CN115760536A - Full convolution blind watermark adding and analyzing system and method based on sub-pixel up-sampling - Google Patents
Full convolution blind watermark adding and analyzing system and method based on sub-pixel up-sampling Download PDFInfo
- Publication number
- CN115760536A CN115760536A CN202211403504.9A CN202211403504A CN115760536A CN 115760536 A CN115760536 A CN 115760536A CN 202211403504 A CN202211403504 A CN 202211403504A CN 115760536 A CN115760536 A CN 115760536A
- Authority
- CN
- China
- Prior art keywords
- module
- secret information
- image
- unit
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Image Processing (AREA)
Abstract
The invention discloses a full convolution blind watermark adding and analyzing system and method based on sub-pixel up-sampling, wherein the blind watermark adding system comprises a carrier image coding module, a DWT unit and a message processing module; the DWT unit is used for performing wavelet transformation on the carrier image and acquiring an LL frequency band generated by each wavelet transformation; the carrier image coding module is used for embedding the message characteristics into the carrier image to obtain a coded image I EN (ii) a The message processing module is used for extracting message characteristics from the original secret information by adopting an up-sampling mode and converting the low-frequency information LL frequency band of the carrier imageAdded to the message feature. The system has higher robustness and imperceptibility for adding blind watermarks to the images.
Description
Technical Field
The invention belongs to the technical field of information security, and particularly relates to a system and a method for adding a blind watermark into a carrier image, and a system and a method for analyzing original secret information from an image added with the blind watermark.
Background
Blind watermarking refers to a technique of embedding secret information into an image without destroying the effect exhibited by the image itself and also destroying the recognition result of the image. In the conventional robust blind watermarking method, a stable hidden space is usually found by means of some transformation to achieve robustness of watermark embedding.
Due to the wide application of deep learning, researchers have proposed some blind watermarking schemes based on deep learning. By means of the deep learning network to find stable embedding space and to perform anti-learning to various attacks, a robust watermark encoder and decoder can be established.
Document 1: jian Z, fang H, zhang W.Mbps of Enhancing robustness of dnn-based watermarking by mini-batch of real and quantized compression [ C ]// Proceedings of the 29th ACM International Conference on multimedia.2021. In this, jia et al propose a completely new model MBRS for JPEG compression, which selects a noise from real JPEG, analog JPEG and noise-free layers to add a noise layer and preprocess the message using a "message processor" in order to resist the cropping noise, MBRS can additionally add a "message diffusion block" to greatly improve the resistance to cropping, and the result of the model is far superior to all current blind schemes based on deep learning, which is the current SOTA scheme.
As shown in fig. 1, the MBRS network architecture disclosed in the above-mentioned document 1 is divided into 5 parts: encoder, message Processor, discriminator, noise Layer, decoder. The encoder receives the carrier image and embeds the features processed by the message processor into the carrier image and outputs a dense image. The message processor receives the secret information, reshapes the message, performs up-sampling and SE blocks by using deconvolution, diffuses and arranges the information, and outputs a characteristic diagram with the information. The discriminator receives the carrier image and the secret image and outputs a result of judging whether the input image is the secret image. The noise layer receives the dense image and outputs a noise image attacked by noise. The decoder receives the noise image and outputs the analyzed secret information.
The following ablation experiments were performed for the MBRS: 1. the full convolution network replaces the original SE blocks, and as a result, the performance of the network is not affected, and the reason for the result is that the feature extraction of the image is mainly embodied in the local feature extraction of the convolution layer, and the attention mechanism of the introduced SE blocks is more overall distribution of the feature map of each channel, so that not only can better local features be obtained, but also the calculation complexity is increased; 2. after replacing all networks in the message processor with simple copy in HiDDeN, and repeatedly copying M bits of secret information M to obtain an information map with the size of M multiplied by H multiplied by W, the network can hardly converge, and the error rate reaches about 50% which is remarkable. That is, the excellence of the MBRS is largely due to the presence of a message processor. Furthermore, the watermark image generated by the MBRS is easy to have regular chessboard effect, which makes the watermark easier to be found and removed.
