracted, and NC = 0.60.

[0112] Fig. 9 For medical images with a compression quality of 21%;

[0113] Fig.10 The watermark extracted for compression quality of 21%, NC=0.60.

[0114] (3) Median filter attack

[0115] Table 4 Watermark anti-median filtering experimental data

[0116]

[0117] Table 4 shows the ability of medical image watermark to resist median filtering. It can be seen from the table that when the median filter parameter is 3x3 and the number of filtering repetitions is 35, the existence of the watermark can still be detected, NC=0.66.

[0118] Fig.11 This is a medical image with a median filter parameter of 3x3 and a filter repetition number of 15;

[0119] Fig.12 The watermark is extracted when the median filter parameter is 3x3 and the number of filter repetitions is 15. NC=0.60, and the watermark can be extracted.

[0120] Fig.13 The median filter parameter is 5x5 and the number of filter repetitions is 15 for medical images;

[0121] Fig.14 The watermark is extracted when the median filter parameter is 5x5 and the number of filter repetitions is 15. NC=0.54, and the watermark can still be extracted.

[0122] 2. Watermark anti-geometric attack capability test

[0123] (1) Spin attack

[0124] Table 5 Watermark anti-rotation attack experimental data

[0125] Rotation degree° -80° -40° -16° -8° 8° 16° 40° 80° PSNR(dB) 13.80 15.03 16.22 17.89 17.89 16.21 15.03 13.81 NC 0.69 0.72 0.60 0.57 0.71 0.64 0.90 0.82

[0126] Table 5 is the experimental data of watermark anti-rotation attack, in which negative is counterclockwise and positive is clockwise. From the table, we can see that when the image is rotated 80° clockwise, NC=0.82, and when it is rotated 80° counterclockwise, NC=0.69, and the watermark can still be extracted.

[0127] Fig.15 It is a medical image rotated 40° clockwise;

[0128] Fig.16 The watermark is extracted by rotating 40° clockwise, NC=0.90, and the watermark can be accurately extracted;

[0129] Fig.17 It is a medical image rotated 40° counter-clockwise;

[0130] Fig.18 The watermark is extracted by rotating 40° counterclockwise, NC=0.72, and the watermark can still be accurately extracted.

[0131] (2) Zoom attack

[0132] Table 6 Watermark anti-scaling attack experimental data

[0133] Scaling Factor 0.5 0.8 1.2 2.4 4.0 8.0 NC 0.81 0.81 0.91 0.81 0.81 0.81

[0134] Table 6 shows the experimental data of medical image watermark anti-scaling attack. It can be seen from the table that when the scaling factor is as small as 0.5, the correlation coefficient NC=0.81, and the watermark can be extracted.

[0135] Fig.19 is a medical image with a scaling factor of 1.2;

[0136] Fig. 20 This is the watermark extracted when the scaling factor is 1.2, NC = 0.91, and the watermark can be accurately extracted.

[0137] (3) Translation attack

[0138] Table 7 Watermark anti-translation transformation experimental data

[0139]

[0140]

[0141] Table 7 Experimental data of watermark anti-translation transformation. From the table, we know that when the image data moves 21% to the left horizontally, NC = 0.91; when the image data moves 21% to the right horizontally, NC = 0.70; when the image data moves 21% upward horizontally, NC = 0.72; when the image data moves 21% downward horizontally, NC = 0.52; the watermark can be accurately extracted from the images after the above attacks, so the watermark method has a strong ability to resist translation transformation.

[0142] Fig.21 This is the image after the medical image is horizontally shifted 28% to the left;

[0143] Fig. 22 The watermark is extracted after horizontally shifting 28% to the left. The watermark can be accurately extracted. At this time, NC=0.82;

[0144] Fig.23 It is the image after the medical image is vertically shifted down by 35%;

[0145] Fig.24 The watermark is extracted after vertically moving down 35%. The watermark can be accurately extracted. At this time, NC=0.77.

[0146] (4) Shear attack

[0147] Table 8 Watermark anti-cut attack experimental data

[0148]

[0149] Table 8 is the experimental data of watermark anti-shearing attack. It can be seen from the table that when the medical image is sheared along the coordinate axis X and the shearing amount is 32%, NC=0.82, the watermark can be extracted; when the medical image is sheared along the coordinate axis Y and the shearing amount is 19%, NC=0.75, the watermark can be extracted, indicating that the watermark algorithm has a strong ability to resist shearing attacks.

