TW200826686A - Method of authentication and restoration for images - Google Patents

Abstract

This patent presents a method of effective image authentication and image restoration by hiding watermarks into DCT coefficients. The basic concept is to embed the selected significant watermarking bits for authentication and restoration into the selected medium- and low-frequency DCT coefficients. Thus, it can detect the illegally tampered region and then the original information in that region is introduced for restoration. Experimental results show that the proposed authentication and restoration techniques can be applied to a DVR system, in which no original image information is involved, and it can effectively detect the illegally tampered region and restore the tampered region up to the human visual perceptual quality by only using a little embedded original information.

Description

200826686 IX. Description of the Invention: [Technical Field] The present invention provides a method for image authentication and restoration, and more particularly to an image authentication and restoration technique that uses discrete cosine transform block coefficients and hides watermark bits. . [Prior Art] Information Hiding technology belongs to

Steganography, steganography stands for "secret writing", and data hiding means hiding information from other information, and hiding information such as text, images, video, sound, etc. into documents such as documents and documents. Information such as document images, photo images, voice, music, videos, advertisements, etc. Data hiding applications include copyright protection, information verification (inf〇rmati〇n authentication), annotation implantation (ann〇tation) Implantation, covert communication, muitipie authorizations, etc. > material hiding mainly includes two major directions: digital watermarking and digital fingerprinting. Watermarking can hide information about intellectual property rights, such as original authors, publishing offices, company addresses, etc., on digital media products. Digital fingerprints give different compilations for each different product. Use this hidden code when being remade or not shared It is possible to find illegal communicators. 6 200826686 目月 ίCompared with the new digital watermark research direction, most of them apply the following two Lier foundations such as cryptography, spread communication (SpreacUpeCtrUffl Communication) and noise theory (N〇ise The〇ry) et al. In fact, the 'digital watermark can be regarded as a noise distribution, the digital watermark is not a noise pattern, and then added to the original signal. In recent years, 'there has been more and more The digital watermarking technology is applied to tamper detection and copyright protection. There are many easy-to-read multimedia video editing products on the market. These softwares use novel technology to edit single images or dynamic digital data. The easier it is to edit the tools, therefore, it is an important issue to protect the original digital image data and ensure that the images obtained are not falsified. A good watermarking technology must meet the following conditions: 1. Invisible · When we add a bunch of watermarks to the original image When you don't want other people to immediately find out that it is different from the original image, the watermark information will be easily extracted or corrupted. Therefore, usually a watermarking technology must achieve the invisibility = degree' purpose in addition to reducing the intention of the person. In addition to being tampered with, the most important thing is to preserve the content information that the original image should have. For the purpose of invisibility, the human vision system (Human Vision System, HVS) can be used for image technology. The smallest change in perception = (JND) hides the information to be hidden in a place where the human eye is less likely to detect. 2. Robustness: A good watermarker must have certain toughness. Signal processing or geometry 200826686 application of the second water = watermark information, even for some specific detection of the original floating h. After the artificial vandalism, can still detect production is not 7 for different aspects of the application, its robustness Strict: Sex is better than the watermark based on the copyright protection of the second, the sub-~ ❹ foreign watermark comes high, even in even Had many thoughts sabotage 'handle you make depend ",, the watermark information to detect a clear recognition of the copyright owner. · Embedded in the file of the watermark message, it is possible to extract the embedded message, so the embedded watermark must have a monthly b, the security of the watermark can be encoded, and then hidden into the watermark to be protected should be avoided The owner of the illegal use can detect the embedded 3. Security Valley. If someone has maliciously attacked the content and tampering it, reproduce it and then publish it with confidentiality, and reduce the malicious attack to use the password. Adding a watermark to the content of a digital file, the probability of a successful person cracking, only the watermarked watermark content. 4, Undeletable (Undeletable): The watermark information embedded in the floating-bit file must be closely integrated with the original file, that is, the watermark information cannot be easily removed because of the general message, the way of processing, in general The watermarking technology with image verification as the main purpose, after the watermark is removed, the image becomes an inactive and untrustworthy image content. In fact, each data hiding application needs to have different degrees of ability to resist data modification and data hiding. If L〇g〇 is used for watermarking purposes, it can be a visual trademark information and an invisible hidden trademark. For the invisible logo, robustness is a necessary ability; for the shadow 200826686, like the verification application rabbit main &, the main printing, the Lin Yin, mainly invisible watermark application, the father of the logo 'Required requirements for broken (also or semi-crushed) materials. Rong (4) has many researches dedicated to researching image visualization, tamper detection technology and publishing many international conference papers (1)-(10), broken! watermark, robust watermarking, and semi-broken type developed in recent years. Sweat =, can be used as a technology for image authentication and copyright protection. The study of sub-watermarks includes image tamper detection methods that are generally applied to the spatial or frequency domain, and sneak detection techniques that can resist image or video reduction methods have been proposed, in addition to vector quantization and passwords. Methods such as learning, amplitude phase shifting, etc. are also applied to tamper detection techniques. In this research field, the more representative ones are: Kwang-Fu Li [1J proposed image tamper detection technology based on wavelet transform technology, which is used to hide the low-frequency 4 injuries in the image into the 鬲 frequency part. Certification basis. Heart Wu and Bede Liu [3] proposed that the meaningful part of the image be used as a watermark and hidden in the image frequency domain to complete the image tampering debt test by look-up table method.

Chlng-Yung Lin and Shih_Fu Chang [5]_[7] proposed a method of authentication based on the inter-block correlation, in order to achieve the anti-compression characteristics of the certification mechanism in the Cong-based compression system green Using two randomly selected block relationships to extract feature information, they found that 'the inter-block relationship is invariant before and after compression, so it can resist JPEG compression. Although Lin uses this unique correlation to ensure that watermark information can be saved on the JpEG compression system, the way they calculate feature information is to randomly select any two areas 200826686, which makes the structural features in the image impossible. Be revealed. Therefore, Chung-Shien Lu and Hong-Yuan Mark Liao [5] improved the method proposed by Lin, discarding the correlation between any two blocks, and generating a meaningful digital signature for the structure of the image for multi-level wavelet transformed information. It is called Structural Digital Signature (SDs). The design of SDS is robust to the image processing of content retention. In other words, it is processed by image processing such as valence, compression and scaling, but if the image content is Tampering can be easily detected. In addition, L. M. Marvel [14] proposed a jpeg-based compression system.

Stego-JPEG is embedded in a watermark and can be applied to compressed images/videos based on the Sip-based system such as the MpEG system. s.Kutter [15] proposes to use the information of the feature position in the image to encrypt and generate the digital sign early. In the authentication, after the information hidden in the image is compared with the feature information of the image to be authenticated, those areas that cannot be correctly aligned are compared. Blocks can be called tamper-evident areas. This authentication method is determined to be resistant to medium-to-high quality JPEG compression because feature information is preserved in JPEG compression.

Phen-Lan Lin [16] proposed a hierarchical detection method in 2005, which enables the tampering area that could not be detected during the first detection process to be detected in the second and third detections. Three different sizes of detection blocks (2x2, 4x4, Uxl2) are used to achieve hierarchical detection. Hierarchical watermarking can improve the performance of tamper detection, but the probability of wrong debt measurement is also Will follow the promotion. Based on M. Kutter and F. Α P. Petitcolasb [18] and Sviatoslav Voloshynovskiy [19] and others, the image attack 200826686 is divided into five different types of hits. ·· 1. Removal type attack: Shixi ^ ^ ^ ^ The main purpose is to remove the information hidden in the image content: the hidden watermark information is regarded as noise and the gas is hidden. In the way of knife analysis or forecasting, all possible certifications are handled in a way that is confusing and clear-minded, so that the shirts can extract any relevant information that can be identified. Disk shearing $attack. Basic image processing as a watermark attack = type 'the original intention is not to remove the embedded watermark message, but the monthly image of the W image processing technology, the watermark for the funeral After the image content is processed, the original pixel value relationship is destroyed, and then the verification fails. The typical methods include: displacement, clipping, filtering, r 疋 、, scaling, blocking or column deletion. Etc. β 3, cryptography Strike: This type of attack is to use the exhaustive method or the = force method to force the guessing of the original embedded watermark location and content. When the confidential information is cracked, the built-in verification information can be easily removed. It is possible to tamper with the video content and still pass the original verification procedure: this type of attack is quite time consuming and consumes resources 'cause & if it is not very familiar with the inspection procedure', this type of attack technology is usually not used. Protocol attack: embedding different authentication information by different watermarking methods, so that the original embedded information is affected, and the original and verification process is invalid. /, 5. Others: such as compression, quantization, format conversion, etc. The furnace can cause the loss of certification information. % 11 200826686 Logo-based watermarking application must be able to resist each of the above attacks, including compression, deformation, conversion, scaling, ... Temple's though The purpose of L〇go watermarking is to be able to identify the original or copyright owner, so you only need to be able to use the human eye to identify the detected flaws. The image-authenticated watermark application can only allow attack patterns that do not change the image theme 'such as compression, scaling, " ..., etc., for the watermarking of image verification, the relative restrictions are relatively large. M·Kutter and F·AP Petitcolasb [18] proposed an image verification method based on wavelet transform in 1999. The main purpose is to detect the spatial and frequency domain changes in the image and to identify the areas that may be tampered with to verify Correctness of the image. (1) Parameter definition: f (m, η): original image