Disclosure of Invention
The purpose of the invention is as follows: aiming at the problems in the prior art, the invention provides a full convolution blind watermark adding system based on sub-pixel up-sampling, and the system has higher robustness and imperceptibility for adding the blind watermark to an image.
The technical scheme is as follows: the invention discloses a full convolution blind watermark adding system based on sub-pixel up-sampling, which comprises a DWT unit, a carrier image coding module 100 and a message processing module 200;
the DWT unit is used for performing wavelet transformation on the carrier image and acquiring an LL frequency band generated by each wavelet transformation;
the input of the carrier image coding module 100 is a carrier image I CO The system comprises a first multi-convolution unit 101, a first connection unit 102, a second convolution unit 103, a second connection unit 104 and a third convolution unit 105 which are sequentially cascaded; the first multi-convolution unit 101 is used for extracting features of a carrier image and is composed of a plurality of cascaded convolution units, wherein each convolution unit is composed of a convolution layer, a batch processing layer and an activation layer in a cascaded mode; the first connection unit 102 is configured to connect features of the carrier image and message features to obtain joint features, and the second convolution unit 103 is configured to extract features of the joint features; the second connecting unit 104 is used to connect the carrier image and the associating feature,the encoded image I is obtained by the third convolution unit 105 EN ；
The input of the message processing module 200 is original secret information M, and includes a message reshaping layer 201, a preprocessing unit 202, a multi-level upsampling module 203 and a second multi-convolution unit 204; the message remodeling layer 201 is configured to convert the one-dimensional original secret information M into two-dimensional secret information M ', and the preprocessing unit 202 is configured to perform preliminary feature extraction on the two-dimensional secret information M'; the multistage upsampling module 203 is configured to perform upsampling on the preliminary features of the two-dimensional secret information, and includes N cascaded upsampling units, where, starting from the 2 nd upsampling unit, each upsampling unit is connected to a connection unit, and the connection unit behind the nth upsampling unit connects the output of the nth upsampling unit with the LL band image generated by the N-N +1 th wavelet transform of the carrier image, and serves as the input of the N +1 th upsampling unit; the structure of the up-sampling unit is a cascade convolution layer, a batch processing layer, an activation layer, a sub-pixel convolution, a batch processing layer and an activation layer; the second multi-convolution unit 204 is configured to extract features of the secret information feature image output by the multi-level up-sampling module to obtain a message feature M ″, where the message feature M ″ is formed by a plurality of cascaded convolution units, and each convolution unit is formed by a convolution layer, a batch processing layer, and an activation layer.
Further, the length L of the original secret information, the length H and the width W of the two-dimensional secret information M', the length H and the width W of the carrier image, and the number N of upsampling units in the multi-level upsampling module 203 have the following relationships:
L＝h×w＝(H/2 N )×(W/2 N )。
further, the training process of the parameters of the carrier image coding module 100 and the message processing module 200 includes:
s1, constructing a blind watermark coding and decoding system, wherein the blind watermark coding and decoding system comprises: the device comprises a carrier image coding module 100, a DWT unit, a message processing module 200, a noise layer 300, a coded image distinguishing module 400, a secret information distinguishing module 500 and a decoding module 600;
the noise layer is used for encoding the image I EN Adding noise to obtain a noise image I NO ；
The input of the coded image discrimination module 400 is a carrier image I CO And coded picture I EN For judging the coded picture I EN And a carrier image I CO Whether the images are the same image or not comprises a third multi-convolution unit and an average pooling layer; the third multi-convolution unit comprises a plurality of cascaded convolution units, and each convolution unit comprises a convolution layer, a batch processing layer and an activation layer;