[0150] Fig.25 This is a medical image after 32% shearing along the X axis;

[0151] Fig.26 It is the watermark extracted after shearing 32% along the X axis, NC = 0.82, which can accurately extract the watermark;

[0152] Fig. 27 This is a medical image after 19% shearing along the Y axis;

[0153] Fig.28 The watermark is extracted after 19% shearing along the Y axis. NC = 0.75, which can accurately extract the watermark. For easier understanding, please refer to Fig.29 Corresponding to the medical image robust watermark embedding method based on the InceptionV3 deep learning network provided in the aforementioned embodiment, the present invention provides an embodiment of a medical image robust watermark extraction method based on the InceptionV3 deep learning network, comprising:

[0154] Step 201: Input the medical image to be watermarked into the InceptionV3 network, use the InceptionV3 network to perform a convolution operation on the medical image to be watermarked, take the output of the fully connected layer of the InceptionV3 network, and obtain the fully connected layer coefficient.

[0155] Step 202: Perform a global discrete cosine transform (DCT transform) on the fully connected layer coefficients, then select low-frequency coefficients from the coefficient matrix obtained by the DCT transform as the visual feature vector of the medical image from which the watermark information is to be extracted, perform binarization on the visual feature vector, and obtain the feature vector of the medical image from which the watermark information is to be extracted.

[0156] In an embodiment of the present invention, a method for obtaining a feature vector V′(i, j) of a medical image to extract watermark information is the same as the method for obtaining a feature vector V(i, j) of a medical image in the aforementioned medical image robust watermark embedding method based on the InceptionV3 deep learning network, and will not be repeated here.

[0157] Step 203: Up-dimension the feature vector of the medical image from which the watermark information is to be extracted, and obtain a 32×32 feature matrix V′m(i, j) of the medical image from which the watermark information is to be extracted.

[0158] Step 204: retrieve the medical image feature matrix V′m(i, j) of the watermark information to be extracted and perform bit-by-bit XOR operation on the logic key Key(i, j) obtained by a third party to extract the encrypted watermark BW′(i, j).

[0159] The encrypted watermark BW′(i, j) can be extracted by performing bit-by-bit XOR operation on the image feature matrix V′m(i, j) and the logic key Key(i, j) obtained by the third party.

[0160]

[0161] This algorithm only needs the key Key(i, j) when extracting the watermark, and does not require the original image to participate. Therefore, it is a zero watermark extraction algorithm.

[0162] Step 205: decrypt the encrypted watermark to obtain restored watermark information.

[0163] In the embodiment of the present invention, the method for obtaining the binary chaotic matrix K(n) of the medical image to be watermarked is the same as the method for obtaining the binary encrypted matrix K(n) of the medical image in the aforementioned medical image robust watermark embedding method based on the InceptionV3 deep learning network, and will not be described in detail here.

[0164] The binary chaotic matrix K(n) and the encrypted watermark BW′(i, j) are subjected to bit-wise XOR operation to obtain the decrypted and restored watermark W′(i, j).

[0165]

[0166] By calculating the correlation coefficient NC between W(i, j) and W′(i, j), the watermark information is determined and the ownership of the medical image is clarified.

[0167] In one embodiment, step 205 further includes:

[0168] Step 205: Obtain the original target watermark, calculate the normalized correlation coefficient between the original target watermark W(i, j) and the restored watermark W′(i, j), and determine the ownership of the target medical image and the embedded watermark information.

[0169] Because this algorithm forms a direct relationship between the feature vectors of the watermark and the image through XOR operation, the similarity between the original watermark and the extracted watermark or the similarity between the feature vector of the original image and the feature vector of the attacked image can measure the robustness of the algorithm. The similarity between two watermark images can be measured by the following formula:

[0170]

[0171] Among them, W(i, j) represents the original watermark, W′(i, j) represents the extracted watermark, is the mean of W(i, j), is the mean of W′(i, j).

[0172] In order to facilitate calculation, this paper uses the method of comparing the similarity between the feature vector of the original image and the feature vector of the attacked image. By calculating the normalized correlation coefficient, the similarity of images can be more accurately and objectively evaluated with data.