Lez+ : the maximum order of wavelet transform

Fk,1 ( m,η ): wavelet image coefficient, k=h,v,d,a,1 = 1,2, L fa,L(m,n): wavelet coefficient w(i) of the highest order lowest frequency · Validation key, i = l, ..., Nw ckey(i) : Coefficient selection key , i=l , qkey(i) · Quantization key 5 i=l , ...... , Nw ( 5 : Quantization parameter (2) Watermark hiding method: First, the original image f (m, n) is L-order wavelet transform, and the frequency 12 200826686 conversion can obtain 3L different wavelet coefficient image information, and the highest order lowest frequency The wavelet coefficients fa, L(ffl, n) 'The wavelet coefficients thus decomposed can be expressed as follows: {Λ>, 4:==βί^Ι/(Μ] (2-1) where k=h, V, d, a (horizontal, vertical, diagonal, highest order lowest frequency region), and 1 = 1, · · ·, L. Next, for the horizontal, vertical, and diagonal wavelet coefficients at each order, by Ckey Select the position of the coefficient to be embedded in the watermark, as follows:

Step 1: If, the value of Lishi 1 is between 1 and Nw, we make the following adjustment: if Qs(fkJ (m5η))^ ^)φ qkey^ ~(m, w) = I 八, (w,w) - Λ, / (所,《) > 〇[/:j{m^«)+ δ2ι, if fkJ (m5 n)<0 (2-2) else zki(m^n )- fkl{m,n) end where function &(7) is defined as: 2^(/)= 0,if f J2l _ Uf f is even is odd (2-3) For the highest order L lowest frequency k = Each of a is not quantized in terms of the wavelet 13 200826686 coefficient of position (m, n). The L-order discrete wavelet inverse transform is performed, and the wavelet coefficient set is added back to 4 to convert back to the original image Z(m, η), which is expressed as follows: » Z(m, n) = ID WTHAAR [{z^; (m? n )}J ( 2 - 4 ) where female = nine f, and /=1,2, . . . i:. (3) Method of watermark extraction and tamper evaluation: Suppose we verify the image z(m,n) that has been hidden in the watermark, we need a string of binary embedded embedding encrypted by key κ The ruler is W(1), 1 = 1, 2, · · · Nw, wherein the original watermark data can also be restored via the key κ. Firstly, the original image Z(m, η) is transformed into Haar discrete wavelet transform of L order, which produces different wavelet image k' 1 (m' η), where k=h, V, d (horizontal, vertical, right Angle), 1 = 1,... L is the difference* with the wavelet coefficients decomposed by the wavelet analysis layer. 'Out...The discrete wavelet coefficient from the decomposition ^^昌庸•叫射, take its watermark mother--take the watermark bit as 砟)=么(4>, "))^知(/), 屮· 愚^, wavelet coefficient; 彳见/?) is selected by ^ & used in the key κ to extract the R returned to the original floating variable =: the watermark comparison is the same, and Use the degree of tampering that is derived from the Xia Box: (4):

If TAF(wfw) > T e/f changes the influence of the image analysis problem (2-5) can be "any video content end 14 200826686 where 33⁄43⁄43⁄4) = 士坌_·) 十邱) Ϊ=1 Τ Acceptable tolerance for image tampering.

The watermarking technology proposed by Kutter can verify whether the image content is burned or not, and the tampering bruise is detected by the recorded watermark embedding position, and the tolerance for jpEG compression is adjusted by using the intensity of quantization 5 In this way, the correctness of the non-malicious attack of the image content subjected to distortion, noise, and scaling is reduced, and the tampering part does not have the ability to recover. Therefore, when the image is damaged, it is necessary. Re-transfer the image and verify it to get the information you need. In addition, Yulin Wang and Alan Pearmain [13] proposed two different watermarking methods to detect and mark tampering areas in 2-4 years. These two ocean watermark methods respectively estimate the change of brightness information in the spatial domain, and The change in the value of the π coefficient in the OCT frequency domain hides the watermark information. Based on human visual research, it is pointed out that the human visual system is less sensitive to changes in single pixel values, and it is not easy to be modified even for a few specific pixel values. (4) Even so, the adaptive part of pixel values for image processing: weak Therefore, Wang uses the relative relationship between the coefficient values of the blocks as the embedded watermark information. "The absolute method achieves the best balance point for the invisibility, information volume, and robustness of the embedded information, and can be applied. M DCT-based video (4) technology, MPEG-2, Η.262, etc. The following is an example of a watermark embedding algorithm. (1) Parameter definition:

Lreal ··Center point pixel value brightness information 15 200826686

Lraean : The brightness of the adjacent pixel value in the selected range: the average value of the body △ 1 △ 2 · The threshold used for the intrusion and verification (2) The watermark hiding method: Please refer to Figure 1, in the In the proposed watermarking method, the block selection can be 3x3(a)7 7r , ^ v J Dxb lb) , 7χ7 ( c), or even 9χ9 (d) pixel value block, and Lreal means “ 〇,, the value of the pixel in the indicated block, and Lmean refers to the value of the surrounding pixel value. When the authentication bit to be embedded is “丨,, we modify the shell information of the intermediate pixel value to make it Meet the following formula: (2-6)

Lreal ^ Lmean + Δΐ Conversely, if the authentication bit to be embedded is “〇,, the information of the intermediate pixel value is modified to satisfy the following formula: (2-7)

Lreal <C Lmean 八 2 2 Based on the results obtained from the experiment, △ ι and △ 2 are set to 5 to 10% of L... as the optimum range. (3) Method of watermarking extraction and tamper evaluation: When extracting the information hidden, the method used and the method of embedding are similar. First, calculate the luminance information average 厶_ of the block and the brightness of the center point. If LReal 2 Lmean, the extracted authentication bit is 1 ' otherwise, if Lreal < Lmean, the extracted bit is '' 〇' 'Compare the extracted authentication bit with the original authentication bit, 16 200826686 This will detect the falsified area. The following is another embodiment of the watermark embedding algorithm. : (1) Parameter definition: AC.., • Low-frequency zone Μ coefficient value of the central block ^q\ ;: Estimated dc coefficient value of the low-frequency zone of the central block △: Threshold value used for embedding and verification ( 2) Watermark hiding method: In the method proposed by Choi and Aizawa [12] in 1999, the DC coefficient values in the 9 8x8 DCT blocks are used to calculate the information to be embedded 'to change the current block ( The DC coefficient value of the central block is hidden as the watermark 'BeiXun. However, Choi's method only uses 9 〇 (^ block up, down, left, right, four blocks as Z) Estimation of the C value (see Figure 2), this way for human vision The high sensitivity, that is, the change in the condition will reduce the pSNR ratio of the image quality. In addition, the estimation method for the threshold (△) is used to control the robustness and invisibility of the watermark. To achieve the best balance point, it will be a difficult choice. To improve the shortcomings of Choi, Wang and Pearmain [12] were in 1990: a method to enhance the accuracy of the prediction of the jc coefficient in the low frequency region (see Equation 2 below). -8), the block used by the estimation is amplified from the original upper, lower, left, right, and four blocks to the adjacent 8 blocks, and the original information is hidden into the Han: The value is changed to 5 values in the low frequency region, which can reduce the image quality due to the increased amount of embedded information. 17 200826686 ^C(05l) = U3884*(Z)C4-Z)C6)/8; = 1.13884 * (DCi - DCs) / 8; AC(0,2) = 0.27881 * {DCa + DCe - 2 * DCs) / 8; AC(250) = 0.27881 * {DCi + DC% - 2 * DCs) / 8; ^C(1,1) = 0.16213 * (DCi + DC9 - DCs - DCi) / 8; ( 2 - 8 ) Next, apply the same judgment condition as above to hide the authentication bit to the central block 5 One , the coefficient value, where represents the five 'values estimated in equation (3), respectively ^(o,1), "c(1,0),"c(0,2), Jc(2, 0), Na, 1) 〇