the input of the secret information discrimination module 500 is the original secret information M and the decoded secret information M output by the decoding module out For judging the decoded secret information M out Whether the secret information is the original secret information M or not, and the secret information comprises a fourth multi-convolution unit and a linear layer;
a decoding module 600 for decoding from a noisy image I NO The original secret information is analyzed to obtain the decoding secret information M out The device comprises a decoding preprocessing unit 601, a multi-level down-sampling module 602, a message extraction module 603 and a message restoration layer 604 which are connected in sequence; the decoding preprocessing unit 601 is used for processing the noise image I NO Performing primary feature extraction, including a cascaded convolutional layer, a batch layer and an activation layer; the multi-stage down-sampling module 602 is configured to down-sample the preliminary features extracted by the decoding preprocessing unit 601, and further analyze secret information; the message extraction module 603 is configured to perform feature extraction on the secret information analyzed by the multi-stage down-sampling module 602, and obtain two-dimensional decoding secret information, which includes a concatenated convolutional layer, a batch processing layer, and an activation layer; the message recovery layer 604 is configured to reshape the two-dimensional decoding secret information into one-dimensional decoding secret information M out ；
S2, adopting countermeasure training to train a carrier image coding module, a message processing module, a coded image distinguishing module, a secret information distinguishing module and a decoding module in the blind watermark coding and decoding system, wherein the training is to minimize the following loss function:
wherein λ E 、λ D 、λ A 、λ dis The weight coefficients are preset weight coefficients which are positive numbers;
for the carrier image encoding module the loss function,
in order to decode the loss function of the module,
a first loss function for the encoded image decision module,
judging a second loss function of the module for the coded image;
a module loss function is determined for secret information, where MSE () is a mean square error function, A (I) EN ) Dis (M) for determining whether the image is a coded image by an image discriminator out ) Is M out Whether the secret information is the original secret information is judged through the secret information judging module.
Further, the preset weight coefficient lambda E 、λ D 、λ A 、λ dis Are positive numbers of equal value.
On the other hand, the invention also discloses a blind watermark analyzing system corresponding to the blind watermark adding system, which comprises a decoding module, wherein the decoding module is a decoding module in the blind watermark coding and decoding system which is trained according to the training process.
The invention also discloses a blind watermark adding method applying the blind watermark adding system, which comprises the following steps:
inputting the carrier image into the carrier image encoding module 100 and the DWT unit;
inputting the original secret information into the message processing module 200;
the output of the carrier image encoding module 100 is the image after the blind watermark is added.
The invention also discloses a blind watermark analyzing method applying the blind watermark analyzing system, which comprises the following steps:
and inputting the image added with the blind watermark into a decoding module, wherein the output of the decoding module is the decrypted secret information after analysis.
Has the advantages that: the full convolution blind watermark adding and analyzing system and method based on sub-pixel up-sampling disclosed by the invention improve the MBRS by adopting up-sampling based on sub-pixel convolution and adopt a coded image and secret information dual-discrimination mode to more robustly and covertly embed blind watermark information into a carrier image, so that the imperceptibility and robustness can be improved, the chessboard effect caused by deconvolution is reduced, and the training speed of a network is accelerated.
Drawings
Fig. 1 is a schematic diagram of a network structure of the MBRS;
fig. 2 is a schematic diagram of the blind watermarking system according to the present invention;
fig. 3 is a schematic diagram of a blind watermark encoding and decoding system constructed in the training process.
Detailed Description
The invention is further elucidated with reference to the drawings and the detailed description.