[0173] The robust watermark extraction method for medical images based on the InceptionV3 deep learning network provided in an embodiment of the present invention corresponds to the aforementioned robust watermark embedding method for medical images based on the InceptionV3 deep learning network, and can achieve the corresponding technical effect of the aforementioned robust watermark embedding method for medical images based on the InceptionV3 deep learning network, which will not be repeated here.

[0174] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A robust watermark embedding method for medical images based on InceptionV3 deep learning network, characterized in that: include: Select the medical image as the input image of the InceptionV3 network, use the InceptionV3 network to perform convolution operation on the medical image, take the output of the fully connected layer of the InceptionV3 network, and obtain the fully connected layer coefficient; Perform global discrete cosine transform (DCT transform) on the fully connected layer coefficients, then select low-frequency coefficients from the coefficient matrix obtained by DCT transform as the visual feature vector of the medical image, perform binarization on the visual feature vector, and obtain the feature vector of the target medical image; Perform chaotic scrambling encryption on the original target watermark to obtain a chaotic scrambled encrypted watermark; The feature vector of the target medical image is dimensionally upgraded to obtain a feature matrix of the target medical image; According to the obtained target medical image feature matrix and the chaotic scrambled encryption watermark, an XOR operation is performed bit by bit to embed the original target watermark information into the target medical image, and a binary logic key sequence is obtained at the same time.

2. According to claim 1, the medical image robust watermark embedding method based on InceptionV3 deep learning network is characterized in that: The original target watermark is encrypted by chaotic scrambling to obtain a chaotic scrambled encrypted watermark, including: Generate chaotic sequence through Logistic Map; Binarize the generated chaotic sequence to obtain a binary encryption matrix; The binary encryption matrix and the converted original target watermark are operated bit by bit to obtain a chaotically scrambled encryption watermark.

3. According to claim 1, the medical image robust watermark embedding method based on InceptionV3 deep learning network is characterized in that: Before selecting a medical image as the input image of the InceptionV3 network, it also includes: The image size of medical images was resized to 299×299.

4. The medical image robust watermark embedding method based on InceptionV3 deep learning network according to claim 1 is characterized in that: The target medical image feature matrix size is 32×32.

5. The medical image robust watermark embedding method based on the InceptionV3 deep learning network according to any one of claims 1-4, characterized in that: Also includes: The binary logic key sequence is stored in a third party, wherein the binary logic key sequence is bound to the corresponding medical image embedded with the original target watermark information.

6. A robust watermark extraction method for medical images based on InceptionV3 deep learning network, characterized in that: include: Input the medical image to be watermarked into the InceptionV3 network, use the InceptionV3 network to perform convolution operation on the medical image to be watermarked, take the output of the fully connected layer of the InceptionV3 network, and obtain the fully connected layer coefficient; The fully connected layer coefficients are subjected to global discrete cosine transform (DCT transform), and then the low-frequency coefficients are selected from the coefficient matrix obtained by the DCT transform as the visual feature vector of the medical image to be extracted watermark information, and the visual feature vector is binarized to obtain the feature vector of the medical image to be extracted watermark information; The feature vector of the medical image to be watermarked is dimensionally increased to obtain a 32×32 feature matrix of the medical image to be watermarked; Retrieving the medical image feature matrix of the watermark information to be extracted and the logic key obtained by a third party, performing an XOR operation bit by bit, and extracting the encrypted watermark; Decrypt the encrypted watermark to obtain the restored watermark information.

7. The method for robust watermark extraction of medical images based on InceptionV3 deep learning network according to claim 6 is characterized in that: Decrypt the encrypted watermark to obtain the restored watermark information, including: Generate chaotic sequence through Logistic Map; Binarize the generated chaotic sequence to obtain a binary chaotic matrix; The binary chaotic matrix and the encrypted watermark are XORed bit by bit to obtain the restored watermark information.

8. The medical image robust watermark extraction method based on InceptionV3 deep learning network according to claim 6 or 7, characterized in that: Also includes: The original target watermark is obtained, and the normalized correlation coefficient of the original target watermark and the restored watermark is calculated to determine the ownership of the target medical image and the embedded watermark information.

Images