JP5482177B2 - Encoding device, program, recording medium - Google Patents

Description

  The present invention relates to an encoding device, a program, and a recording medium.

  Generally, an image is composed of (1) a picture and (2) a character / line drawing (including a symbol different from the picture, hereinafter referred to as “non-picture”). In recent years, when encoding this image, it has been proposed to separate (1) a picture and (2) a character / line image and perform encoding on each.

  For example, Patent Literature 1 proposes a method called an MRC (Mixe Straster Content) model. FIG. 1 shows a conceptual diagram of the MRC model. In the example of FIG. 1, the original image α shows (2) character A “Buy now fastest printer” and (1) picture B of the copier. In the following, “Buy now” is referred to as a character A1, and “fastest printer” is referred to as a character A2. And the characters A1. Character A2 is collectively referred to as character A.

  A foreground P (hereinafter referred to as “non-pattern part color image P”), a mask Q (hereinafter referred to as “non-pattern part shape image Q”), and a background R (hereinafter referred to as “picture part image R”). And 3).

  The non-design part color image P is an image showing a color A character image. The non-pattern part shape image Q is an image showing an image of the shape of the character A1. The pattern part image R is an image showing a background. As described above, the MRC model is divided into three layers, and encoding is performed for each layer.

  In Patent Document 2, a mask is divided using an MRC model to improve accessibility to partial images.

  There are roughly two methods for creating the non-pattern part color image P in the MRC model. As a first method, as schematically shown in FIG. 2, a gradation G (see FIG. 1) is formed between the color S (see FIG. 1) of the character A1 and the color T (see FIG. 1) of the character A2. Thus, the non-pattern part color image P is created by blurring. In order to simplify the drawing of FIG. 2, the color S is shown in black and the color T is shown in white. In FIG. 2, pixels other than the characters A1 and A2 are referred to as non-character pixels.

  As shown in FIG. 2, this first method is a method in which the colors U1 to U3 of the non-character pixels between the character A1 and the character A2 are gradually changed (gradation colors). When this first method is used, only a small amount of edges are generated, so that the compression rate of the non-pattern part color image P can be increased. That is, in this first method, the non-pattern part color image P is composed of low-frequency color components.

  FIG. 3 shows a schematic diagram in the case of dividing into three layers by the second method. FIG. 3 differs from FIG. 1 only in the non-pattern part color image P ′. The letters S1 and A2 in the non-pattern part color image P ′ are colored with the colors S and T, respectively. The characters A1 and A2 are given a background color C.

  FIG. 4 schematically shows the second method. For simplification of the drawing of FIG. 4, the color S is shown in black and the color T is shown in white. As shown in FIG. 4, this method is located between the character A1 (character pixel) and the character A2 (character pixel), has the smallest code amount, and is a specific pixel C (for example, black or black) that is a non-character pixel. Gray). A set of specific pixels C is a specific image. Therefore, hereinafter, the set of specific pixels C is also referred to as a specific image C.

  Then, the non-character pixel region is filled with the specific image C to generate the non-pattern part color image P ′. In other words, the non-pattern part color image P ′ is composed of a color image (a set of color pixels) of the non-pattern (characters) of the original image α and a specific image occupying a position other than the color image of the non-pattern. It is. A set of specific pixels C becomes a specific image.

  That is, regardless of the color S of the character A1 and the color T of the A2, the specific image C having a small code amount is used between the characters A1 and A2. Since the specific image C has a small code amount, the compression rate of the non-picture part color image P ′ can be increased by the second method. Further, in the second method, the transition between the character A1 and the character A2 is not smoothly performed, and an edge is generated. Therefore, the non-pattern part color image P ′ includes a high frequency.

  Here, the first method and the second method are compared. When the ratio of characters / line drawings in the original is small, in the first method, a wide transition range E1 to E3 (gradation area) occurs between the characters A1 and A2, and the code of the non-pattern part color image P is generated. The amount tends to increase. On the other hand, in the case of the second method, since the specific image C having a small code amount occupies most of the original, the code amount is small.

Further, when the ratio of the character or line drawing occupying the document is large, in the first method, the narrow transitive range E1~E3 occurs between the characters A1 and character A2, the code amount of the non-picture portion color images P Tend to decrease. On the other hand, in the second method, there is a problem in that the amount of code increases because a specific image with a small code amount has a small ratio of originals and an edge is generated.

  The encoding device of the present invention aims to reduce the amount of codes and perform appropriate encoding even when the second method is used.