The invention discloses a full convolution blind watermarking adding system based on sub-pixel up-sampling, which comprises a DWT unit, a carrier image coding module 100 and a message processing module 200, as shown in figure 2; in fig. 2, the solid rectangular box represents ConvBNRelu, which is the cascade of convolutional layers, batch layers, and active layers; the trapezoid box represents ConvBNRelu + PixelShuffleBNRelu, namely the cascade of the convolution layer, the batch layer, the activation layer, the sub-pixel convolution, the batch layer and the activation layer;
the DWT unit is used for performing wavelet transformation on the carrier image and acquiring an LL frequency band generated by each wavelet transformation;
the input of the carrier image coding module 100 is a carrier image I CO The system comprises a first multi-convolution unit 101, a first connection unit 102, a second convolution unit 103, a second connection unit 104 and a third convolution unit 105 which are sequentially cascaded; the first multi-convolution unit 101 is used for extracting features of a carrier image and is composed of a plurality of cascaded convolution units, wherein each convolution unit is composed of a convolution layer, a batch processing layer and an activation layer in a cascaded mode; the first connection unit 102 is configured to connect features of the carrier image and message features to obtain joint features, and the second convolution unit 103 is configured to extract features of the joint features; the second connection unit 104 is used for connecting the carrier image and the joint feature, and obtaining a coded image I through a third convolution unit 105 EN ；
The input of the message processing module 200 is original secret information M, and includes a message reshaping layer 201, a preprocessing unit 202, a multi-level upsampling module 203 and a second multi-convolution unit 204; the message remodeling layer 201 is configured to convert the one-dimensional original secret information M into two-dimensional secret information M ', and the preprocessing unit 202 is configured to perform preliminary feature extraction on the two-dimensional secret information M'; the multistage upsampling module 203 is configured to perform upsampling on the preliminary features of the two-dimensional secret information, and includes N cascaded upsampling units, where, starting from the 2 nd upsampling unit, each upsampling unit is connected to a connection unit, and the connection unit behind the nth upsampling unit connects the output of the nth upsampling unit with the LL band image generated by the N-N +1 th wavelet transform of the carrier image, and serves as the input of the N +1 th upsampling unit; the structure of the up-sampling unit is a cascade convolution layer, a batch processing layer, an activation layer, a sub-pixel convolution, a batch processing layer and an activation layer; the second multi-convolution unit 204 is configured to extract features of the secret information feature image output by the multi-level up-sampling module to obtain a message feature M ″, where the message feature M ″ is formed by a plurality of cascaded convolution units, and each convolution unit is formed by a convolution layer, a batch processing layer, and an activation layer.
The length L of the original secret information, the length H and the width W of the two-dimensional secret information M', the length H and the width W of the carrier image, and the number N of the upsampling units in the multi-level upsampling module 203 have the following relations:
L＝h×w＝(H/2 N )×(W/2 N ) (1)
from the relation of equation (1), the multi-level upsampling module 203 converts the two-dimensional secret information M' into a feature map with the same size as the carrier image, and this enables the message to be spread to each corner of the feature as much as possible, thereby enhancing robustness.
Compared with document 1, the above-described blind watermarking system employs a sub-pixel convolution layer instead of deconvolution as a new upsampling manner, thereby mitigating the checkerboard artifacts of deconvolution due to repeated stacking of features and obtaining better image quality. The sub-pixel convolution layer does not contain an anti-convolution layer, and the influence caused by the checkerboard artifact is reduced. Furthermore, the sub-pixel convolution layer can effectively amplify the two-dimensional secret information M', and can reduce the computational complexity of the up-sampling operation. The sub-pixel convolution layer is used for changing a low-resolution image into a high-resolution image, obtaining a characteristic diagram of r ^2 channels (the size of the characteristic diagram is consistent with that of the input low-resolution image) through convolution, and then obtaining the high-resolution image through a periodic screening method, wherein r is an up-scaling factor (i.e. the expansion magnification of the image). The periodic screening has no convolution operation, the screening process is as if the image pixels are convolved by stride =1/2, and the operation in the whole pixels is performed, so that filter parameters which can be learned and any multiply-add operation are not involved, and the training speed is improved.
