JP2007221388A - Data inserting device and method, and image compressing coding device and method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To prevent deterioration in image quality, an increase in data quantity to be embedded in image data and scale-up of a file in embedding optional data into the image data. <P>SOLUTION: A quantization representative value R1 (X) of a division result value D(X)/Q1(X) of a DCT coefficient D(X) of an original EOB is changed to a value R2(X) most approximate to the division result value D(X)/Q1(X) out of -1 and 1, as a data finish marker. Since the data finish marker is at a low-frequency side by one coefficient position of a new EOB, a decoder side can discriminate an insertion bit "0" from a zero quantization representative value sequence under EOB (end of block) by detecting the data finish marker in sequentially decoding DCT coefficients from the low-frequency side. The new EOB advances by one coefficient position to the high-frequency side from the original EOB, however, the EOB itself never disappears. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a technique of embedding arbitrary data in compression-encoded still image data or moving image data.

  Various compression algorithms have been proposed in the field of image data compression. For example, in the case of JPEG compression, the original image data is subjected to discrete cosine transform (DCT), bit allocation is determined, and quantization is performed. In the case of the baseline method, the quantized data is encoded by Huffman encoding to obtain encoded data.

  As a technique for adding arbitrary data to the compression-encoded image data, there is steganography. Steganography is a process that directly superimposes additional information (messages, images, etc.) on images, etc. For example, data is inserted into the least significant bit of bit information indicating the value of each pixel of image information. (Patent Document 1), or changing the DCT coefficient of the quantized Y component of the macroblock (Non-Patent Document 1), or embedding data by the remainder obtained by dividing the sum of the quantized DCT coefficients by a certain value (Non-Patent Document 2) representing various aspects.

  As an information embedding technique similar to steganography, there is a digital watermark. For example, one or more DCT coefficient blocks (embedded blocks) for embedding a digital watermark are selected from the DCT coefficient block group, and one bit constituting the digital watermark is embedded in each selected embedded block ( Patent Document 2). However, this technique is used for information addition and copyright management of the image itself, and the amount of data embedded as a digital watermark is small, and resistance to data modification is required. Also, extraction of embedded data is not necessarily 100% possible, and differs from steganography in these respects.

Moreover, image quality degradation may occur due to data embedding. As a technique for preventing this, for example, there is a technique in which a place where an image is embedded is positioned at the highest frequency of the DCT coefficient (Patent Document 3, Non-Patent Document). 3).
JP 2003-289435 A JP 2000-151973 A JP 2002-330279 A JPO Standard Technology Collection “Steganography for JPEG Coded Sequences” Internet URL [http://www.jpo.go.jp/shiryou/s_sonota/hyoujun_gijutsu/denshi_sukashi/2_e_2.htm] “Data Embedding Method in JPEG Image That Does Not Need Specification of Embedding Position”, IEICE Transactions 2005/10 Vol. J88-D-II No. 10 pp. 2037-2045, The Institute of Electronics, Information and Communication Engineers “Data Embedding Binary Data in JPEG Coded Sequence”, IEICE Transactions 2000/6 Vol. J83-D-II No. 6 pp. 1469-1476, The Institute of Electronics, Information and Communication Engineers

  In the techniques of Patent Document 1, Non-Patent Document 1, and Patent Document 2, only one bit can be embedded per macroblock, and the amount of data that can be embedded is limited. Even in the technique of Non-Patent Document 2, the amount of data that can be embedded is limited. Further, as in Patent Document 3 and Non-Patent Document 3, since the EOB (End of Block) code disappears, the encoding efficiency is reduced and the JPEG file becomes very large.

  An object of the present invention is to prevent image quality deterioration, increase the amount of data to be embedded in image data, and prevent an increase in file length when embedding arbitrary data in image data.

  A data insertion device according to the present invention inputs a DCT coefficient input unit that inputs DCT (discrete cosine transform) coefficients calculated for each of a plurality of blocks obtained by dividing image data, and is input to the DCT coefficient input unit. Of the DCT coefficients, the DCT coefficients from the start position, which is a predetermined low frequency side coefficient position at which data insertion is started, to the end position, which is a low frequency side coefficient position by one coefficient position from EOB (End of Block). Each of the least significant 1 bits of the quantized representative value, which is a value obtained by rounding the division result value obtained by dividing each by the corresponding quantization width, is changed to a bit constituting arbitrary data to be inserted, and A data insertion unit that changes the quantization representative value of the DCT coefficient of the EOB to a value closer to the division result value of the DCT coefficient of the EOB of −1 or 1.

  According to the present invention, the image data is embedded by changing the least significant 1 bit of each DCT coefficient calculated for each block obtained by dividing the image data to a bit of arbitrary data. Instead of inserting 1-bit data into each block as in the prior art, it is possible to insert 1-bit data into each DCT coefficient, which greatly increases the amount of data that can be inserted into image data.

  In addition, as will be described later, if the quantization width of the DCT coefficient in the range where data is likely to be inserted is set to a value smaller than 1/2 of a normal quantization width where no data is assumed to be inserted, Since the displacement of each DCT coefficient is 1 bit at most, the influence on the image quality is small.