In order to achieve the above object, a pattern image in which a pattern of the original image is displayed, a non-picture color image of the original image, and a specific image occupying a position other than the non-picture color image are displayed from the original image. A generating unit that generates a non- image part color image and a mask image on which the shape of the non-pattern is displayed ;
A frequency converter that converts the frequency of the pattern part image and the non-pattern part color image;
A quantization unit that quantizes the frequency coefficient of the picture part image and the frequency coefficient of the non-picture part color image obtained by the frequency conversion unit ;
An encoding unit that encodes the frequency coefficient of the quantized picture part image and the frequency coefficient of the non-picture part color image;
The quantization, except for normalization for frequency gain,
To provide an encoding apparatus, wherein the quantization step number of the frequency coefficients of the high frequency band of the non-picture portion color images, small Kusuru than the quantization step number of the frequency coefficients of the high frequency band of the picture portion image .

  With the encoding apparatus of the present invention, even when the second method is used, it is possible to reduce the amount of codes and perform appropriate encoding.

The figure which shows the case where an original image is divided into three layers by the 1st method. The figure which showed the 1st method typically. The figure which shows the case where an original image is divided into three layers by the 2nd method. The figure which showed the 2nd method typically. An example of the configuration of an MRC code. The structural example of a JPM code | symbol. 1 is a configuration example of an encoding apparatus according to the present embodiment. The figure which showed the function structural example of CPU of a present Example. The figure which showed the flow of the main processes of the encoding apparatus of a present Example. The figure which shows the case where an original image is divided into five layers. The figure which shows an example of a Sobel operator (the 1). The figure which shows an example of a Sobel operator (the 2). The figure which showed the original image and coordinate system after DC level shift. An array of coefficients after filtering in the vertical direction. An array of coefficients after horizontal filtering. A sorted array of coefficients. The figure which showed the arrangement | sequence of the rearranged coefficient after the 2nd conversion. The figure which showed the relationship between a decomposition level and a resolution level. The figure which showed the image, tile, subband, etc. The figure which showed Visual Weight of Visual distance1000 [dot]. The figure which showed Visual Weight of Visual distance3000 [dot]. The figure which showed the square root of the frequency conversion gain of 9 * 7 wavelet transform. The figure which shows the square root of the gain of color conversion. The figure which shows the quantization step number for normalization. The figure which shows the substantial quantization for the foreground P. FIG. 6 shows substantial quantization for background R. The figure which showed the substantial quantization step number of the substantial quantization step number / brightness | luminance Y of color difference Cb. The figure which showed the substantial quantization step number of the substantial quantization step number / brightness | luminance Y of color difference Cr. The figure which showed the number of substantial quantization steps for foreground P '. The figure which showed the substantial number of quantization steps of the brightness | luminance of the foreground P 'of the modification 1, and the background Q. The figure which showed the substantial quantization step number of the substantial quantization step number / brightness | luminance Y of the color difference Cb of the modification 1. FIG. The figure which showed the substantial quantization step number of the substantial quantization step number / brightness | luminance Y of the color difference Cr of the modification 1. FIG. The figure which showed the number of sub bit planes which the foreground P 'of the modification 2 and the common background Q do not encode. The figure which showed the square root of the frequency conversion gain of 5 * 3 wavelet transform. The figure which showed the gain of color conversion. The figure which showed the divisor for the normalization in the modification 3. FIG. The figure which showed the number of bit planes which are not encoded in the foreground P ′ in Modification 3. The figure which showed the bit plane number which is not encoded in the background Q in the modification 3. The figure which showed the layout object of the JPM file in a present Example (the 1). The figure which showed the layout object of the JPM file in a present Example (the 2). The figure which showed the hardware block diagram which shows the main structures of the status notification apparatus of a present Example.

  Prior to the description of the “DETAILED DESCRIPTION”, the background of this embodiment will be described. In recent years, with the development of networks such as the Internet and public lines and the advancement of image compression technology, the transmission speed of image data has been dramatically improved. Along with this, higher resolution, higher gradation, and colorization of image data that can be transmitted are progressing, and 400 dpi (dot per inch) full-color images are becoming practical transmission targets. By the way, when transmitting such a full-color image through a network or a public line, the image is usually compressed. As a compression algorithm at that time, the JPEG algorithm is generally used. Also, image filing and other image accumulation is actively performed, and data compressed using a JPEG algorithm or the like is similarly used for color image accumulation.