In the process of up-sampling, in order to better guide the message processing module to hide information at positions which are not easy to be perceived, such as the edge of a carrier image, a carrier image low-frequency information LL frequency band is added to facilitate fusion learning of secret information and the carrier image by a network, so that blind watermark information is embedded into a low-frequency part of the carrier image to obtain a better visual effect. Specifically, the carrier image is subjected to DWT, and the LL frequency band of DWT is a thumbnail of the image content, which is the frequency band in which the energy of the image data is concentrated. The image can be seen from the display of the coefficients due to the regularization processing of the wavelet coefficientsThe content is recorded. The HH band stores high-frequency detail information of the image, and the LH and HL bands also contain only a small amount of low-frequency texture edge information. Therefore, DWT performs multiple wavelet transforms on the carrier image and each time the resulting LL band is added to the characteristics of the upsampling process. The LL band image generated by DWT conversion is 1/4 of the size of the LL band image generated by last DWT conversion, and the up-sampling is to enlarge the input image by 4 times, so different up-sampling positions are put into the LL band images with different sizes. Since the image features after the first upsampling are too coarse to facilitate the extraction of local features, LL band information is added from the second upsampling. Specifically, a connecting unit behind the nth up-sampling unit connects the output of the nth up-sampling unit with the LL band image generated by the N-N +1 th wavelet transform of the carrier image, and the LL band image is used as the input of the (N + 1) th up-sampling unit, wherein N =1,2,3, \ 8230; \8230;, N. As shown in fig. 2, if there are 4 upsampling units, N =4, the LL band image YL obtained by 3 rd DWT conversion is added after the 2 nd upsampling unit 3 Adding LL band image YL obtained by 2 nd DWT conversion after 3 rd up-sampling unit 2 Adding LL band image YL obtained by 1 st DWT conversion after 4 th up-sampling unit 1 . This mechanism brings about an improvement in the quality of the encoded image, since information such as edges, contours, etc. of the image greatly affects the subjective quality of the human eye when viewing the image.
The training process of the blind watermarking system comprises the following steps:
s1, constructing a blind watermark coding and decoding system, as shown in FIG. 3, wherein the blind watermark coding and decoding system comprises: the device comprises a carrier image coding module 100, a DWT unit, a message processing module 200, a noise layer 300, a coded image distinguishing module 400, a secret information distinguishing module 500 and a decoding module 600;
the noise layer is used for encoding the image I EN Adding noise to obtain a noise image I NO ；
The input of the coded image discrimination module 400 is a carrier image I CO And coded picture I EN For judging the coded picture I EN And a carrier image I CO Whether it is the same image, packetA third multi-convolution unit and an average pooling layer; the third multi-convolution unit comprises a plurality of cascaded convolution units, and each convolution unit is formed by a convolution layer, a batch processing layer and an activation layer in a cascade mode;
the input of the secret information discrimination module 500 is the original secret information M and the decoded secret information M output by the decoding module out For judging the decoding secret information M out Whether the secret information is original secret information M or not comprises a fourth multi-convolution unit and a linear layer; the fourth multi-convolution unit comprises a plurality of cascaded convolution units, and each convolution unit comprises a convolution layer, a batch processing layer and an activation layer;
a decoding module 600 for decoding a noisy image I NO The original secret information is analyzed to obtain the decoding secret information M out The decoding device is the same as the decoder of the MBRS model in document 1, and includes a decoding preprocessing unit 601, a multi-level down-sampling module 602, a message extraction module 603, and a message restoration layer 604, which are connected in sequence; the decoding preprocessing unit 601 is used for processing the noise image I NO Performing primary feature extraction, including a cascaded convolutional layer, a batch processing layer and an activation layer; the multi-stage downsampling module 602 is configured to downsample the preliminary features extracted by the decoding preprocessing unit 601, and further parse secret information, in this embodiment, the multi-stage downsampling module 602 is composed of multiple cascaded SE Blocks; the message extraction module 603 is configured to perform feature extraction on the secret information analyzed by the multi-level down-sampling module 602, and obtain two-dimensional decoding secret information, which includes a concatenated convolutional layer, a batch layer, and an activation layer; the message recovery layer 604 is configured to reshape the two-dimensional decoding secret information into one-dimensional decoding secret information M out ；
S2, adopting countermeasure training to train a carrier image coding module, a message processing module, a coded image distinguishing module, a secret information distinguishing module and a decoding module in the blind watermark coding and decoding system, wherein the training is to minimize the following loss function:
wherein λ E 、λ D 、λ A 、λ dis The weight coefficients are preset weight coefficients which are positive numbers; in this embodiment, the weighting coefficients have equal values.
Compared with the document 1, the invention adopts a double-discrimination mode, namely, a secret information discrimination module is added, so that a decoder is helped to better analyze correct information, and the network performance is improved.
After training, the method for blind watermarking by using the trained blind watermarking adding system comprises the following steps:
inputting the carrier image into the carrier image encoding module 100 and the DWT unit;
inputting the original secret information into the message processing module 200;
the output of the carrier image encoding module 100 is the image after the blind watermark is added.