  In addition, since the quantization representative value of the EOB DCT coefficient is changed to a value closer to the division result value of the EOB DCT coefficient of −1 or 1, the quantization representative of the insertion bit “0” and the EOB or less is represented. A column with a value of 0 can be distinguished.

  The original EOB quantized representative value is changed to the end marker, so that the new EOB is moved up by one coefficient position from the original EOB, but the EOB itself disappears as in the prior art. However, since it is only shifted by one coefficient position, the decrease in coding efficiency and the increase in code amount due to the decrease in zero run length are not large.

  The data insertion unit is a value obtained by dividing the quantization representative value of the EOB DCT coefficient by −1 or 1 by a predetermined value equal to or less than ½ of the quantization width corresponding to the EOB DCT coefficient. It is preferable to change to a value closer to.

  By doing this, the noise power of the error due to changing the quantization representative value to a non-zero data end marker is equal to or less than the noise power of the quantization error of the DCT coefficient of the EOB.

  In the present invention, image data is divided into a plurality of blocks, and a DCT transform unit that calculates DCT (discrete cosine transform) coefficients for each block, and a DCT coefficient of each block is divided by a corresponding quantization width A quantization unit that determines a quantization representative value of each DCT coefficient by rounding a division result value that is a value, a quantization data table that stores the quantization representative value of each DCT coefficient, and creates a quantization data table. The present invention relates to an image compression encoding apparatus including an encoding unit for encoding.

  Here, the quantization unit starts from a start position that is a predetermined low frequency side coefficient position where data insertion is started to an end position that is a coefficient position that is one coefficient position lower than EOB (End of Block). Change each of the least significant 1 bit of the quantized representative value, which is a value obtained by rounding the division result value obtained by dividing each DCT coefficient by the corresponding quantization width, to a bit constituting arbitrary data to be inserted. In addition, the quantization representative value of the DCT coefficient of EOB is changed to a value closer to the division result value of DCT coefficient of EOB, which is −1 or 1.

  The data insertion method according to the present invention includes a step of inputting a DCT (discrete cosine transform) coefficient calculated for each of a plurality of blocks obtained by dividing image data, and a step of inputting a DCT coefficient input to a DCT coefficient input unit. Among them, each DCT coefficient from the start position which is a predetermined low frequency side coefficient position where data insertion starts to the end position which is a low frequency side coefficient position by one coefficient position from EOB (End of Block) is supported. The least significant 1 bit of the quantized representative value that is a value obtained by rounding the division result value that is a value divided by the quantization width to be changed is changed to a bit that constitutes arbitrary data to be inserted, and the DCT of the EOB Changing the coefficient representative quantization value to a value closer to the division result value of the DCT coefficient of EOB among −1 or 1.

  An image compression encoding method according to the present invention performs block formation by dividing image data into a plurality of blocks, calculates DCT (discrete cosine transform) coefficients for each block, and a quantum corresponding to the DCT coefficients of each block. Rounding a division result value, which is a value divided by the quantization width, and determining a quantized representative value of each DCT coefficient, and starting from a start position that is a predetermined low frequency side coefficient position at which data insertion is started, EOB (End Quantization representative value that is a value obtained by rounding a division result value that is a value obtained by dividing each of the DCT coefficients up to the end position that is a coefficient position on the lower frequency side by one coefficient position than of the block). Each of the least significant 1 bits is changed to a bit constituting arbitrary data to be inserted, and the quantization representative value of the DCT coefficient of EOB is set to −1 or 1, which is the DCT coefficient of EOB. A step of changing to a value closer to the division result value of the number, and a step of creating a quantized data table storing the quantized representative value of each DCT coefficient and encoding the quantized data table.

  According to the present invention, the quantization representative value of the DCT coefficient of EOB (End of Block) is changed to a value closer to the division result value of the DCT coefficient of EOB of −1 or 1, so that the insertion bit “0” ”And a column of quantization representative value 0 below EOB can be distinguished.

  Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

<First Embodiment>
FIG. 1 is a block diagram of an image compression coding apparatus according to the first preferred embodiment of the present invention. Although details are omitted, each block of the image compression encoding apparatus described in the present embodiment can also be realized by program control.

  The image data creation unit 11 performs gamma correction of the image signal output from the image signal output unit 10 such as an analog front end built in the digital still camera, defective pixel interpolation by pixel interpolation, white balance correction, image signal Data for each color of RGB representing a subject image is generated by performing processing such as sharpening of the image, correction of the influence of camera shake, and color space conversion from the CMYG color space to the RGB color space, and linearity by a 3 × 3 matrix The image data is output after conversion into the YUV color space of the luminance signal (Y) representing luminance, the color difference signal (U) representing the “blue” component, and the color difference signal (V) representing the “red” component by the conversion.

  For example, Y = 0.2990R + 0.5870G + 0.1140B U = −0.1687R−0.3313G + 0.5000B + 128 V = 0.5000R−0.4187G−0.0813B + 128 is used for conversion to the color space of YUV. be able to.

  A black and white image may be considered as U = V = 0, and the present invention can be applied to both a color image and a black and white image.