Set ACi > AC\ + Δ to embed bit Ί

Set ACi < AC\ - Δ to embed bit Ό1. ( 2 - 9 ) (3) Method of watermark extraction and tamper evaluation: At the time of authentication, the coefficient extracted is compared with the estimated by neighboring blocks. If the extracted feature bit is " Γ , otherwise if A: /, then the extracted feature bit is "〇". Then the extracted feature bits are compared with the original authentication bits, and then Knowing whether the block has been maliciously tampered with. However, this embedding method has several shortcomings. First, when the JCV risk is equal to XC, it cannot be determined that the authentication bit is 0 or 1, which causes blind spots on authentication. Secondly, in the 8 8x8 DCT blocks, only the middle block is embedded in the watermark, which will cause block effect in the authentication process, and it is impossible to clearly identify the area that is exactly modified by Xia. Ching-Yung Lin Shih-Fu Chang [5]-[7] proposed anti-JPEG compression and MPEG compressed image authentication methods from 1999 to 2002. The authors propose two image correlation invariance theories based on 8x8 DCT blocks, as follows: :

1. Define Fp and Fq are two 8x8 DCT 18 200826686 blocks that do not overlap in image X. Q is the quantization table used for JPEG distortion compression, V ", ve[0,...7], VMe[1 ,···川, 夕 is the number of all 8x8 DCT blocks in the image, △', F {μ, v) ^ integer roundi Fp Q(u,v) Q(u,v) , the following properties are always true ·· ^ ^p,q (U^v) > 〇, then AFp q (u, v) > 05 (2-10) elseif AFpq(u,v)<Q,thenA?pq(u,v )sQ, else AFpq (u, v) = 〇5 then AF (w5 v) = 0 2. Extend the definition of 1 above, Are and a fixed threshold, for ^ 1; Kv ^integerround any w, v , 〇, v) 2elseif AFpq, u, v) <k, k , the following properties are true:

elseMp,q(u,v) = k, Ο,): person ν·δ(“,ν), |,vl).0(W,V), k 〃 2(w,v) e elsewhere KyQ(u ,v), ,fc,v+l).0(w,v), ,k ,sZ Ωψ,ν) elsewhere ^ Kv -Q{u,v\ i~ \ k Q{u,v) eZ (2 -11) Based on the above two invariance theorems, Lin and Chang propose the following image verification method to solve the uncertainty of the watermark information on JpEG compression. (1) Parameter definition: f (m, η): original image 19 200826686 maximum order of wavelet transform

V fk,l(m,n) ··wavelet image coefficient,

......, L (2) watermark hiding method ·· First, take the original image out a = lt spleen 1 r different eigenvector sets to make the original image especially 丨 | into 8 s ΠΓΤ .s ^ Block, and CT differentiation of each block, all blocks ~ two groups, of which - ^ t. ~, consider each block in P v = lt 〇 6 to collect A n coefficient values 'for N different special τ, the critical value of moxibustion and 乂, #中走为在在在在

The critical value of Theorem 2. When you = 1 , specify (=0 Μ , retain the sign of Μ, and then the process: ' -I:"" 疋々 is the dynamic binary decision range to ensure the accuracy of the verification ' When we extract more feature vector sets, we can do more accurate verification. Set:: 疋广 is a mapping function, so we can find two / heart 2, ... 2} and V", ..., ~}, make and satisfy 4 fly% 1, such as ^{l,3,w_l}, eight={24, small, one of the functions can be selected as a strong watermark authentication security. It is enough to add M (L after calculating the H of each block in P to calculate its H if... then the extracted feature code is ''〇,, otherwise Z„ is “i, right 隼, The feature code set is called the feature data of the original image, which is the basis for the subsequent image verification. (3) Image verification method: 20 200826686 Based on the above image feature data acquisition method, for each block and The DCT coefficient value P, using the judgment formula of The〇r(4)2, extracts the characteristic value Z„ from the image, and then stores the (four) value with the original z"Do the comparison" If the two do not match, it is called the area to be tampered with. In the method proposed by Lin and Changm, the correlation between the blocks is used as the extraction and comparison of the feature values, by adjusting The threshold value and the number of coefficients to be judged in the 8x8 DCT block can enhance the verification strength or accuracy for the image content to be applied, and can be applied to JPEG or MPEG, etc., mainly based on DCT quantization. Image compression technology. In other image restoration methods, J. 讦丨 丨 与 and ^ G〇ijan [8] proposed the method mainly to obtain image features in the DCT frequency domain, and when the image is suspected of being tampered with This feature can be used to verify or tamper with the image. The process is shown in Figure 3. The image restoration method proposed by J. Fridrich and M. Go 1 jan is to extract the feature values in the DCT frequency domain. After the original image is divided into non-repeating 8χ8 blocks, each block is DCT-converted, and the energy in the frequency domain is obviously concentrated in the upper left corner in the Zig-zag mode (see Figure 4). Every DCT turn The latter block uses the jpEG standard quantization table to perform the deuteration action. Most of the information retained after quantization is the coefficient of the low frequency region. This is because the design of the JPEG quantization table is less sensitive to the change of the high frequency £. And the asymmetric quantization method designed.] Firdrich and M. Gol jan use this feature to design their image restoration methods. The image restoration methods proposed by J. Fridrich and M. Goljan have two 21 200826686 feature coding methods. They are 64 bits and 128 bits respectively. The feature vector generated by the 128-bit encoding method is larger, but the relative image quality that can be restored is also better. Referring to FIG. 5, the meaning of the number of the coding matrix is the number of bits. For example, the meaning of 7 in the first column of the first row is that the quantized DCT coefficient of the corresponding position is the maximum allowable coding range of 7. Bits, when the coefficient value is greater than 127, is regarded as 127. r Use this principle to record the block features of each block in an image. After all the block features are obtained, the results are hidden in the space domain [if 128 bits are encoded, then Hide to a minimum of 2 bits). Since the method proposed by J. Fridrich and M. Gol jan is to record the coefficient value of the low frequency in the frequency domain', the texture of the detail part is subject to the low-pass filter (Low-Pass Filter). The treatment has been general, and the display is more vague. See Figure 6 for a comparison of the experiments performed by the method proposed by J. Fridrich and M. Gol jan: (a) is the original image; (b) is the method proposed by J. Fridrich and M. Gol jan. Feature hiding; (c) License plate that has been tampered with; (d) Reduced license plate. In 2002, Hsien-Chu Wu and Chin-Chen Chang [9] proposed a spatial domain-based image restoration method. The special feature of this method is that it only restores the outline of the image, that is, the edge area, and cannot restore the tamper image to the grayscale value of the original image. Based on this feature, this method has a very low image feature vector length. Its characteristics produce a private graph '凊 see Figure 7, at Hs i en-Chu 22 200826686