Meanwhile, the decoding module constructed and trained in the training process is a blind watermark analyzing system, and the method for analyzing the blind watermark by using the system comprises the following steps:
and inputting the image added with the blind watermark into a decoding module, wherein the output of the decoding module is the decrypted secret information after analysis.
This embodiment compares the blind watermarking system disclosed in the present invention with the prior art. To ensure the stability of the experimental results, the average of the results of 20 epochs after model convergence was selected as our final result. For the index, PSNR and BER are selected, wherein PSNR is peak signal-to-noise ratio and can represent the quality of the coded image, and PSNR ≧ 33.5 generally belongs to higher image quality. The BER is a bit error rate representing a percentage of different bits of information of the decoded information and the original information to the total number of bits of the information. The results of the experiment are shown in table 1.
TABLE 1
In table 1, hiddenn is adopted: zhu J, kaplan R, johnson J, et al.Hidden: high data with deep networks [ C]The blind watermarking network disclosed in/Proceedings of the European conference on computer vision (ECCV) 2018; TSDL is adopted literature: liu Y, guo M, zhang J, et al. A novel two-stage separable deep learning frame for reactive marking [ C]// Proceedings of the 27th ACM International Conference on multimedia.2019, 1509-1517Printing and adding a network; TRDH is adopted literature: zhang C, karjauv A, benz P, et al Towards Robust Deep mining Non-differential diagnosis for Practical Blind Watermarking [ C]// Proceedings of the 29th ACM International Conference on multimedia.2021. Blind watermarking networks disclosed in [ 5158-5166 ]; both MBRS and MBRS (diffusion layer added) are available from literature: jia Z, fang H, zhang W.Mbps, enhancing robustness of dnn-based watermark by mini-batch of real and related jpeg compression [ C]The network disclosed in// Proceedings of the 29th ACM International Conference on multimedia.2021. Ours is a blind watermarking system applying the present disclosure. Clipping (R = 0.3) indicates that the noise layer adds noise in a clipping manner, with a clipping ratio R =0.3; gaussian filter (σ) 2 =0.5, w = 7) indicates that the noise layer adds noise by means of gaussian filtering, the square of the variance is 0.5, and the window size is 7; the median filtering (w = 3) indicates that the noise layer adds noise in a median filtering manner, and the window size is 3; mixed noise means that clipping (R = 0.3) and gaussian filtering (σ) are used for the noise layer 2 =0.5, w = 7) and median filtering (w = 3).
As can be seen from table 1, the system disclosed by the present invention exhibits excellent performance in both robustness and imperceptibility. The MBRS requires an additional diffusion layer to be substantially equal to the system disclosed in the present invention, but this results in a reduced robustness of the MBRS to other noise. Moreover, under other noise attacks, the system disclosed by the invention can basically keep the same or even better performance, and the training time is shortened by about 30 percent relative to the MBRS. Therefore, the system disclosed by the invention is more excellent in universality, extraction accuracy and image quality.