  Here, the JPEG algorithm will be described. First, the compression algorithm will be described. The input image data is subjected to color space conversion, for example, from RGB color space to L * a * b * color space, if necessary, and then divided into 8 × 8 pixel blocks. It is converted to the frequency domain by DCT (Discrete CosineTransform), quantized with a quantization table having the same size as the block, then converted from block data to raster data by zigzag scanning, and becomes compressed data by two-dimensional Huffman coding. . The decompression algorithm is an inverse algorithm of the compression algorithm, and image data is restored by Huffman decoding, raster-block transformation, inverse quantization, and IDCT (Inverse DCT).

  However, as can be seen from the above description, since the JPEG algorithm performs quantization, which is an irreversible process, the decompressed image becomes an image with a degraded image quality compared to the image data before compression. Usually, a quantization table is used in which the quantization step becomes coarser as the high frequency component increases in order to increase the compression ratio. For this reason, particularly in a portion where there is an edge containing a lot of high-frequency components such as characters and line drawings, the image quality is significantly deteriorated and the compression rate is also reduced.

  As described with reference to FIG. 1, in the MRC method, a plurality of “foreground P (non-pattern part color image P) + mask Q (non-pattern part shape image Q)” are superimposed on one background. It is common to go. The mask Q has shape information for cutting out the foreground P (for example, in the form of characters). In some cases, the value of the mask Q has the transmittance of the foreground P (the mixing ratio of the foreground and the background Q (the pattern portion image Q)). Further, the resolution of the foreground P, the mask Q, and the background R can be arbitrarily selected.

  FIG. 5 is an explanatory diagram of a code configuration based on the MRC model. The code format using MRC is an overall header indicating that it is an MRC code, etc., one background code and its header, and one or more “pairs of foreground codes and mask codes” and their pairs. It is typically composed of headers for

For example, in this embodiment, JPEG is used as the background, foreground, and mask encoding method.
JPM (JPEG2000 Multi Layer) which makes 2000 selectable is used. Of course, in JPM, JPEG, MMR, JBIG, or the like can be selected as an encoding method other than JPEG2000, and the format of the JPM code also has a code configuration as shown in FIG.

  For example, an encoding method called JPM that enables selection of JPEG2000 as a foreground, mask, and background compression method of the MRC model has been standardized.

FIG. 6 is an explanatory diagram of a configuration example of a JPM code. Since the dotted line portion is an option, a simple description will be given focusing on the solid line portion. In FIG. 6, “JPEG2000 Signature box” is an entire header indicating that the code belongs to the JPEG2000 family. "Fi
“Le Type box” is an entire header indicating that the code is in JPM format. The “Compound Image Header box” is an entire header including general information of the code. “Page Collection box” is a table of contents indicating the order of each page when the code consists of multiple pages. “Page box” is an overall header indicating the resolution of the page. Here, the page is a canvas for sequentially superimposing (combining) images, and has the same size as the image after the composition is finished. In the case of JPM, “layout object (layout object)” composed of a pair of foreground and mask is sequentially drawn on the page. “Layout Object box” is a header for the foreground and the mask indicating the size and position of the foreground and the mask. The “Media Data box” and “Contiguous Codestream box” are portions including foreground and mask codes. In JPM, the background (BasePage) is handled as an initial page before the layout object is drawn.

In JPM, as described above, as a method of combining the foreground and the background in the process of combining the images by sequentially overlaying the layout objects defined as the foreground and mask pairs on the background,
(I) Method for selecting either foreground or background (ii) A method for taking a weighted average of foreground values and background values can be selected. In the case of (i), the mask is binary, and when the value is 1, the foreground is selected, and when the value is 0, the background is selected. In the case of (ii), the mask has a positive value of 8 bits and is synthesized by the weighted average of the foreground and the background according to the following equation (1).

Composite image = (mask value / 255) × foreground + {(255−mask value) / 255} × background (1)
Which of these synthesis methods is used can be specified for each pair of foreground and mask, and is described in the header for each pair.

In the encoding with the above MRC model, the foreground and the background are compressed at a high compression rate (for example, 1/40) in order to increase the compression rate. Under such a high compression rate, the skill of quantization greatly affects the image quality, so the quantization applied when compressing the foreground and background needs to be optimized.
Here, as a method of creating the MRC code, a method of reducing the foreground image and expressing the foreground in several colors (in the extreme case, one color) is typical. If such color reduction is used, the foreground image becomes an artificial image, and a high compression rate can be obtained by using an encoding method for the artificial image.