The edge detection method used by Wu and Chi η-Chen Chhongng uses four different edge detection masks to obtain the edges of the image with the following formula: 9kG\H, D\D\v) 2] X z{i, j) x Mask(U j) (2-12) where the four masks of //, /), D+, and F represent the horizontal edge mask, the vertical edge mask, and the +45° slope. Diagonal matte with -45. Oblique diagonal mask, see Figure 8. After calculating the λ values of the four masks separately, the largest value is compared with the threshold value r, and if it is greater than the threshold value, it is set to 1 (edge), otherwise it is 〇 (background). Therefore, 1 bit can be used to represent 4><4 blocks are edges or background, and an 8x8 block only needs 4 bits to represent edge features, wherein two different threshold values are used as Edge Judgment Conditions (See Figure 8) 'Experimental results of edge extraction and reduction of tampered images using the method proposed by Hsien-Chu Wu and Chin-Chen Chang (see Figure 9). In addition, Phen-Lan Lin et al. [16] proposed an image restoration method for obtaining image features in the spatial domain in 2005, and hashed the obtained feature vectors to obtain verification information with a length of 16 bits, which can be used for images. Verify and restore the image that was tampered with. The flow chart is shown in Figure 1. In the image feature extraction method proposed by Phen-Lan Un, the original image is first cut into non-overlapping 8X8 blocks (see Figure 1 1). Cut each block into "2 sub-blocks for processing, find the average of each block and record the highest six bits to become the restoration information of the sub-block. So an 8x8 block will Generate 96-bit restoration information. After taking the feature value in the image at 23 200826686, use the lowest two-digit element of each pixel to hide the _ _ a, so that the block can just store the (10) bits of the image feature. (block location + restore information + verification information), and when hiding, in order to avoid the feature information of the block hidden in the block to which it belongs, use the key to scramble the block and then hide the information. Please refer to Figure 1 2, (a) and (c) for the image to cut off the left and above image information, (b) and (4) to use the hidden information, (a) and (c) carry out tampering and recovery actions, observe the original image of the modified area and after tampering The restored image content can be found that the image after the restoration of Xiachang has obvious block effect. This is because the record of the restored information records the average value of the sub-blocks, so after the tamper-reduced area is restored, there will be obvious blocks. [Invention] The present invention relates to a method for image authentication and restoration, which is designed to effectively detect an illegally tampered area and perform information restoration to achieve image quality that can be directly recognized by the human eye. In order to achieve the foregoing object, the present invention provides a method for image authentication and restoration, which utilizes discrete cosine transform (Discrete. "This Transform" block coefficient and image authentication and restoration technology by hiding watermark bits, mainly It is an action of performing image verification on a single image obtained in the Digital nde〇Recorder System. 'The characteristic information obtained by the relationship between the blocks is used to solve the certification of the 24 200826686. Information on the impact of calculation errors in the DCT conversion and quantization process, and after detecting the falsified area, using the digital signature The image feature information, the area was tampered with to make undone action. With the above method of image authentication and restoration, the following effects can be achieved: ^ 1. Can accept image compression within a reasonable range: 疋DCT conversion technology used by JPEG compression's corresponding dct between blocks After the coefficient is quantized, quantized, and converted, there is still a large majority, and the small one is constant. However, when the feature value is taken, the correlation between the blocks is utilized, so the method proposed by the present invention can resist A reasonable range of image compression. . 2. It is possible to find out the part of the image that has been tampered with and mark the possible tampering area: because this method first cuts the image into 8 x 8 blocks, and after DCT conversion, the DC and AC eigenvalues are calculated. The monument is the information to be hidden, and for verification, it is also for each 8 x 8 DCT block.

By doing the calculation and comparison of the eigenvalues, it can be pointed out that the 8×8 DCT block has been tampered with, so that the portion of the image that has been tampered with can be found and marked. 3. Take the characteristics of the surrounding blocks when taking the features: because the inter-block relationship will remain consistent before and after the conversion (according to Chlng~Yung Lin and Shih-Fu Chang[5] [6] [7] The corresponding coefficient values between the proposed rationalized blocks will retain the same property regardless of the number of discrete cosine transforms and the inverse transforms. Therefore, the present invention uses this characteristic to take the feature values to achieve the semi-breaking property. The effect of watermarking. . 4. The image after hiding the information still retains high-quality image content: because for each 8×8 DCT block, the information we want to hide is 1 DC eigenvalue and 4 AC eigenvalues. After that, I used the method of table lookup to hide information. Therefore, in our method, the hidden information bits are not in the same place and can retain the characteristics between the blocks. Therefore, the images hidden in the information still retain high quality. Image content. 5. Use different block coefficient scanning methods to obtain the information needed for image restoration. 'Storage space required for effective simplification: In addition to JpEG compression using Zig-zag scanning method, Η·261/Η·263 in Intra_Frame In the coding mode, two different coefficient scanning modes of rotating vertical scanning and rotating horizontal scanning are added, so the flatness index can be calculated and classified according to different blocks, according to the energy distribution of different blocks. Different coefficient scanning methods are used to reduce the energy loss caused by only the Zig-zag scanning and only the coefficient values of the low frequency fixed position, and DPCM is used to store the important coefficients of the block to reduce the amount of stored data required. Θ, the tampering block can be effectively restored: the information to be restored is stored as a digital signature in an additional digital storage space. When the present invention utilizes the method of verification 26 200826686 to find the falsified block, the present invention can The original recovery > sfl is retrieved from the digital signature to be restored. The recovered information includes the block DC coefficient and m AC coefficients, plus a block horizontal/vertical decision bit, so this The invention can use this information to effectively recover the tampered blocks. [Embodiment] The image compression technology is divided into two types of compression methods: distortion (L〇ssy) and distortion-free (L〇SS-Less), wherein the distortion-free compression includes Huffman coding, shannon-Fano coding, arithmetic coding, and The encoding method of the Lz series. The modes that are generally used for distortion-free image compression are "statistical mode" and "dictionary basic mode". Among them, the statistical mode is based on the probability of occurrence of each symbol, giving the probability of occurrence of each symbol. The established coding table has the following important properties: 1. Different codes use different number of bits. 2. Codes with low probability symbols use more bits, and codes with higher probability symbols use fewer bits. But it can be decoded uniquely. In fact, it is 3. The length of the code is not the same. · Typical statistical mode coding methods are Huffman coding and arithmetic coding. These two coding methods are used in many international standards such as JPEG, JPEG2000, MPEG, H.26x, etc. The dictionary-based mode of the distortion-free image reduction method takes a completely different 27 200826686 method to compress the image 'for the different lengths of the symbols are represented by the same token, and the attack mark represents a phrase in the dictionary. Position, as long as the number of bits required by the record is smaller than the number of bits required for the phrase it replaces, we have done image compression. In order to pay a lower data rate than the distortion-free image compression method, we allow the reconstruction signal to have some distortion, which may be visually obvious or not easily detectable. Generally, the image transmitted on the network allows a large distortion to achieve a lower data rate, and the distortion compression method commonly has vector quantization coding, conversion coding, wavelet coding, etc., among which is a very basic distortion image compression. There are many important image compression techniques (such as jpEG) applied to the basic concept of VQ. Therefore, the improved VQ compression result will be helpful to other related image compression technologies, so VQ is the most widely used in academia. To study the important format of image compression. The basic practice of traditional VQ first divides the compressed image into many small squares of the same size, for example, a 512-pixel image. We usually divide it into small squares of 4x4 points, which are completed in advance according to the query. Code book, the code vector closest to the original vector (c〇de w〇rd ) will be selected, and then use the index values of these closest code vectors to form an index table, thus Complete image compression. This index table is the result of VQ (four), because the volume of the index table is usually much smaller than the size of the original image, so VQ can have a good compression effect. - In general, the quality of the VQ (four) image is determined by the code. The number of thin inner code vectors and the pros and cons of representativeness. 28 200826686 Conversion coding is to convert the original signal into another representation. This representation can be restored to the original signal via inverse conversion, and its monthly b 塁 original signal is concentrated, so it is easier to do image compression. Imagine image conversion methods such as Karhunen-Loeve conversion (KLT), digital Fourier transform (DFT), and digital cosine transform (DCT), etc., but for images that generally exhibit cumbersome sampling between heights, DCT The performance is almost the same as that of KLT. Because it does not produce high-frequency terms like D{?T, DCT can achieve higher compression performance. Also popular in the cross-frequency coding method is the use of wavelet (Wavelet) as a frequency filter; wavelet refers to a functional group formed by a function of amplification and translation, in speech processing, image processing, computer vision research The field has quite successful applications. The J wave knife solution can be regarded as a special case of the frequency division coding method, and the most representative of the ten generations is the use of red ΤΠΤ^^ΟΛΛ A .