Claims (7)
1. A full convolution blind watermarking system based on sub-pixel upsampling, comprising a DWT unit, a carrier image encoding module (100) and a message processing module (200);
the DWT unit is used for performing wavelet transformation on the carrier image and acquiring an LL frequency band generated by each wavelet transformation;
the input of the carrier image coding module (100) is a carrier image I CO The device comprises a first multi-convolution unit (101), a first connecting unit (102), a second convolution unit (103), a second connecting unit (104) and a third convolution unit (105) which are sequentially cascaded; the first multi-convolution unit (101) is used for extracting features of a carrier image and is composed of a plurality of cascaded convolution units, and each convolution unit is composed of a convolution layer, a batch processing layer and an activation layer in a cascade mode; the first connection unit (102) is used for connecting the characteristics of the carrier image and the message characteristics to obtain joint characteristics, and the second convolution unit (103) is used for extracting the characteristics of the joint characteristics; the second connection unit (104) is used for connecting the carrier image and the joint feature, and a coded image I is obtained through a third convolution unit (105) EN ；
The input of the message processing module (200) is original secret information M, and the message processing module comprises a message reshaping layer (201), a preprocessing unit (202), a multi-level upsampling module (203) and a second multi-convolution unit (204); the message remodeling layer (201) is used for converting one-dimensional original secret information M into two-dimensional secret information M ', and the preprocessing unit (202) is used for performing primary feature extraction on the two-dimensional secret information M'; the multistage upsampling module (203) is used for upsampling the preliminary features of the two-dimensional secret information, and comprises N cascaded upsampling units, wherein each upsampling unit is connected with a connecting unit from the 2 nd upsampling unit, and the connecting unit behind the nth upsampling unit connects the output of the nth upsampling unit with an LL band image generated by the N-N +1 th wavelet transform of a carrier image to serve as the input of the N +1 th upsampling unit; the structure of the up-sampling unit is a cascade convolution layer, a batch processing layer, an activation layer, a sub-pixel convolution, a batch processing layer and an activation layer; the second multi-convolution unit (204) is used for extracting the features of the secret information feature image output by the multi-level up-sampling module to obtain a message feature M' and is composed of a plurality of cascaded convolution units, and each convolution unit is composed of a convolution layer, a batch processing layer and an activation layer in a cascaded mode.
2. Blind watermarking system according to claim 1, wherein the length L of the original secret information, the length H and width W of the two-dimensional secret information M', the length H and width W of the carrier image, and the number N of upsampling units in the multi-level upsampling module (203) have the following relations:
L＝h×w＝H/2 N )×(W/2 N )。
3. blind watermarking system according to claim 1, characterized in that the training process of the carrier image encoding module (100) and message processing module (200) parameters comprises:
s1, constructing a blind watermark coding and decoding system, wherein the blind watermark coding and decoding system comprises: a carrier image encoding module (100), a DWT unit, a message processing module (200), a noise layer (300), an encoded image discrimination module (400), a secret information discrimination module (500), and a decoding module (600);
the noise layer is used for encoding the image I EN Adding noise to obtain a noise image I NO ；
The input of the coded image distinguishing module (400) is a carrier image I CO And coded picture I EN For judging the coded picture I EN And a carrier image I CO Whether the images are the same image or not comprises a third multi-convolution unit and an average pooling layer; the third multi-convolution unit comprises a plurality of cascaded convolution units, and each convolution unit comprises a convolution layer, a batch processing layer and an activation layer;
the input of the secret information discrimination module (500) is the original secret information M and the decoding secret information M output by the decoding module out For judging the decoding secret information M out Whether the secret information is original secret information M or not comprises a fourth multi-convolution unit and a linear layer;
a decoding module (600) for decoding a noisy image I NO The original secret information is analyzed to obtain the decoding secret information M out The device comprises a decoding preprocessing unit (601), a multi-stage down-sampling module (602), a message extraction module (603) and a message restoration layer (604) which are connected in sequence; the decoding preprocessing unit (601) is used for processing the noise image I NO Performing preliminary feature extraction, including cascading volumesBuild-up layer, batch layer and active layer; the multi-stage down-sampling module (602) is used for down-sampling the preliminary features extracted by the decoding preprocessing unit (601) and further analyzing the secret information; the message extraction module (603) is used for extracting the characteristics of the secret information analyzed by the multi-stage down-sampling module (602) and acquiring two-dimensional decoding secret information which comprises a concatenated convolutional layer, a batch processing layer and an activation layer; the message recovery layer (604) is configured to reshape the two-dimensional decoding secret information into a one-dimensional decoding secret information M out ；
S2, adopting countermeasure training to train a carrier image coding module, a message processing module, a coded image distinguishing module, a secret information distinguishing module and a decoding module in the blind watermark coding and decoding system, wherein the training is to minimize the following loss function:
wherein λ E 、λ D 、λ A 、λ dis The preset weight coefficients are positive numbers;
for the carrier image encoding module the loss function,
in order to decode the loss function of the module,
a first loss function for the encoded image discrimination module,
judging a second loss function of the module for the coded image;
discriminating a module loss function for the secret information, where MSE is a mean square error function, A (I) EN ) Is I EN The image discriminator judges whether the image is a coded image, dis (M) out ) Is M out Whether the secret information is the original secret information is judged through the secret information judging module.