  However, the method of performing color reduction requires processing time for color reduction and causes image quality deterioration for color reduction. In order to prevent deterioration in image quality due to this color reduction, it is necessary to narrow the area represented by one color. For this purpose, it is necessary to divide the foreground image into many partial images, and in any case, processing time is required. (In the first place, since the MRC model requires a lot of processing time to generate the foreground / background, there is a circumstance that it is desired to reduce the time required for encoding as much as possible). In addition, it is difficult to completely separate the foreground and the background, and when a portion that should originally be the background is separated as the foreground, if the color is reduced, it is recognized as an image defect.

  Therefore, if an MRC code is to be created at high speed, there is a need for a method that can encode the foreground image at a high speed and with a high compression rate without reducing or dividing the foreground image much.

  Thus, when omitting the color reduction of the foreground image, it is necessary to compress the foreground image in a state where noise is superimposed by scanning, and thus a high compression rate cannot be obtained with the coding method for artificial images. Therefore, in order to obtain a high compression ratio by omitting color reduction, a method using frequency conversion (encoding method for natural images) is applied to both the foreground and the background.

  As long as color reduction is not performed (= high compression but we want to maintain image quality and processing speed), we would like to optimize the foreground image quantization, but the reason for not performing color reduction is speeding up. The processing time cannot be spent on optimization. For example, there is known a method of analyzing feature amounts and contents of an image and selecting an encoding method or a quantization method according to the analysis. However, such analysis generally has a large processing amount and is difficult to adopt.

  Therefore, the applicant paid attention to the structure and function of the MRC model when encoding the foreground image without color reduction at high speed and optimally. In other words, when the encoding using frequency transformation and quantization is adopted for the foreground, considering the function of the layer itself “foreground is color”, the difference between the functions is not analyzed without analyzing the contents of the image. Can be reflected in quantization. However, so far (as far as the inventors know), foreground has not been quantized to reflect these functions.

  In addition, the same number is attached | subjected to the process which performs the same process in the structure part and flowchart figure which have the same function in a block diagram, and duplication description is abbreviate | omitted.

Further, in the following description, the non-picture refers to something that is not a picture, for example, a character, a diagram, and a symbol.
[Overall flow]
First, the overall flow will be described. FIG. 7 is a block diagram showing a hardware configuration of the PC according to the present embodiment. A PC 100 according to this embodiment is a general personal computer in which a CPU (Central Processing Unit) 101, a RAM (Random Access Memory) 102, an HDD (Hard Disk Drive) 103, and the like are interconnected by a system bus 104.

  The CPU 101 is a microprocessor for controlling the entire PC 100. Since the RAM 102 has the property of storing various data in a rewritable manner, it functions as a work area for the CPU 101 and functions as a buffer. The HDD 103 is an information recording medium that stores images and various programs executed by the CPU 101.

  Information recording media include not only the HDD 103 but also various optical disks such as CD (Compact Disk) -ROM and DVD (Digital Versatile Disk), various magnetic disks such as various magneto-optical disks and flexible disks, semiconductor memories, and the like. Various types of media can be used. Alternatively, the program may be downloaded from a network such as the Internet via an external I / F device (not shown) and installed in the HDD 103. In this case, the storage device that stores the program in the server on the transmission side is also the information recording medium in the present invention. Note that the program may operate on a predetermined OS (Operating System), in which case the OS may perform execution of some of various processes described later, It may be included as a part of a group of program files constituting the application software or OS.

  The CPU 101 that controls the operation of the entire system executes various processes based on programs stored in the HDD 103 used as the main storage device of the system.

Next, of the functions that the program stored in the HDD 103 of the PC 100 causes the CPU 101 to execute, the characteristic functions provided in the PC 100 of the present embodiment will be described. Next, an outline of characteristic functions of the PC 100 will be described with reference to FIG. (1) An original image to be processed is read from the HDD 103 into the RAM 102 according to a command from the CPU 101. (2) The CPU 101 reads the original image on the RAM 102 and performs encoding using the MRC model. (3) The CPU 101 writes the encoded data in another area on the RAM 102. (4) The encoded data is stored in the HDD 103 according to a command from the CPU 101.
[Embodiment 1]
Next, a functional configuration example in the CPU 101 of the first embodiment is shown in FIG. FIG. 9 shows the flow of processing of the CPU 101.