The color model used is different, vrh~, VTA,

Indicates the difference between the blue component and a reference to a reference value. The difference in value, Cr means that the red component and the other conversion formula are as follows: 29 200826686 "16 ' '65.481 128.553 24.966" A R - Cb = 128 + a 37.797 — 74.203 112.000 G Cr 128_ 112.000 -93.786 -18.214 B For example, ^^6,17,...,235}, C0,Cre{l6,17"..,24〇}, and in the (3-1) J Jl IL Vj compression specification, allow co^ Ed·, 255}. YCbCr is not an absolute color space, it is a way to encode RGB color information, in fact, it still needs to be converted to RGB color space. Discrete Cosine Transform (DCT) It converts digital image information in the spatial domain into a frequency domain, which is a Fourier-related transform, similar to Discrete F〇urier Transform (DFT), but only uses real numbers. The most commonly used variant of discrete cosine transform Conversion is a second type of DCT, often referred to simply as DCT, and its inverse conversion is often referred to as inverse DCT (IDCT). 〇DCT is often used in image processing, such as capturing features or hiding data 'especially Distorted Image compression method, because it has a strong "energy concentration" nature, most of the signal information will be concentrated in several low-frequency areas of the DCT's image compression technology JPEG, μ jpeg, mpeg, H. 2 6x It is shown that the two-dimensional DCT conversion is used for signal compression, and the compression effect is achieved by quantization and entropy coded. Under normal circumstances, N will be equal to 8, so the result after DCT conversion will be &8<8 conversion coefficient values, where (〇, 〇) represents the % coefficient (Zero-Frequency), and the vertical and horizontal directions The coefficient values indicate a higher vertical or horizontal spatial frequency, respectively. 30 200826686 The two-dimensional DCT formulas are shown in 3-2 and 3-3, where VIII)) is the input block information, and F(",v) is the output data. for w = 〇(3-2 C(w) = ^ λ/2/ΛΓ /orw = l?25...57V-l ^v)=C(M)C(v)||/(,y)cos^cos^ (3.3) The full name of JPEG is j〇int ph〇t〇graphic Experts Group, whose main function is image compression, and then extended to image exchange protocols, scanners, printers, digital cameras, and widely distributed image formats on the Internet. In addition to reducing the image storage space, it also speeds up the transmission speed of images on the network. Static images are mostly used as compression standards. For the flow of jPEG compression, please refer to Figure 13. First, the pixel values are 〇~255 through the hierarchy. Level Offset processing, subtracting the pixel value by 128, so that the value range is between -128 and 127, and then the image is cut into blocks of 仏8, and the discrete cosine transform is used to convert the signal of the spatial domain to the frequency. Since the DCT-converted value belongs to a real number, jpeg quantizes the value into an integer using a quantization table. The converted value is divided into DC and AC. After each block is processed, all DC parts in the image are reused by Differential Pulse Code Modulation (DPCM) to record each DC. The difference is further added to the variable length coding method (Variable Length Coding), and the corresponding data of the DC Huffman table is used to find the coded data corresponding to the difference 31 200826686, and then the compressed bit stream (Bit-Stream) is sent. For the AC part, JPEG uses the Zigzag Scan method to convert the two-dimensional block frequency domain coefficients into one-dimensional type information, and then uses Run-Length Coding to record data continuously. The number of times, finally use the AC Huffman table to find the compression code by using the look-up table method. In general, JPEG compression will undergo a quantization process, and the purpose of quantization is to simplify the amount of stored information required, by the process of quantization. So that the coefficients of the middle and low frequency regions in the original frequency domain are changed to zero, so that the number of coefficients to be stored can be reduced, and the above Dc and AC coding methods are adopted. The quantization table used for quantization and quantization is also divided into two kinds of quantization tables of luminance and color information, as shown in Fig. 14. In the case of JPEG, the quantization part is quantized using the Unif〇rm method. The quantization formula is as follows, where _ is the set quality factor (Quality Factor), W corpse e{i, 2"..i〇〇}, each of the 'prepared JQF is larger, representing the better the quality of the compressed image. '❿JQF approaches zero, the quantized coefficient value contains more zero coefficients, that is, the amount of energy that can represent the image: the less the amount of image quality, the more the image quality is obtained when restoring the image. JQF&gt 50 qs = 2-JQFx〇.〇2 else qs = 50/JQF (3-4) QTable = QTable x qs F(u, v) = IntegerRound (3-5)

Qiu, v) 32 200826686 By means of the quantization table, we can quantize the real value after DCT conversion into integers. In the quantized coefficients, most of the high frequency region is quantized as 〇, and because of the high frequency part of the natural signal The energy is usually less than the low frequency, which means that the information in the frequency domain is less important for the image, so the amount of coefficients needed to be stored can be reduced by losing less high frequency energy. Video compression standards include MPEG (Moving Pictures Expert)

Group) The MPEG series (MpEG_4, MpEG_7, etc.) developed by the organization, and the H. 26x series (Η.263, Η.264, etc.) developed by CCITT. Video is composed of a series of time-correlated images, so the video compression method is estimated using the relationship between the front and the back (Estimati〇n, see Figure 15). The quality of the estimation technique controls the image quality yield after decoding and the compression ratio of the video compression technology. Please refer to the i.5, i6, and 17 diagrams for the H.263 video compression technology. Dividing the picture into

Intra-Frames (I-frames), Inter-Frames (P-ffames) and PB-Frames, I-frame is a separate image content, JpEG, compresses image information applied in a single screen, hence the name Intra, and The P-frame is the facet content estimated by the I-frame; the PB-frame is encoded by two images, so the Macro Block in the PB-frame contains 12 blocks (6 p-blocks, 6 B -blocks); B-block If it is encoded, it must be encoded in Inter, even in one

The same is true for INTRAMB'. H.263 supports five resolutions, in addition to QCIF and CIF supported by η·261, as well as SQCIF, 4CIF and 16CIF. The resolution of SQCIF is about half of QCIF, and the resolution of 4CIF and 16CIF 33 200826686 are 4 times and 16 times of CIF respectively. For the detailed specifications of support, please refer to Figure 16. The DCT coefficient quantization method in H. 263 adopts the non-uniform method. Compared with the JPEG DCT method, Η·263 uses fast DCT to accelerate the conversion of the spatial frequency domain, if it is divided into dct and qdct. For the pre-quantization and quantized coefficients, the DC coefficient of the Intra-Frame can be expressed as: qdct = IntegerRound = IntegerRound] 8 (3-6)

Kstep size y and for other coefficients can be expressed as follows, where scale is an integer value between 1 and 31: qdct = = L; therefore, |_2 heart is also "" (3-7) In JPEG compression, Image quality is quantified by Quantizati〇n

Table) and quality factor (Quality Fact〇r) control, when the image is processed, the result of DCT coefficient quantization will cause the image quality to decrease,

The method of hiding the sub-watermark is to hide the watermark bit to the LSB bit of the pixel value. This method has a great advantage in image quality, it only modifies the lowest bit of the pixel value, so the PSNR value It will be higher than other watermarking methods. In general, the watermarking method in the spatial domain is relatively simple: it mentions: a more convenient way to use, only to adjust the pixel value of the image =: not considered The effect of image processing on pixel values is so that the sub-watermarking method of the spatial domain is gradually ignored. Moreover, this method of watermark hiding has many disadvantages, except that it is not resistant to the basic image processing (filtering, wave, rotation, scaling, etc.) of 34 200826686, and it is impossible to take the opposite direction. Tibetan material watermark information 0 For the image of the valley, the image features can be fully expressed in the frequency domain conversion I, the image itself is rotated, zoomed, noise filtered, etc. does not change the image content. After processing (the so-called not changing the image content here means that the subject to be expressed by the image, the object will not disappear or lose its original meaning after processing), and the original feature information can still be retained, so it is for image processing. The degree of tolerance is higher than that of the floating method in the spatial domain. At the same time, it can more effectively resist the loss of information caused by compression. Therefore, in the image verification method proposed by the present invention, we apply the characteristics of the frequency domain. Digital watermarking and digital signature are used as the basis for information hiding, except that the verification information is hidden in the frequency domain of the original image (DCT doma). In), and save the required recovery information into a digital signature (extra storage space), please refer to Figure 18 and Figure 2. The total energy after discrete cosine conversion is unchanged, so it is not reduced, so DCT itself is a distortion-free spatial domain to frequency domain conversion method. If an image is converted into a frequency domain by DCT, it must be restored to the original spatial domain information by discrete cosine inverse transform (IDCT). Therefore, in our During processing, k{l6,17"..,235},〇)5〇^{16517,""240} Although the same coefficient number is retained before and after the coefficient conversion, the pixel value at the time of IDCT conversion sometimes Will exceed the allowable range (〇~255), first fix the YCbCr coefficient value to reduce the error caused by the inverse conversion. 35200826686 Poor' Then we cut the image into several blocks that do not overlap each other, and for each The block is transformed into DCT, and the coefficient value after the (10) conversion will show the phenomenon of energy concentration, and the distribution of energy is more concentrated. Since the DCT conversion is a distortion-free conversion process, the inter-block relationship is before the conversion. After the conversion, the consistency will be maintained. If the DC coefficient value % of the first block is greater than the Dc coefficient value % of the second block, the discrete cosine transform and inverse transform of the ~: will retain the same property. The ^ term characteristic can be derived to the quantized dct coefficient value. In the theory proposed by Ching-Yung Lin and Rufu (5) 叩 [5] _ [7], when the DCT coefficient value of the block is quantized, In the present invention, the eigenvalues of the sink coefficient and the AC coefficient are calculated separately, and in the part of the De coefficient, we use the current block 1 as the medium~ To expand outward, select other m blocks as the basis for comparison. The distance between each block is Di stb. The larger the D 丨 s ^ b is, the lower the correlation between the blocks is. When ρ 丨s 7 b is 1, the range selected is 3 > 0, and then the central block is calculated. The difference between the dc coefficient value ης and the other m DC coefficient values DCy, where 7 L ′···′Μ, if the difference is greater than the set threshold value, the DC eigenvalue of the block is 1, otherwise Why?

After determining the extent and number of blocks to be referenced, the feature values of the block are extracted for DC and AC coefficients, respectively. The characteristic extraction method of D(: coefficient is as follows: 36 200826686 ifDCi>\/DCjJ^{\,2,...m} the feature bit (Featurefc) of DC. else reature^ (3-8) and when Distb= l time (please refer to the Macr〇 shown in Figure 2 (a)

Block) will make it possible to increase the speed of extracting the authentication bit if it is equal to 〇. If Feature fc = 1 then Feature^ = 〇(3-9) else Feature^ -〇or\ We use the above method to extract the DC coefficient feature in the image block, one image of the size of ><512 It is said that a total of (feyh23) #DC coefficient feature values are taken out, and the non-〇 of each DC coefficient feature value is 1. In addition, for the AC coefficient (see Figure 20), the 3X3 block of (a) in the mouth chart is used as the working block, and four AC eigenvalues are calculated in each block. Representation, where i is the index value of the current block, and j is the obtained feature index value, #{ο,ι,2,3}. When calculating the eigenvalues, four eigenvalues are calculated using the horizontal, vertical, and diagonal g formulas based on the intermediate block i (see Figure 21). Each eigenvalue is not 〇 1 The eigenvalues are calculated as shown in Equations 3-10 to 3-13, where a' 丨 is the number of non-zero AC coefficients in the horizontal (vertical) and vertical (verticai) (see Figure 2) ), for the complete certification information extraction process, please refer to Figure 2 3. If (NQAC.h >NQAC2h)&(NQAC,h> NQAC^h)

Feature^ = 1 else (3-10)

Featurefc = 〇 37 200826686 (3-11) if{NQACiv>NQACj&(NQACUv>NQACj Feature^ =1 else

Feature^ = 0 (3-12) if(NQACuh>NQAClh)&{NQAC,h > NQACj &(NQACih >NQAC6,h)&(NQACih> NQACJ Feature^ = 1 else

Feature^ = 0 if{NQACiv >NQACj&(NQAC.v > NQACj ^QACiv>NQACj&^QACiv>NQAC,v) (3-13)

Feature^ =1 else

Featuref^ = 0 When selecting the hidden position, taking into account the effect of the DCT coefficient change on the image quality, on the DC coefficient, choose the Lookup Table method to determine the DC coefficient value to be changed when the authentication bit is hidden. Μ Figure 2 4 shows. For the hiding method of the AC coefficient, in order to reduce the influence of the judgment formula in calculating the AC characteristic value, the AC characteristic value is selected to be hidden in the four positions of "c?, dCll'dCl3 of the Zig-zag scan, see the second 5th. The figure shows.

Describe the DC coefficient, assuming that the feature is to be hidden in the kth coefficient, then we refer to the kth quantized coefficient of the QLB to adjust the coefficient value of the DCT to hide the feature in the image. Suppose the circle in the figure indicates the original DCT coefficient value of the hidden position. If the hidden feature bit is “ Γ , the DCT coefficient is adjusted to 3×L (9); if the hidden feature bit is “0”, the DCT will be 38 200826686 The coefficient is adjusted to (4). Similarly, the same hiding method can be used for the AC coefficient'. The point to note is that when the Ac coefficient changes, it may change to the block calculated when the authentication bit is generated. For the potential problem, we will use a boundary value for the tampering (4). When performing image authentication, use the same pre-processing method as the hiding method to convert the image to The YCbCr color space is normalized for the luminance component and the color component, and then the image is cut into 仏8 blocks and subjected to DCT conversion, and then calculated for each block using the same calculation method as the extraction block verification bit. The characteristic value of the DC and the carrier coefficient. The above mentioned invention mentions that when the authentication bit is hidden to the low frequency region of the AC coefficient, the non-zero Ae when calculating the AC coefficient feature of the block may be changed. In view of the above potential problems, in the AC coefficient eigenvalue calculation of the verification stage, the formula of the formula 3-1Q is modified to add a critical value, in order to solve the above problem. The corrected AC coefficient eigenvalue calculation method will be as follows: · if WACUh - NQAClih) + [NQACUh -NQACl tl)> 2t Feature^ = 1 else (3-14)

Feature^ = 〇 where the critical value i depends on the number of authentication bits to be hidden. In the course of the experiment of the present invention, dip is a better empirical value, as shown in Figure 26. Next, the authentication bit element hidden in the image is obtained from the image. According to the invention, 39 200826686 'receives the DC system, and the DC and AC coefficients use the authentication bit of the table check method of the 25th figure, and m ACs. The authentication bit of the coefficient. For the DC coefficient, due to the DC coefficient of its own block,

Because the Dc authentication bit is hidden when hiding, it is easy to take out the hidden DC authentication bit', and in the authentication bit of the AC coefficient, the private key is hidden in the authentication bit. Select the corresponding block to perform the hidden action, so

The correct block can obtain the authentication bit of the AC coefficient from the relative block. Since the coefficient value after DCT conversion will exhibit energy concentration, in the DCT-based compression technology, the z-type scanning method (zig-Scan) is the most common coefficient scanning method, which can be used first by Zig-zag scanning. Obtaining the coefficient value of the larger energy, the coefficient of the smaller energy can be abbreviated. After the transformation, it seems that only a small amount of data is needed to represent the Zig-zag, except the JPEG image compression technology. , Η · 261 / Η. 263 In the Intra-Frame encoding method, the addition of alternating vertical scanning (Alternate-Vertical Scan) and alternate horizontal scanning resistance (Alternate-Horizontal Scan) two different coefficient scanning methods. In the above-mentioned restoration method, the present invention stores the information required for restoring an image by means of a digital signature. First, the horizontal and vertical months b畺' are calculated for the block, and the energy distribution characteristics of the block are used differently. The coefficient scanning method takes the left quantized coefficient value and calculates the difference from the coefficient value taken out from the previous block, and stores the important number of the block 200826686 in the DPCM manner, which can reduce the required number. Store the amount of data. Moreover, according to different block energy distribution situations, the present invention uses different coefficient scanning methods to solve the energy loss caused by storing only the coefficient values of the fixed positions of the low frequency region when scanning in z-shape. In the process of selecting the block recovery information, the DCT of the image is first converted to the block. The converted Dc coefficient value will be between [―(9) you (9)47], and the AC coefficient value will be between [ - between Tao Tao 23], if there is no special place to store DC, the product will use id 12 bits, and an AC coefficient value occupies 11 bits, so if it is from 8><8DCT block Select a dc coefficient and m AC coefficients as the recovery information. The storage space occupied by the single-block is (i2 + mXll) bits. The light is a heart attack image. Select one coefficient and five AC coefficients. It requires (ΐ2 + 5χΐι)χ212 bits, which is equivalent to 34,304 Bytes. This kind of recovery information takes up too much storage space. If it is a continuous image content, it does not meet the needs of streamlining. Furthermore, it has been found during the course of the experiment of the present invention that the retention of five coefficients is not sufficient to represent the content of the block, and therefore, the present invention uses different bit retention patterns for different block properties. First, the flatness index is calculated for each SxSDCT block. The definition of the flatness index is as follows: Di, (3-15) where Qiao and & are the vertical and horizontal flatness indicators of the block, respectively, a and b are the scaling factors. , used to adjust the ratio of vertical and horizontal flatness to obtain the flatness index of the block. In general, 1, and Fv and 仄 are defined as follows, ~() is the flatness test function of adjacent pixels·· 41 200826686

Fv=iF^)=iiFM(h+ui, /=〇/=〇y=o Fh=iF^hitMiuj+^^ /=〇i=n /=n where FM(x) - i=0 j=0 fl3x&gt ;r 1〇,χ<Γ (3-16) In general, if the mean square error of the image is large, the critical value T is also larger, otherwise the mean square error is smaller and the critical value τ is smaller. The critical value τ can be determined by the following formula: σ2 =^ΣΣ(/^-^)2^=^ΣΣ/υ· , 04 /=〇y=o 64,,y Γ = σ2/δ + 1 (3- 17) When the obtained flatness index Fv-仄>7> and Q>7>2, the present invention selects the method of using the vertical scanning in turn (see (7) shown in Fig. 27) to select the desired The m AC coefficient values are used. Conversely, when 尽-仄>7> and 夂>7>2, the horizontal scanning method is used alternately (please refer to (b) shown in Fig. 27). To smooth the block (see Figure (a) in Figure 27). Obtain the DC coefficient and the Ac coefficient of the melon, where ^{^••,^ use the difference method to record the current coefficient and The difference of the previous coefficient; first calculate the difference between the strict and the thin < the recovery information stored in the block and then filter out each仏Core _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ Encrypt the action, and save it to the extra storage space, please refer to Figure 28. When the receiving end receives the image, the field uses the verification program to determine whether the content has been illegally falsified (four) - When the suspicious area 42 200826686 is found in the image, the required restoration information is obtained from the received recovery information. The restoration information obtained from the digital signature includes the DC coefficient of the block and two The AC coefficient, plus a block horizontal/vertical decision bit, is used to fill the DC/AC coefficient values in the 8x8DCT block for the different block properties by using the horizontal scan or the vertical scan, respectively. Convert and then fill in the block that detected the error. In the test sample of the present invention, different types of image content are used, including Baboon images with more details (see Figure 29 (a)) , A larger image of the peter in the sliding zone (see Figure 29 (c)), and a Lena image with the same proportion of the smoothing zone as the smoothing zone (see Figure 29 (b)), plus The five monitors captured by the DVR system (see Figure 29 (d~h)) and four self-portrait images (see Figure 29 (i~1)) In order to achieve a fair and objective test angle, in addition to the test image with high brightness, it also contains images taken at night. The characteristic is that the brightness in the face is darker, and the bright part contains more noise. In terms of evaluation criteria, the present invention uses PSNR (Peak Signal to Noise Ratio) as a basis for evaluating the performance of the algorithm; the unit of pSNR is 疋dB, and the difference between the higher the signal-to-noise ratio and the original surface is smaller. The closer to the original surface, the better the quality of the compression. The definition of psNR is as follows: PSNR = 201og1〇

255 MSE MSE = Σ\/(^Αυ)]2 Ν~2 (4-1) where MSE (Mean Square Error) is the mean square error of the original signal and the signal after the inclusion of 43 200826686 information. Generally, PSNR Below 25dB, most people can't be satisfied with such quality; higher than 35dB, generally think that the picture quality is good; above 40dB, if you do not freeze, you can't usually distinguish it from the original picture. In addition to using PSNR as the evaluation standard, the present invention also uses the error detection rate to evaluate the accuracy of the proposed image verification method. In the case of false detection, there are two definitions. 1. Loss detection (Miss) -Detection Rate) ·· Has been tampered with, but not taken out. I have not been tampered with 2. Error detection (Error_Detecti〇nRate) and mistakenly thought that the block was tampered with. . Because in the application of image verification, it is mainly to avoid the situation that the ghosts of the Xiachang District cannot be detected. 纟 Some scope of error detection can be seen in our evaluation standard, which is the first species. Definition (Leak 2 is used as the verification item for the accuracy of the different method. The leak detection rate used is calculated by the 4-2 type. Loss detection rate = (no detected tampering number / detected) The number of tampering) (4-2) In the analysis of the energy distribution, the indicators, first of all; & mentioned the flatness of 1 〇, after the autumn heart Η 263 Quantification, quantized gradation for the second application 3 12~3-14 to calculate the flatness index. After completing the calculation, use 卩捡8 & ^ and ping, 、, and 4 knives as the horizontal detail area and the vertical detail area, using red and blue respectively. Color and black are marked, please refer to 200826686, see Figure 30, which is the result of horizontal energy and vertical energy distribution experiments. See Figure 31 for details, except for smooth and detail blocks in the image. In the detail block, it is still classified into a horizontal detail area and a vertical detail area. In the invention of the recovery information extraction, three different coefficient scanning methods are used, so that the selection of DCT coefficients can be more accurately performed. This is also the scanning used in the H. 263 and MPEG series video compression technologies. In the experimental evaluation of the present invention, the image quality after hiding the watermark is the main consideration; the basic requirement of the digital watermark is the image quality after the information is hidden, and there is not much visual difference with the original image. Therefore, the present invention uses the PSNR value as a basis for evaluation, and the image quality after hiding the information can be found, and the proposed watermark hiding algorithm can effectively retain the image information 'avoiding the original watermark after hiding the watermark There are too many differences. Please refer to Figure 3 2 and Figure 3, which are the images after watermarking and the image quality (pSNR), in which the image quality after watermarking is hidden. We compare it with the other three methods (marked by *), followed by Hsien-Chu Wu and Chin-Chen Chang 2002 [9], Yulin fang and Alan pearmain 2〇〇4 [13] and

Ching-Yung Lin and Shih-Fu Chang 2001 [6]. In the verification process of the present invention, firstly, the acquired authentication information is obtained from the image to be verified, and the feature information of the block is calculated. Similarly, the present invention performs experiments from 12 test images and calculates errors. For the proportion of loss detection, please refer to Figure 34 and Figure 5, which is the image test 45 200826686, Figure 3 4 (b), which is the result of the tampering test. Figure (a) is the original image, which is the tampering image after adding the watermark, Figure 34 (c tampering area (red mark), picture 34 (d) original image. From the experimental results, it can be found that The image verification method used can effectively mark the tampering block, reduce the possibility of missed detection, and effectively restore the tampered block to approximate the original image through the block restoration method. Image authentication and restoration technology can be applied to various image authentication processing systems, digital video recording devices such as DVR (Digitai)

ReC〇rder)......etc., and is applicable to various compression formats, such as 26. 26X series, MPEG series, JPEG, JPEG2000, etc., and the present invention can be stored in a computer readable medium (such as a magnetic disk, a compact disc). The program in the film, and the like, is achieved by mounting the program from the aforementioned medium to a computer system. The DVR monitoring system with this tamper-tested and certified restoration can be installed in general buildings and important institutions, and can be extended to the establishment of a community safety main system to enhance the added value of the product and enhance its international competitiveness. In relation to the concept of the present invention, in August 2006, at the International Conference on Innovative Computing (Information and Control) held in Beijing, China, by Chen Zhaohe ( "An Effective 46 200826686 by Th〇u-Ho Chen), Chung-Yih Lee, Tsong-Yi Chen and Da-Jinn Wang

Authenticating Method on the Compressed Image Data (pp. 249_252) and published. [Simple description of the figure] Figure 1 shows the center point and adjacent points under different size blocks. Figure 2 shows the Dc values of the four blocks using up, down, left, and right to estimate the c value of the middle block. Figure 3 shows the image restoration flowchart proposed by J. Fridr ich and M. Gol jan. Figure 4 is (a) is the block coefficient before the CT conversion; (b) is the coefficient after the block is converted to the frequency domain by DCT; (c) is the jpEG standard quantization table; (d) is the use of quantization The table will be shown in Figure 4; (b) the quantized coefficient values. The picture shows the coding matrix proposed by J. Fridrich and Μ Gol jan. Figure 6 (a) is the original image; (b) is the feature hiding using the method proposed by Fridrich and M·G〇i; (c) is the license plate that has been tampered; (d) is the restored license plate. For 1^1611-(]111^11 and (:1^11-(:11611(]11& called the proposed feature selection flow chart. Figure 8 is the edge detection result of Lena. The qth, the picture is the use The experimental results of the method proposed by Hsien-Chu Wu and Chin-ChenChang. The $1 π map is an image restoration method proposed by Hsien-ChuWu and Chin-ChenChang. 47 200826686 The first picture is the generation method of feature information. 1 2 is the experimental result of the method proposed by Phen-Lan Lin. Figure 13 is the JPEG compression flow chart. Figure 14 is the JPEG compressed DCT coefficient quantization table; (a) luminance information quantization table; (b) color Information quantization table. Figure 15 is the picture estimation method in video coding. Figure 16 shows the image format supported by H.263 video coding. Figure 17 shows the H-263 video coding Intra-Frame.