4. Blind watermarking system according to claim 1, characterized in that the preset weighting factor λ E 、λ D 、λ A 、λ dis Are positive numbers of equal value.
5. A blind watermark parsing system comprising a decoding module in a blind watermark encoding and decoding system trained according to the training process of claim 3.
6. A blind watermarking method of the blind watermarking system according to claim 1, comprising:
inputting a carrier image into a carrier image encoding module (100) and a DWT unit;
inputting original secret information into a message processing module (200);
the output of the carrier image coding module (100) is the image added with the blind watermark.
7. The blind watermark parsing method of the blind watermark parsing system as claimed in claim 5, comprising:
and inputting the image added with the blind watermark into a decoding module, wherein the output of the decoding module is the decrypted secret information after analysis.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211403504.9A CN115760536A (en) | 2022-11-10 | 2022-11-10 | Full convolution blind watermark adding and analyzing system and method based on sub-pixel up-sampling |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211403504.9A CN115760536A (en) | 2022-11-10 | 2022-11-10 | Full convolution blind watermark adding and analyzing system and method based on sub-pixel up-sampling |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115760536A true CN115760536A (en) | 2023-03-07 |
Family
ID=85368096
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211403504.9A Pending CN115760536A (en) | 2022-11-10 | 2022-11-10 | Full convolution blind watermark adding and analyzing system and method based on sub-pixel up-sampling |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115760536A (en) |
-
2022
- 2022-11-10 CN CN202211403504.9A patent/CN115760536A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wu et al. | Reversible image watermarking on prediction errors by efficient histogram modification | |
| Huan et al. | Exploring stable coefficients on joint sub-bands for robust video watermarking in DT CWT domain | |
| CN111541894B (en) | Loop filtering method based on edge enhancement residual error network | |
| CN112714231A (en) | Robust steganography method based on DCT (discrete cosine transformation) symbol replacement | |
| Chammem et al. | DWT-based stereoscopic image watermarking | |
| Qi et al. | Generic reversible visible watermarking via regularized graph fourier transform coding | |
| CN116452401A (en) | Reversible robust watermark embedding and extraction model construction method for resisting image attack | |
| CN113612898B (en) | Robust covert communication device for resisting JPEG image downsampling | |
| CN116503230A (en) | Robust image watermarking algorithm based on double channels | |
| Lin et al. | A reversible data hiding scheme for block truncation compressions based on histogram modification | |
| CN115760536A (en) | Full convolution blind watermark adding and analyzing system and method based on sub-pixel up-sampling | |
| CN114363719B (en) | Content self-adaptive robustness video watermarking method based on Bessel-Fourier moment | |
| CN115880125A (en) | Soft fusion robust image watermarking method based on Transformer | |
| CN115545998A (en) | Blind watermark embedding and extracting method and device, electronic equipment and storage medium | |
| CN116029887A (en) | Image high-capacity robust watermarking method based on wavelet neural network | |
| Hussein | Luminance-based embedding approach for color image watermarking | |
| Ernawan et al. | A blind watermarking technique based on DCT psychovisual threshold for a robust copyright protection | |
| Jia et al. | Deep convolutional network based image quality enhancement for low bit rate image compression | |
| Kekre et al. | Robust watermarking using Walsh wavelets and SVD | |
| Jamali et al. | Robust image watermarking by multi resolution embedding in wavelet transform coefficients | |
| Yang et al. | High‐Quality and Robust Reversible Data Hiding by Coefficient Shifting Algorithm | |
| CN114449296B (en) | Loop filtering method and device based on convolutional neural network | |
| Khedr et al. | A novel blind and robust watermarking technique of multiple images | |
| Wu et al. | Enhancing Robustness and Imperceptibility of Blind Watermarking with Improved Message Processor | |
| Naheed et al. | Intelligent reversible digital watermarking technique using interpolation errors |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |