KR100468009B1 - Feature-based robust image watermarking scheme using bispectral invariant - Google Patents

Abstract

A new watermark embedding method and detection method are proposed. The proposed watermarking algorithm uses the Radon transform with smaller aliasing in the inverse transform process, unlike the conventional Fourier-Melin transform method. The watermark embedding method in the algorithm calculates a two-dimensional Fourier transform (FT) for the original image, maps it on polar coordinates, extracts a feature vector composed of double spectra of the orthogonal projections, and inserts it. After selecting a particular projection using the key for embedding, modifying the phase component for feature transformation using the intensity parameter and orthogonalizing the changed polar Fourier spectra Mapping on Cartesian coordinates, inverse FT again, and passing the watermark detector described below to use the extracted signal as a watermark, while the detection method is a two-dimensional image of the test image. The Fourier transform of the image is calculated and mapped to polar coordinates. A feature vector is extracted from the mapped Fourier spectrum and a key is obtained. By determining the insertion position is determined by using the similarity between the watermark and the extracted feature vectors.

The watermarking algorithm using the double spectrum and the Radon transform has an excellent invariant to geometric deformations such as rotation, scale, and translation. It has been shown experimentally that it shows robustness against attempts to modify or delete.

Description

FEATURE-BASED ROBUST IMAGE WATERMARKING SCHEME USING BISPECTRAL INVARIANT} FEATURE-BASED ROBUST IMAGE WATERMARKING SCHEME USING BISPECTRAL INVARIANT}

The present invention relates to digital watermarking. More specifically, a method of concealing specific code values inside multimedia data including audio, images, and video, and related to a means for providing effective protection of copyright and copyright of images and music, and an algorithm providing the above means. Is associated with.

Digital watermarking is a technique of embedding a certain signal in the original multimedia data. The method of creating a new image using a watermark is, firstly, when a watermarking signal is generated irrespective of the original image. It can be divided into the case of making with the original image. In general, for watermarking to work effectively, the following requirements must be met.

Difficult to notice: Difficult to notice: Watermarking is the most important factor in the case of watermarking, so that the original data should not be transformed and lossy compression is used such as JPEG or MPEG. If possible, the watermarking should be maintained when lossy compression of the original data is performed.

Robustness: Watermarking should not change the alignment state of the original image, such as improving the sharpness of the image or changing the color, and if necessary, the watermark viewer ( You should be able to read it using the viewer.

Tamper-resistance against damage or deletion of the watermark: If the fact that the watermark is included in the multimedia is known, there will be an attempt to damage or delete the included watermark. Watermarking should have a mechanism to prevent such attempts.

Watermark Scalability: Even in the case of watermarking, where current computer digital technology cannot solve it, it can be broken with the development of decryption technology in the future. In this case, the existing watermarking should be able to be extended rather than developing a new watermarking.

In order to satisfy the above requirements, conventionally proposed watermarking algorithms are disclosed in Korean Patent Application No. 10-2001-7012060 'Public Watermarking for Rotation, Scale, and Displaced Image' and Application No. 10-2001-7013591 'Watering Images. How to use the mark. "

The 7012060 application relates to a method for inserting or detecting a watermark such that the watermark is recovered from geometric distortions such as rotation, scale, and translation. Create a log-polar Fourier spectrum using a polar Fourier transform, project the Fourier spectrum into a lower-dimensional space, calculate and compare the pre-extracted signal to a target watermark signal, and confirm the presence of the target watermark signal. The process is done.

The 7013591 application includes inserting a message into an image by changing a data characteristic of the image when transmitting an image, and when receiving the image in which the message is inserted, determine the coordinates of the origin of the original image and read the inserted image to obtain the origin. It is performed using steps of periodically inserting a binary matrix P into the image and co-watermarking to allow registration of the received image with respect to.

The above inventions can be applied to any indexing application of invariant to geometric distortion, performance complexity on the system, image or image sequence using Fourier transform, cosine transform or wavelet transform as a transform based method of watermarking. Although solved to some extent, there was a problem in that the characteristics of the watermarking algorithm cannot be solved more clearly. In particular, it could not effectively solve the problems of invariant to geometric distortion, ability to cope with watermark, and detectability after separation from original image.

The present invention is to solve such problems.

Conventional image watermarking algorithms are sensitive to geometric deformations. An image may be distorted due to geometric deformation such as rotation, scale, and translation, and thus embedded watermark may not be detected due to a change in alignment. Watermarks embedded for the original purpose of watermarking are watermarking within the data in the case of general distortion of the signals constituting the image, i.e. increasing the sharpness of the image, or modifying the base frequency of the audio signal or compressing the image. It must be able to remain unbroken and the watermark viewer should be able to read the watermark inside the modified data as above.

In other words, the watermark embedded in the data should be invariant.

An object of the present invention is to provide an algorithm for having a robustness (robustness) that can be maintained when the watermark inside the image is deformed as described above.

That is, an object of the present invention is to provide an algorithm of watermarking invariant with respect to rotation, scale, and translation.

Recently, watermarking methods with invariances to geometric deformations have been reported (Lin, CY, Wu, M., Bloom, JA, Cox, IJ, Miller, ML, and Lui, YM: 'Rotation, scale, and translation resilient watermarking for images ', IEEE Trans.Image Process., 2001, 10 (5), pp.767-782; Ruanaidh, J., J., and Pun, T.,:' Rotation, scale and translation invariant spread spectrum digital image watermarking ', Signal Processing, 1998, 66, pp.303-317). Since the methods described above are based on the Fourier-Mellin Transform, difficulties arise in the process of embedding and inverse log-polar-mapping procedures. The present invention proposes a new watermarking algorithm having RST (Rotation, Scale, Translation) invariant characteristics for images. The proposed algorithm differs from the conventional method using the Fourier-Melin transformation and uses a Radon transform with small aliasing compared to the log-pole-mapping during the inverse procedure. Is the point. The proposed method embeds a watermark signal in the Fourier phase spectrum, which embeds an attempt to eliminate watermarking compared to the Fourier-Meline transformation based on methods in which the proposed system uses the Fourier amplitude spectrum. Make your coping skills excellent.

That is, another object of the present invention is to provide a watermarking algorithm having excellent ability to cope with an attempt to remove a watermark using a radon transform.

The method proposed in the present invention has excellent invariability using the bispectrum used for pattern recognition (Chandran, V., Carswell, B., Bosahash, B, and Elgar, S., : 'Pattern recognition using invariants defined from higher order spectra: 2-D image inputs', IEEE Trans.Image Process., 1997, 6, (5), pp703-712). According to the algorithm proposed in the present invention, the watermark is embedded by forming a bispectrum feature vector defined from the radon transform of the image and uses the modified feature vector as the watermark.

The watermark may be detected in the absence of an original image.

That is, an object of the present invention is to provide a watermarking algorithm that can be detected in the absence of an original image by using a modified feature vector as a watermark.

1 illustrates a watermarking plan according to the present invention.

(a) is a block diagram showing an insertion procedure;

(b) is a block diagram showing a detection procedure.

2 shows a watermarked Lena image

(a) is watermarked at r = 35 ° and 125 °

(b) shows the amplified difference between the watermarked and the original image.

A robust watermarking algorithm with RST invariant characteristics based on the Radon transform is described below.

Definition of feature vector

The double-spectrum, B (f _1, f ₂ ) of a sequence with one-dimensionally determined real values, is defined as follows;

B (f ₁ , f ₂ ) = X (f ₁ ) X (f ₂ ) X ^* (f ₁ + f ₁ ). … … (One)

X (f) denotes a discrete-time Fourier transform of the sequence at frequency f.

The parameter p is defined as the phase of the integrated bispectrum along a straight line such that f ₁ = f ₂ ;

(2)

An invariant feature vector P of length N is defined as

P = (p ₁ , p ₂ , ..., p _N ) (3) In the above, p _i denotes a parameter p of projection i generated by radon transformation of an image. P is invariant to dc-level shift, amplification, translation, scaling, and Gaussian noise (Chandran, V., Carswell, B., Boashash, B). And Elgar, S .: 'Pattern recognition using invariants defined from higher order spectra: 2-D image inputs', IEEE Trans.Image Process., 1997, 6, (5), pp.703-712). Image rotation results in a cyclical variation of P depending on the nature of the radon transform.

Watermark insertion

The embedding procedure is shown in Fig. 1 (a). Hereinafter, a watermark embedding process will be described with reference to the drawings. First, a two-dimensional Fourier transform F (w ₁ , w ₂ ) for the original image I (x, y) is calculated (S1). Mapping the values shown in the calculation process (S1) on the polar coordinates (S2), to make F _p (w, θ) for N θ values uniformly distributed by 1 degree between 0 ° and 180 ° (S2) ). This process has the same result as the Fourier transform after the Radon transform. It is used to select r (0 ° to 180 °), which is the angle of the projection at which the watermark is to be inserted using an arbitrary key k. From this orthographic projection, the feature vector P is obtained as shown in equation (3), and the watermark is inserted in such a manner that the feature vector is transformed. This feature shaping process is performed by changing the phase components of F _p (w, r) using the strength parameter δ according to the following equation (S3);

∠ F _p '(w, r) = ∠ F _p (w, r) + δ... … … (4)

After the phase is corrected as described above, a feature vector P ' having a watermark inserted therein is created (S3).

After the watermark is inserted, the inverse Fourier transform is calculated after the inverse polar coordinate transformation S4 to obtain the watermarked image I '(x, y) (S5). The feature vector P " is extracted from the watermarked image (S6-S8) and passed to a detector to detect the watermark. The vector P ' is transformed into a feature forming process during insertion. therefore, the detector will be used in place of these variations are considered a final vector, P "P '.

Watermark detection

The detection process is shown in Figure 1 (b). The embedded watermark can be played back without the original image. The two-dimensional FT of the test image I ^* is calculated (P1) and mapped to polar coordinates (P2). The feature vector P ^* is mapped from polar coordinates and then evaluated from the Fourier spectrum (P3).

Using the key k, the insertion location r is determined. The similarity s between the extracted feature vector P ^* and the embedded watermark P " is calculated in the following manner (P4);

… … … (4)

Where L determines the length of the interrelated adjacent feature parameters. If s is smaller than the threshold T, it is determined that a watermark is present.

Experimental results

Experiments were conducted using 256 × 256 Lena and 512 × 512 Fishingboat images. Two perpendicularly intersecting feature vectors are corrected using δ = 5 °. When the watermark signal is embedded at two angles, s is measured at two intervals with L = 30 and the average of the two measurements is taken. The threshold T = 2.5 is used as the detection threshold. 2A shows a watermarked Lena image and FIG. 2B shows an amplified difference between the original image and the watermarked image. The watermarked image exhibited a PSNR of 36 dBd and no embedded signal appeared.

The Stirmark 3.1 program was used to evaluate the algorithm (Fabien AP Petitcolas, Ross J. Anderson, Markus G. Kuhn: 'Attack on copyright marking systems', in David Aucsmith (Ed), Informing Hiding, Second International Workshop, IH'98, Protland, Oregon, USA, April 15-17, 1998, Precessings, LNCS 1525, Springer-Verlag, ISBN 3-540-65386-4, pp.219-239). For each case, a successful watermark detection rate was calculated. Rate 1 means 100% detection success and rate 0 means detection failure completely. The experimental results are shown in Table 1.

Table 1 Watermark Detection Rate for Multiple Attacks

Attacks Detection rate Scaling 1.00 Small angle rotation and output result 0.95 Random geometric deformation 0.93 JPEG compression 1.00 Gaussian Noise1 1.00

For scaling the watermark was successfully detected in the range between 0.5-2. The above facts show that the feature vector used in the present invention has sufficient invariance for proportions up to 50%. When rotating at small angles in the range of -2 ° to 2 °, the watermark was successfully detected without a time synchronization procedure. However, for angles larger than this range, it was necessary to make the feature vector time consistent before measuring s. The algorithm used here has shown a fairly good performance up to 10 quality factors for JPEG compression. The above results are shown because the characteristics of the image are used as the watermark rather than the general embedding method. The feature vectors used herein also showed good performance against Gaussian noise.

A new RST invariant watermarking method based on the invariant features of the image has been invented. The dual spectrum feature vector is used as a watermark and the watermark has high robustness when RST is applied. The embedding procedure used is designed to reduce the failure rate in the watermark detector. Experimental results show that the scheme used is good for a wide range of applications, including RST applications, JPEG compression, arbitrary geometrical applications, and Gaussian noise.

The methods and experimental results presented herein are exemplary only and the scope of the present invention is not limited by the methods and experiments but only by the claims below.

Claims (4)

Calculating a two-dimensional Fourier Transform (Fourier Transform) of the original image and mapping it on the polar coordinates; Modifying a phase component for feature shaping using a strength parameter to create a feature vector; Using a key to select a projection for embedding; Mapping the original image mapped to the polar coordinates onto Cartesian coordinates using an inverse Fourier transform to obtain a bispectrum vector; Using the dual spectrum as a watermark; To a digital image or sequence of images that is invariant to rotation, scale, and translation using a Radon transform or equivalent Fourier transform and polar mapping. How to insert a watermark. The method according to claim 1, And wherein said double spectrum is a sequence having a one-dimensional determined real value. The method according to claim 1, Wherein said one feature vector is created from a projection generated by radon transformation of an image. Calculating and mapping Fourier transforms on a two-dimensional image of a test image to polar coordinates; Evaluating a feature vector from the mapped Fourier spectrum; Determining an insertion position using a key; Determining similarity between the extracted feature vector and the embedded watermark; Watermark detection method comprising a.