Inter-Frame, and PB-Frame. Figure 18 is a method for hiding watermarks proposed by the present invention. Figure 19 is a video verification and restoration procedure proposed by the present invention. Figure 20 shows the block selection method. Figure 21 shows the eigenvalue calculation method. Figure 2 2 shows the calculation of the non-zero AC coefficient. Figure 2 3 shows the flow chart of the certification information extraction. Figure 24 shows the table lookup method. Figure 25 shows the selected AC coefficient hiding position. Figure 26 is a flow chart of image verification. Figure 27 shows the DCT coefficient scanning method. Figure 28 is a flow chart for selecting image restoration information. Figure 29 is the image used in the experiment. Figure 30 is a statistical chart of block classification. Figure 31 shows the results of block classification. 48 200826686 Figure 3 2 shows the image after hiding the authentication information. Figure 3 3 is an image quality chart after hiding the authentication information. Figure 34 shows the results of image verification and restoration. Figure 35 is a block verification chart. [Main component symbol description] 49

Claims (1)

200826686 X. Patent application scope: 1. A method for image authentication and restoration, which comprises the following steps: converting the image to be authenticated into a YCbCr color space; setting the coefficient value of YCbCr (}^{16,17,···, 235}, α?, α^{ΐ6,17,···,24〇}); cut the image into a plurality of 8x8 blocks that do not overlap each other; perform DCT conversion on each block; select and calculate Select the DC and AC coefficient characteristics of the block; obtain the array coefficient value; determine the authenticity of the set of coefficient values; if the coefficient value is reasonable, the image is not falsified, and it is a real image; if the coefficient value is incorrect The image has been tampered with and the following steps are required; the original image is converted to the YCbCr color space; the coefficient value of YCbCr is set (7e{16,17"."235}, C6, Oe{l6,17"."24〇 }) · Cut the image into a plurality of 8x8 blocks that do not overlap each other, and calculate the horizontal and vertical flatness indicators for the parent 8×8 blocks; perform DCT conversion for each block; classify the DCT blocks; DC and AC coefficient recovery information; stored in DPCM The quantized coefficient value; encrypts the DC and AC coefficient recovery information; and fills the encrypted restoration information into the block of the falsified image. 2. For the method of image authentication and restoration as described in item 1 of the patent application, wherein the selected block and A (: coefficient 50 200826686 sign, the DC and AC coefficients are each block) i is the center, and the other m blocks are selected as the basis for comparison; the distance of each block is Distb. When the value of Distb is 1, it indicates that the selected block is 3x3. The larger the Distb value, the more The lower the block correlation; the difference between the juice center block' and the other m DC coefficient values, where 4{l, 2"."m}, if the difference is greater than the set threshold, Then, the DC eigenvalue of the block is !, otherwise it is 〇; the AC eigenvalue acquisition method is also the same. 3. The method for image authentication and restoration according to claim 1, wherein the horizontal and vertical flatness The degree index is defined as factory=<+, &, b is the scaling factor, and F v and Fh are defined as follows: FM () is the flatness test function of adjacent pixels: τ is the critical value and can be determined by the following formula: f = cr /0 丄 丄 所 所 平坦 平坦 F rr rr rr rr rr h h h h h h h h 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直 垂直Select the method of rotating horizontal scanning; eight >?> and Q>7>2, select to use as smoothing block. The value of the AC coefficient; the way of the other; in other cases, the use of the patent model 4, an image authentication processing system, operates according to the method described in item 1. 51 200826686 5. A digital video recording device that operates using the method described in item 1 of the patented vehicle color. 6. A computer readable medium storing a program code for causing a computer system to execute a method of image authentication and restoration as described in the patent application U.S. Patent Application "I g Μ f π yoke. 52 200826686 Attachment: [1] . Tsong-Yi Chen, Chien-Hua Huang and Thou-Ho Chen, “Authentication of Lossy Compressed Video Data by Semi-Fragile Watermarking”, IEEE 2004 International Conference on Image Processing (ICIP Singapore), 2004 [2] . Kwang-Fu Li, Tung-Shou Chen and Seng-Cheng Wu, '' Image tamper detection and recovery system based on discrete wavelet transformation" , Communications, Computers and signal Processing, 2001 IEEE Pacific Rim Conference on, Vol 1, Aug. 2001, Page(s): 164 -167 vol. 1· [3] . Min Wu and Bede Liu, “Watermarking for Image Authentication”, Proceedings of IEEE International Conference on Image Processing, Oct· 4-7 , 1998, Chicago, Illinois, USA, Vol· 2, Page(s): 437-441· [4] · Tae-Yun Chung, Min-Suk Hong, Young-Nam Oh, Dong-Ho Shin and Sang-Hui Park , "Digital Watermarking for Copyright Protections of MPEG2 Compressed Video", IEEE Trans. on Consumer Electronics, Vol. 44, No. 3, Aug. 1998, Page(s): 895-901. [5] · Ching-Yung Lin and Shih-Fu Chang, "Semi-Fragile Watermarking for Authenticating JPEG Visual Content", SPIE International Conf. on Security and Watermarking of Multimedia Contents II, 53 200826686 Vol. 3971, No. 13, El J 00, San Jose, USA, Jan 2000. [6] . Ching-Yung Lin and Shih-Fu Chang, “SARI: Self-Authentication-and-Recovery Image watermarking system”, ACM Multimedia 2001 Workshops - 2001 Multimedia Conference, 2001, Page ( s): 628-629. [7] Ch Ch Ch Ch Ch Ch Ch Ch Ch Ch Ch Ch Ch , El '99, San Jose, USA, Jan 1999. [8] . J. Fridrich and M. Goljan, “Images with Self-Correcting Capabilities”, IEEE International Conference on Image Processing (ICIP '99), Vol· 3, Oct· 1999, Page(s): 792-796· [9] . Hsien-Chu Wu and Chin-Chen Chang, “Detection and Restoration of Tampered JPEG Images” The Journal of System and Software, Vol. 64, NO. 25 15 Nov. 2002 Page(s): 151-161. [10] . Phen-Lan Lin, Po-Whei Huang and An-Wei Peng, “A Fragile Watermarking Scheme for Image Authentication with Localization and Recovery" , Multimedia Software Engineering, 2004. Proceedings. IEEE Sixth International Symposium on 13-15 Dec. 2004, Page(s): 146-153· 54 200826686 [11] . Chung-Shien Lu and Hong-Yuan Mark Liao, "Structural Digital Signature for Image Authentication: An Incidental Distortion Resistant Scheme", IEEE Transactions on Multimedia, Vol. 5, No. 2, Jun. 2003. [12] . Choi, Y. and Aizawa, I ., "Digital watermarking using interblock correlation", In: Proc. International Conference on Image Processing (ICIP J 99), Vol. 2, 24- 28 Oct. 1^99, Page(s): 216 - 220· [13] Yulin Wang and Alan Pearmain, "Blind image data hiding based on self reference" , Pattern Recognition Letters, Vo. 25, Issue 15, Nov. 2004, Page(s): 1681-1689. [14] · LM Marvel, G . Hartwigand C. G. Bonce let Jr., "Compression Compatible Fragile and Semi-Fragile Tamper Detection", SPIE International Conf. on Security and Watermarking of Multimedia Contents II, Vol. 3971, No. 12, El ' 00, San Jose, USA , Jan 2000· [15] . S. Bhattachar jee and M. Kutter, “Compression Tolerant Image Authentication”, IEEE International Conference on Image Processing (ICIP '98), Chicago, USA, Oct 1998. [16] · Phen-Lan Lin, Chung-Kai Hsieh and Po-Whei Huang, “A Hierarchical Digital Watermarking Method for Image Tamper 55 200826686 Detection and Recovery”, Pattern Recognition, Vol. 38, Issue: 12, Dec· 2005, Page(s): 2519- 2529,· [17] . Hae-Yeoun Lee, Heung-Kyu Lee and Junseok Lee, “Comparison of Feature Extraction Techniques for Watermark Synchronization”, 9th International Conference, KES 2005, Vol. 3, Sep. 2005, Page(s) : 309-316. [18] . M. KutterandF. AP Petitcolasb, ff A fair benchmark for image watermarking systems11, Electronic Imaging 1999. Security and Watermarking of Multimedia Contents, Vol. 3657, Sans Jose, CA, USA, 25-27 Jan. 1999. The International Society for Optical Engineering. [19] . Sviatoslav Voloshynovskiy, Shelby Pereira, Victor Iquise and Thierry Pun "Attack modeling: Towards a second generation benchmark" , Signal Processing, Special Issue: Information Theoretic Issues in Digital Watermarking, May 2001. [20] . Jiri Fridrich, Miroslav Goljan, Rui Du, "Invertible Authentication Watermark for JPEG Images" , Information Technology: Coding and Computing, Proceedings. International Conference, Apr. 2001, Page(s): 223-227. [21] · Lin Yiji, Lai Xisong, Digital Watermark Technology, Information Security Newsletter, Volume 4, Issue 3 , 56 200826686 1998. 06. [22] · Zhang Chengxin, Chen Tongxiao, Huang Guofeng, Digital Image Processing Technology, Songgang Computer Books and Materials Co., Ltd., 2001, Taipei. [23] . Rafael C. Gonzalez and Richard E. Woods, Digital Image Processing 2nd Edition, Prentice Hall , 2002. [24] . Digital Watermarking World, http://www. watermarkingworId. org/ [25] · The Checkmark Watermark Benchmarking Project, http://watermarking, unige. |ch/Checkmark/index, html [26] The Sti mark Benchmark, http://www. petitcolas. net/fabien/watermarking/ stirmark/ 57