  The DCT processing unit 12 divides the image data output from the image data creation unit 11 into a plurality of blocks each consisting of 8 × 8 elements (here, 64 blocks for convenience of explanation), and each block DCT (Discrete Cosine Transform) is performed. The DCT processing unit 12 creates a DCT coefficient table in an 8 × 8 table format storing DCT coefficients for each block, and outputs the DCT coefficient table to the quantization unit 13.

  The quantization unit 13 divides each DCT coefficient of the DCT coefficient table for each block output from the DCT processing unit 12 by a corresponding section width (referred to as a quantization width or a quantization step) of the quantization table, and the value Each DCT coefficient is quantized by rounding off the decimal places.

  The quantized representative value Q is, for example, x: a value of a DCT coefficient before quantization, Δ: a quantization width, g: a function that performs rounding operations such as rounding, rounding down, and rounding up.

  It is represented by In the mid-tread linear quantization, g is a function that performs rounding processing. In JPEG etc.

And Δ are encoded separately. In the present specification, for convenience of explanation, a value obtained by Equation 2, that is, a value obtained by quantizing a DCT coefficient is referred to as a quantized representative value.

  The DCT processing unit 12 creates an 8 × 8 table-format quantized data table storing the quantized representative values of the DCT coefficients determined in this way for each block, and outputs them to the entropy coding unit 14.

  The entropy encoding unit 14 entropy-encodes the quantized data table for each block output from the quantization unit 13 using a Huffman code table or the like, and outputs a bit stream. At this time, the number of bits is effectively reduced by performing encoding in consideration of the length (zero run length) of a sequence of zero quantized representative values continuous up to the coefficient position of the highest frequency.

  An EOB (end of block) code is added to the encoded data. Here, EOB is a code meaning that all subsequent quantized representative values in the block are zero.

  In the DCT coefficient table obtained by performing DCT, a low spatial frequency DCT coefficient in which deterioration of image quality is likely to be manifested is arranged in the upper left, and a high spatial frequency DCT coefficient in which degradation of image quality is less likely to be apparent toward the lower right. Be placed.

  FIG. 2 is a diagram illustrating the order in which the quantized representative values stored in the quantized data table are entropy-encoded and the order in which the entropy-decoded quantized representative values are stored in the quantized data table. As shown in the drawing, the quantized representative values are entropy-encoded or stored in the zigzag order from the upper left to the lower right, and this order is generally called a zigzag order or zigzag sequence. Therefore, for example, when entropy decoding is performed in the order of the zigzag order, the quantized representative value representing the DCT coefficient having a lower spatial frequency is restored earlier. EOB is added to the head of the column of zero quantized representative values.

  Hereinafter, the normal interval width for dividing the DCT coefficient D (m) of the Y component image data when no data is inserted is represented by the first quantization step Q1 (m), and when the data is inserted, the Y component image is displayed. A section width for dividing the DCT coefficient D (m) of the data is represented by a second quantization step Q2 (m) (m is a subscript indicating a zigzag order, and 0 ≦ m ≦ 63). Also, the EOB determined by the quantization in the first quantization step Q1 (m) is called an original EOB, and the order of the zigzag scan of the original EOB quantization representative value is X (0 ≦ X ≦ 63). . Further, the representative quantization value obtained by the quantization of the DCT coefficient D (m) in the first quantization step Q1 (m) is represented by R1 (m). Further, the representative quantization value obtained by the quantization of the DCT coefficient D (m) in the second quantization step Q2 (m) is represented by R2 (m).

  Also, an EOB that is determined by quantization in the second quantization step Q2 (m) and is one frequency higher than the original EOB position is called a new EOB, and a new EOB quantization representative value is obtained. The order of zigzag scanning is W (0 ≦ W ≦ 63).

  The data insertion unit 15 includes each of the quantized representative value R2 (Z) of the Y component image data that starts embedding a bit string of arbitrary data to be inserted to the quantized representative value R2 (Y) that ends the bit string embedding. One bit is embedded in the least significant bit in order from the beginning of the bit string of data (0 ≦ Z ≦ m ≦ Y ≦ X ≦ 63).

  Specifically, the data insertion unit 15 discards the quantization step Q1 (m) and the quantization representative value R1 (m) corresponding to the DCT coefficient D (m), and then corresponds to the DCT coefficient D (m). Set the quantization step to Q2 (m) so that the evenness (parity) of the rounding operation result value of the division result value D (m) / Q2 (m) corresponds to the data bit value (0 or 1) to be embedded Is set to the rounding direction of the division result value D (m) / Q2 (m), the value obtained by rounding the division result value D (m) / Q2 (m) in the set rounding direction is set to the DCT coefficient D (m) By setting the quantized representative value R2 (m), bits are inserted into each quantized representative value R2 (m).

  Although details are omitted, it goes without saying that data can be embedded in the U / V component in the same manner as the Y component.

  In the following, for the sake of simplicity, data insertion for one quantized representative value R2 (m) of a block will be described, but data insertion is similarly performed for each quantized representative value R2 (m) of a plurality of blocks. Can do.

  In the present specification, the Z-th coefficient position at which data insertion is started is assumed to be fixed (that is, Z is a predetermined value), but may be variable according to various parameters. The Y-th coefficient position at which data insertion is completed is preferably the coefficient position immediately before the original EOB (Y = X−1), and changes depending on the original EOB position. This is to prevent the zero run length from being shortened and the encoding efficiency from being lowered by inserting bits after the original EOB. If the original EOB position is on the lower frequency side than the data embedding start position (zigzag scan order Z-th coefficient position), data insertion is not performed.

  In the conventional example, only one bit of data can be inserted per macroblock. However, in this embodiment, one bit of data is assigned to each quantized representative value R2 (m) (0 ≦ Z ≦ m ≦ Y ≦ 63) of one macroblock. Therefore, the amount of data that can be inserted is significantly increased as compared with the conventional example.

  FIG. 3 shows an example of bit insertion by the data insertion unit 15. In this figure, for convenience of explanation, the second quantization step Q2 (m) is 1 (Q1 (m) = Q2 (m) = 1) and is the same as the first quantization step Q1 (m). The result value D (m) / Q2 (m) is assumed to be within the range of -0.5 to 0.5.

  The data insertion unit 15 assumes that the bit value to be inserted is “0” and the division result value D (m) / Q2 (m) is −0.5 <D (m) / Q2 (m) ≦ 0. Is set to + ∞ side, and the quantization representative value is changed from R1 (m) to even R2 (m) = 0 in the nearest neighborhood of the division result value D (m) / Q2 (m). Also, if the bit value to be inserted is “0” and the division result value is 0 ≦ D (m) / Q2 (m) <0.5, the rounding direction is set to −∞, and the quantization representative value is From R1 (m), is changed to an even number R2 (m) = 0 in the nearest neighborhood of the division result value D (m) / Q2 (m). These rounding operations correspond to so-called rounding, and the error that occurs when the quantization representative value is changed from R1 (m) to R2 (m) is half of the first quantization step Q1 (m) (= 0.5). It is less than or equal to the rounding error at the time of quantization in the first quantization step Q1 (m) (rounding error due to rounding of the division result value D (m) / Q1 (m)).

  On the other hand, if the bit value to be inserted is “1” and the division result value D (m) / Q2 (m) is −0.5 <D (m) / Q2 (m) ≦ 0, the rounding direction is set to −∞. And the quantization representative value is changed from R1 (m) to the nearest odd number R2 (m) = − 1 of the division result value D (m) / Q2 (m). Also, if the bit value to be inserted is “1” and the division result value is 0 ≦ D (m) / Q2 (m) <0.5, the rounding direction is set to + ∞, and the quantization representative value is From R1 (m), the odd number R2 (m) = 1 nearest to the division result value D (m) / Q2 (m) is changed. These rounding operations correspond to so-called rounding down and rounding up, and the error that occurs when the quantization representative value is changed from R1 (m) to R2 (m) is the first quantization step Q1 (m) (= 1). It becomes twice the rounding error (rounding error due to rounding of the division result value D (m) / Q1 (m)) at the time of quantization in the first quantization step Q1 (m).

  As described above, if the data bit to be inserted into the quantized representative value R2 (m) is 0, the quantized representative value R1 (m) is an even number R2 () nearest to the division result value D (m) / Q2 (m). If the data bit to be inserted is 1, the quantized representative value R1 (m) is changed to the nearest odd number R2 (m) of the division result value D (m) / Q2 (m). On the decoding side, if the quantized representative value R2 (m) is an even number (that is, the least significant bit is 0), data bit 0 is inserted, and the quantized representative value R2 (m) is an odd number (that is, the least significant bit is If 1), it can be identified that data bit 0 is inserted.

  By the way, when Q1 (m) = Q2 (m), the quantization error due to the change from R1 (m) to R2 (m) to the quantization representative value is due to the first quantization step Q1 (m). It becomes larger than an error (quantization error) that occurs during quantization, and may cause image quality degradation.

  Therefore, in this embodiment, the second quantum is further reduced in order to keep the image quality degradation accompanying the change of the quantization representative value due to the data embedding within the quantization error of the quantization by the first quantization step Q1 (m). The quantization step Q2 (m) is set to 1/2 or less of the first quantization step Q1 (m).

  That is,

The second quantization step Q2 (m) is set in advance so as to satisfy

  Then, the quantization error of the quantization by the second quantization step Q2 (m) is equal to or less than the quantization error of the quantization by the first quantization step Q1 (m), and the data bit Image deterioration due to insertion can be made difficult to occur.

  The best Q2 (m) value determined based on the CSF (Contrast Sensitivity Function) representing human perception characteristics is

It is. A method of obtaining this Q2 (m) will be described in an example described later.

  However, if the quantization step Q2 (m) is a value smaller than Q1 (m) as in Equation 3 or Equation 4, the quantized representative value R2 (m) after the change on the high frequency side is a non-zero value. The probability that a zero run length will occur until a new EOB is reduced or the zero run length is shortened as compared with the quantization in the first quantization step Q1 (m). For this reason, the encoding efficiency is lowered and the JPEG file is lengthened.

  For this reason, in the present embodiment, the second quantization step Q2 (m) is sequentially increased in a zigzag order from the low frequency side to the high frequency side.

  As an example, quantization step Q2 (m)

And the denominator for decreasing Q1 (m) is sequentially changed from ^71/2 to 2. Or

とし、Q1(m)を小さくする分母を７^1/2から１へ順次変化させてもよい。あるいは、ｋ<Ｙとなる定数ｋを予め設定しておき、

And the denominator for decreasing Q1 (m) may be sequentially changed from ^71/2 to 1. Alternatively, a constant k that satisfies k <Y is set in advance,

When m is between Z and k, the denominator for decreasing Q1 (m) is sequentially changed from ^71/2 to 1, and when m is between k + 1 and Y, the denominator for decreasing Q1 (m). May be set to 1 (that is, Q2 (m) is the same value as Q1 (m)).

  Originally, the first quantization step Q1 (m) is set so that the probability that a quantized representative value of 0 is gradually generated increases from the low frequency side to the high frequency side coefficient position. In this connection, a value obtained by multiplying the first quantization step Q1 (m) by the “monotonically increasing function of m” as shown in Equations 5 to 7 is defined as a second quantization step Q2 (m), and Q2 ( Increase the value of m) in zigzag order. In this way, the probability that the quantized representative value R2 (m) on the high frequency side will be a value other than 0 compared to the quantized representative value R1 (m) due to the quantization in the second quantization step Q2 (m). As a result, the zero run length is shortened as data is inserted, and the JPEG file can be prevented from becoming too long.

  As described above, if the quantized representative value R2 (m) is an even number, the least significant bit is 0. From the decoding side, a new EOB quantization representative value “0” must be distinguished from the insertion bit “0” of the highest frequency side coefficient position.

  Therefore, the division result of −1 or 1 is used with the quantization representative value R1 (X) of the division result value D (X) / Q1 (X) of the original EOB DCT coefficient D (X) as the data end marker. Change to the value R2 (X) closest to the value D (X) / Q1 (X). Since the data end marker is at a coefficient position one frequency lower than the new EOB, the decoding side inserts the data end marker by detecting the data end marker when sequentially decoding the DCT coefficient from the low frequency side. It is possible to distinguish between the bit “0” and the zero quantized representative value sequence below EOB (End of Block).

  The original EOB quantized representative value is changed to the end marker, so that the new EOB is moved up to the coefficient position one frequency higher than the original EOB, but the EOB itself as in the prior art. Is not eliminated, but is only shifted by one coefficient, so that the decrease in coding efficiency and the increase in code amount due to the decrease in zero run length are not large.

  It is also conceivable that the original EOB quantization representative value is changed from 0 to -1 or 1 to deteriorate the image quality. Therefore, Q2 (X) <Q1 (X) is changed to the original EOB DCT coefficient D (X) quantization step, and the quantized representative value R1 (X) is a division result value of -1 or 1. If you change to the value R2 (X) that is closest to D (X) / Q2 (X), the error due to changing the original EOB to the end marker will be reduced accordingly, and the effect on image quality will also be reduced. .

This is because the DCT coefficient of the original EOB from the end marker on the decoding side is
Data end marker x Q2 (X)
This is because the smaller the Q2 (X), the closer to the original EOB DCT coefficient = 0.

  In particular, if Q2 (X) ≦ Q1 (X) / 2, the error caused by changing the quantized representative value R1 (X) to the data end marker is quantized by the first quantization step Q1 (X). The same or less than the quantization error of.

1. Basic Algorithm First, FIG. 4 shows an example of a steganographic encoder using an uncompressed image as an image for concealing data. The secret data sequence whitened by the randomizer is embedded in the lower n bits of the requantized value in which the lower n bits become 0 by the quantizer Qn. At the same time, by performing error diffusion with dither and feedback filter F (z), the perception of image quality degradation is prevented as much as possible. The human perceptual characteristic has a characteristic such as CSF (Contrast Sensitivity Function) shown in FIG. 5, for example (see Table 1, documents [1] and [2]), and is perceived as shown in FIG. A difficult noise power spectrum shape can be assumed. In order to make it difficult to perceive image quality degradation, it is desirable that the noise power spectrum and the noise power spectrum caused by the requantization error have a similar shape.

  Next, the outline of the method of concealing data in the JPEG data devised this time will be described. JPEG encoding is performed in the order shown in FIG. Even if data concealment is performed on the pixel value, since DCT conversion and quantization are performed thereafter, the concealed data is destroyed. Therefore, instead of concealing data in pixel values, consider concealing data when quantizing DCT coefficients. In this research, at most one bit of data is concealed in one DCT coefficient. This is because JPEG encoding originally compresses the amount of information using human perceptual characteristics, and concealing data of 2 bits or more in the DCT coefficient increases the quantization noise power as will be described later, resulting in deterioration in image quality. This is because it becomes difficult to prevent perception of the image.

  FIG. 8 shows a case where data concealment is performed on DCT coefficients that are even quantization representative values if normal quantization is performed. In the case of normal quantization, the maximum quantization error is half of the quantization step, but in FIG. 8, the maximum quantization error is achieved. The same applies to the case of DCT coefficients that are odd quantization representative values in normal quantization. The average quantization noise power is also described. If the quantization step of normal quantization is Δ, the probability of occurrence of the DCT coefficient x is uniform p (x) within each quantization boundary, and the quantization representative value of x is q (x), the average noise power is From Max's quantizer,

  When data concealment is performed as shown in FIG. 8, the quantization step is δ, and in the case of normal quantization, the probability that the quantized representative value is an even number, the probability that it is an odd number is 1/2, and the embedded data bits are '0', ' If each occurrence probability of 1 'is 1/2, the average noise power is

  Since the image data has already been transformed into the frequency domain by DCT transformation, the quantization step is determined so that the quantized noise power spectrum quantized while concealing the data in the DCT coefficient is directly in the shape shown in FIG. I need to do it.

  In addition, the influence of the quantization step on the JPEG data size must be considered. In JPEG encoding, DCT coefficients are zigzag scanned and linearly quantized from low to high as shown in FIG. 9, so that the quantized values of DCT coefficients are continuously set to 0 particularly in the high band. The probability of becoming high. If the quantization value continues to be 0, the JPEG data size is reduced by the zero-run length code during Huffman coding. However, if the quantization step is small, the probability that the quantized value becomes 0 decreases, so that it is difficult to expect a decrease in the JPEG data size due to the zero-run length code. Therefore, it is necessary to determine the quantization step and the position of the DCT coefficient that conceals the data in consideration of both the perception of image quality degradation and the fact that the JPEG data size becomes larger than usual.

2. Confidentiality in luminance data Arbitrary data is embedded in luminance data of an image. In order to make it difficult to perceive image deterioration due to embedding, the following measures were taken.
1. Implant so that it is difficult to perceive by examining CSF.
2. The maximum noise power generated by embedding is suppressed to the same level as when no embedding is performed.

Details are omitted.
2.2 Relationship between quantization noise power and image quality deterioration When data is embedded in an image, an error more than the quantization error caused by normal JPEG compression occurs. By making the noise power resulting from the error equal to the image without embedding, the deterioration of the image is less perceived.

  As shown in Equation 9, the quantization step size may be halved in order to make the average noise power of errors due to embedding equal to the normal average quantization noise power. In 2, it was found that image degradation was easily perceived.

  Therefore, we inferred that the cause is the worst noise power due to embedding, and examined a quantization step size that makes the worst noise power equal to the normal average quantization noise power.

Quantization noise power in the JPEG compression is delta ^2/12 than the number 8.

  Average noise power due to error when data is embedded

(Single underline has no embedding error, double underline has embedding error)
Here, considering only the case where an embedding error occurs,

It becomes. The quantization step size for equalizing the maximum noise power with embedding and the quantization noise power with no embedding is:

Thus, it can be seen that the quantization step size at the position where the embedding is performed should be 7 ^-1/2 in the case where the embedding is not performed.

However, if the quantization step size of all the embedding positions is set to 7 ^-1/2 , the compression rate is lowered and the file size is increased. Therefore, by setting the quantization step size to 7 ^{-1/2 in the} frequency band with high human eye sensitivity and 1/2 in the frequency band with low sensitivity, the increase in file size is suppressed while maintaining the image quality.

3. Subjective evaluation of image quality Subjective evaluation of image quality between steganographic images and normal JPEG images was performed. Since data embedding relates to the image quality because the position of the DCT coefficient at which data embedding starts (FIG. 10), evaluation was performed using an image in which the JPEG Quality value and the embedding start position were changed. The image used in this evaluation was created by a method of halving the quantization step size corresponding to the embedding location.

Here, the JPEG Quality value is a numerical value (0 to 100) related to the JPEG compression rate. The larger the value is, the lower the compression rate and the less the deterioration of the image. For IJG encoders, the default Quality value is 75.
■ Image used for evaluation
25 sheets * 16 types (change in embedding start position, quality change (Fig. 11))
■ Evaluator
6 people ■ Evaluation method JPEG images without embedding and JPEG steganography images are compared, and evaluations are made with numbers from 1 to 6.
1I'm very worried about image quality degradation ...
5I don't mind image quality degradation at all
As shown in FIG. 12, as a result of the evaluation experiment, it was found that the deterioration of image quality due to embedding was perceived more than expected. Therefore, the experiment was performed again using a method of setting the quantization step size to 7 ^-1/2 so that the average maximum noise power by embedding becomes equal to the average quantization noise power without embedding.

  This experiment was conducted with one evaluator, 5 images used for evaluation * 19 types (16 types in the previous time + normal JPEG images of each Quality), and the same evaluation criteria. As a result, as shown in FIG. 13, it is understood that it is less perceived than when the quantization step size is halved. Also, if the embedding start position is set to No. 20 or later, the evaluation score becomes 4.5 or more, and it can be determined that even if embedding is performed, degradation of the image is hardly noticed.

  FIG. 14 shows the increase in the amount of confidential data and the JPEG file size when data is embedded after No. 20.

  It was possible to embed up to 15% at maximum for each JPEG image. Also, the increase in JPEG file size with respect to the amount of confidential data was 1.1 to 1.7 times. Except for the case of Quality95, it can be seen that the embedding efficiency is better when the embedding amount is increased.

  FIG. 15 shows an increase in the amount of confidential data and the JPEG file size when data is embedded in a standard image Lena having 512 × 512 pixels. The relationship between the Q-factor and the quality of the present embodiment in Non-Patent Document 2 “Data Embedding Method in JPEG Image That Does Not Need Specification of Embedding Position” is as follows.

Quality82⇔Q-factor20
Quality28⇔Q-factor80
Quality15⇔Q-factor140
Quality9Q-factor200

Second Embodiment
FIG. 16 is a block diagram of an image compression coding apparatus with a data insertion function according to a preferred second embodiment of the present invention. Unlike the first embodiment, this apparatus further includes a decoding unit 21, a requantization unit 22, and an entropy encoding unit 23. The requantization unit 22 includes a data insertion unit 15. The image signal output unit 10, the image data creation unit 11, the DCT processing unit 12, the quantization unit 13, and the entropy encoding unit 14 are the same as those in the first embodiment. The first processing block 101 includes an image signal output unit 10, an image data creation unit 11, a DCT processing unit 12, a quantization unit 13, and an entropy coding unit 14. The second processing block 102 includes a decoding unit 21, a requantization unit 22, and an entropy encoding unit 23. The first processing block 101 and the second processing block do not necessarily have to be incorporated into one device, and may be incorporated into separate devices.

  The decoding unit 21 decodes the DCT coefficient D (m) of the encoded image data output from the entropy encoding unit 14. The data insertion unit 15 uses the DCT coefficient D (m) decoded by the decoding unit 21 from the DCT coefficient D (Z) at a predetermined start position to the low frequency side of the original EOB (non-zero maximum conversion coefficient). The DCT coefficient D (Y) = D (X−1) one before) is divided by a quantization width Q2 (m) smaller than the quantization width Q1 (m) of the encoded image data of the quantization unit 13. In addition, as in the first embodiment, the DCT coefficient D (m) is inserted while inserting the data bit into the least significant bit by rounding the division result value D (m) / Q2 (m) according to the inserted data bit. Quantize If the quantized representative value at the position immediately before the new EOB determined by this quantization is 1 or -1, no change is made. The zero-value DCT coefficient at the position after the new EOB (above the non-zero maximum conversion coefficient) is left as it is and no data bits are embedded.

  The reason why the data insertion unit 15 makes the quantization width Q2 (m) smaller than the quantization width Q1 (m) is to prevent image quality deterioration due to quantization while inserting data as much as possible.

  The requantization unit 22 divides each DCT coefficient other than the DCT coefficient for which the data insertion unit 15 has determined the quantized representative value by the corresponding quantization width Q1 (m) and performs a rounding operation of rounding to quantize Obtain a representative value and quantize again.

  The entropy encoding unit 23 encodes the quantization width quantization table used by the requantization unit 22 and the quantization representative value determined by the requantization unit 22.

  In this way, the second processing block 102 inserts arbitrary data into the image data once compressed and encoded in the first processing block 101 and performs compression encoding again.

<Third Embodiment>
The requantization unit 22 of the image compression coding apparatus with data insertion function of the second embodiment, based on the quantization table Q1 of the quantization unit 13, adjusts the basic quantization table Q2b for inverse quantization ( F which is a Quality Factor) may be determined.

  As a premise, the requantization unit 22 includes a basic quantization step including a basic quantization step Qb2 (m) (0 ≦ m <64) for quantizing a DCT coefficient when data is not inserted as in the second embodiment. Suppose that it has a conversion table Qb2.

  In addition, the requantization unit 22 has a basic quantization table Qb3 including a basic quantization step Qb3 (m) (0 ≦ m <64) for quantizing DCT coefficients when data is inserted. And

  The decoding unit 21 dequantizes the DCT coefficient quantized by the quantization unit 13 based on the quantization table Q1 based on the quantization table Q1, and then requantizes the inversely quantized DCT coefficient and the quantization table Q1. To the conversion unit 22.

The requantization unit 22 determines f based on the quantization table Q1. Then, the requantization unit 22 creates a quantization table Q2 for actually inversely quantizing the DCT coefficient by multiplying Qb2 (m) uniformly by f. That is, the requantization unit 22
Q2 (m) = Qb2 (m) × f (0 ≦ m <64)
To create a quantization table Q2.

  In the present embodiment, the requantization unit 22 creates the quantization table Q3 used for data insertion by uniformly multiplying the determined f by the basic quantization step Qb3 (m).

That is, the requantization unit 22
Q3 (m) = Qb3 (m) × f (0 ≦ m <64)
To create a quantization table Q3.

However, the quantization steps from the DC component position to the DCT coefficient position where the data embedding starts (coefficient position Z-1 on the lower frequency side of Z) are common to the quantization tables Q2 and Q3. That is,
Q2 (m) = Q3 (m) (0 ≦ m ≦ Z-1)
It is.

  As in the first and second embodiments, the data insertion unit 15 performs the division result value that is a value obtained by dividing each DCT coefficient from the decoding unit 21 by the quantization step Q3 (m) (Z ≦ m ≦ Y). By changing the rounding direction according to the value of the insertion data, data is inserted into the least significant 1 bit of each DCT coefficient.

After each DCT coefficient from the decoding unit 21 is divided by the quantization step Q2 (m) to determine the original EOB, data is inserted using the quantization step Q3 (m) (Z ≦ m ≦ Y). May be. The value of the quantization step Q3 (m) (Z ≦ m ≦ Y) is preferably 1/2 or less of the quantization step Q2 (m), and more preferably (7) ^−1/2 or less. .

1 is a block diagram of an image compression coding apparatus according to a first embodiment. Conceptual diagram of data insertion in JPEG data quantization Conceptual illustration of data insertion by changing the least significant 1 bit of the quantized representative value Main block diagram showing an example of a steganography encoder using an uncompressed image Diagram showing an example of CSF Diagram showing an example of noise power spectrum Block diagram of JPEG encoder Conceptual illustration of data insertion by changing the least significant 1 bit of the quantized representative value Conceptual illustration of DCT coefficient quantization zigzag scan The figure which shows the position of DCT coefficient Diagram showing the relationship between Quality and embedding start position Diagram showing the evaluation of steganographic images Figure showing the evaluation of 7 ^-1/2 and 1/2 Diagram showing the relationship between the amount of confidential data and the increase in JPEG file size Diagram showing the relationship between the amount of confidential data and the increase in JPEG file size in Lena (512x512 pixels) Block diagram of an image compression coding apparatus according to the second embodiment

Explanation of symbols

10: Image signal output unit, 11: Image data creation unit, 12: DCT processing unit, 13: Quantization unit, 14: Entropy coding unit, 15: Data insertion unit, 21: Decoding unit, 22: Requantization Part 23: Entropy encoding part

Claims (5)

A DCT coefficient input unit for inputting DCT (discrete cosine transform) coefficients calculated for each of a plurality of blocks obtained by dividing the image data;
Among the DCT coefficients input to the DCT coefficient input unit, the coefficient position on the low frequency side by one coefficient position from the start position, which is a predetermined low frequency side coefficient position at which data insertion is started, from EOB (End of Block) Arbitrary data to be inserted into each of the least significant 1 bits of the quantized representative value, which is a value obtained by rounding a division result value, which is a value obtained by dividing each of the DCT coefficients up to the end position by the corresponding quantization width A data insertion unit that changes the quantization representative value of the DCT coefficient of the EOB to a value closer to the division result value of the DCT coefficient of the EOB, of -1 or 1,
A data insertion device comprising:   The EOB DCT coefficient quantization representative value of the EOB DCT coefficient is a predetermined value equal to or less than ½ of a quantization width corresponding to the EOB DCT coefficient of −1 or 1. The data insertion device according to claim 1, wherein the value is changed to a value closer to a value obtained by dividing. A block that divides image data into a plurality of blocks, a DCT transform unit that calculates a DCT (discrete cosine transform) coefficient for each block, and a division that is a value obtained by dividing the DCT coefficient of each block by a corresponding quantization width A quantization unit that determines the quantization representative value of each DCT coefficient by rounding the result value, a quantization data table that stores the quantization representative value of each DCT coefficient, and a code that encodes the quantization data table An image compression encoding device comprising a conversion unit,
The quantization unit calculates DCT coefficients from a start position that is a predetermined low frequency side coefficient position at which data insertion is started to an end position that is a low frequency side block by one coefficient position from EOB (End of Block). Changing each of the least significant 1 bit of the quantized representative value that is a value obtained by rounding a division result value that is a value obtained by dividing each by a corresponding quantization width into a bit that constitutes arbitrary data to be inserted; An image compression coding apparatus that changes a quantization representative value of the DCT coefficient of the EOB to a value closer to a division result value of the DCT coefficient of the EOB, which is -1 or 1. Inputting DCT (discrete cosine transform) coefficients calculated for each of a plurality of blocks obtained by dividing the image data;
Among the DCT coefficients input to the DCT coefficient input unit, a block on the low frequency side by one coefficient position from the start position, which is a predetermined low frequency side coefficient position at which data insertion is started, from EOB (End of Block). Arbitrary data to be inserted into each of the least significant 1 bits of the quantized representative value, which is a value obtained by rounding a division result value obtained by dividing each DCT coefficient up to a certain end position by the corresponding quantization width, And changing the quantization representative value of the DCT coefficient of the EOB to a value closer to the division result value of the DCT coefficient of the EOB of −1 or 1,
Data insertion method including Performing block formation to divide image data into a plurality of blocks, and calculating DCT (discrete cosine transform) coefficients for each block;
Determining a quantization representative value of each DCT coefficient by rounding a division result value that is a value obtained by dividing the DCT coefficient of each block by a corresponding quantization width;
Quantities corresponding to respective DCT coefficients from a start position which is a predetermined low frequency side coefficient position at which data insertion is started to an end position which is a low frequency side coefficient position by one coefficient position from EOB (End of Block). Each of the least significant 1 bits of the quantized representative value that is a value obtained by rounding the division result value that is a value divided by the conversion width is changed to a bit that constitutes arbitrary data to be inserted, and the DCT coefficient of the EOB Changing the quantization representative value of −1 or 1 to a value closer to the division result value of the DCT coefficient of the EOB.
Creating a quantized data table storing quantized representative values of each DCT coefficient and encoding the quantized data table;
A method for compressing and encoding images.

Images