In the example of FIG. 8, the CPU 101 of the first embodiment includes a generation unit 2, a DC level shift unit 4, a frequency conversion unit 8, a quantization unit 10, a bit plane encoding unit 12, and a packet generation unit 14. Further, in the case of decrypt the original image by an arrow order in FIG. 8, the processing is performed when decoding encoded original image, the arrow direction opposite (i.e., packets It is performed in order from the generation unit 14). In this embodiment, an example in which JPEG2000 is used will be described.

  First, an original image α to be encoded is input to an input unit (not shown) (step S2). The input original image α is input to the generation unit 2. An image generated from the original image α by the generation unit 2 is shown in FIG. The generation unit 2 generates, from the original image α, a pattern portion image R (also referred to as “background R”) on which the pattern of the original image is displayed, and a non-pattern portion color image representing the color of the non-pattern of the original image α. P ′ (also referred to as “foreground P ′”) and a non-picture part shape image (also referred to as “mask Q”) are generated. Further, the generation unit 2 generates a second mask V and a second background W.

  Next, details of the processing of the generation unit 2 will be described. The generation unit 2 determines whether or not the original image input in step S2 is a pixel (hereinafter referred to as “character pixel”) constituting a character / line drawing in units of pixels. The character pixel indicates, for example, the characters A1 and A2 in FIGS.

The discrimination is performed by a known image area discrimination technique. For example, a Sobel filter known as an edge detection operator is applied to each pixel of the original image. The Sobel filter multiplies the first weight matrix shown in FIG. 11 by multiplying the top, bottom, left, and right nine pixels centered on the pixel of interest to calculate the sum HS. Similarly, the second filter shown in FIG. Multiply the matrix and calculate its sum VS. Then, the square root of (HS ² + VS ² ) is used as the output value of the filter.

  For example, when the filter output value is 30 or more, it is determined that the target pixel is a character pixel, the value of the target pixel position is set to 1, and the values of the other pixels are set to 0. The pattern part shape image Q) is generated (in the case of a binary image, black = 1 and white = 0 are common) (step S4).

  Also, the color of pixels that do not belong to the character / line drawing area (hereinafter, non-character pixels) is replaced with a fixed value (128, 128, 128) to generate a foreground P (non-picture part color image P) (step S6). ). This is because the second method is used as the foreground P generation method in the present embodiment as described above.

  Here, as described above, the non-pattern part color image P is composed of the non-pattern color image of the original image (the locations of the colors S and T in FIG. 10) and the specific image C that occupies a position other than the non-pattern color image. Composed.

  Next, the color of the pixel in the character / line drawing area in the original image α is replaced with the color of the non-character pixel located closest to the pixel to generate a picture portion image R (background R) (step S8).

  Subsequently, a second mask V is created as an image having a pixel value 1 having the same size as the original image (step S10). A second background image W is created as a white image having the same size as the original image (step S12).

  Note that the flow of processing from step S4 to step S12 is not limited to these.

Among these, the non-picture part color image P ′ and the picture part image R are input to the DC level shift unit 4.
<Processing of DC Level Shift Unit 4>
Next, the processing of the DC level shift unit 4 will be described. In the present embodiment, “embodiment 1 in which color conversion is not performed on the image signal by the DC level shift unit 4” and “embodiment 2 in which color conversion is performed on the image signal by the DC level shift unit 4”. Will be explained. In the paragraph <Processing of DC level shift unit 4>, color conversion processing is also described. The luminance signal after color conversion is Y, and the color difference signals are Cb and Cr.

  First, the image is divided into rectangular tiles (number of divisions ≧ 1). Next, for example, when a color image composed of three RGB components is encoded, the tile is subjected to DC level shift and color conversion to luminance / color difference components. More precisely, DC level shift is performed before color conversion. DC level shift is a level shift process that subtracts ½ of the dynamic range of a signal from each signal value using a predetermined conversion formula when the input image signal is a positive number such as an RGB signal value. In the color conversion, the RGB image is converted into the luminance signal (image) Y and the color difference signals Cr and Cb images, whereby the processing for increasing the compression efficiency of the color image can be performed.

In the DC level shift, when the image signal is a positive number (unsigned integer) such as an RGB signal value, as shown in Expression (1), half of the dynamic range of the signal is obtained from each signal value by forward conversion. In the inverse conversion, the level shift for subtracting is performed by adding half of the dynamic range of the signal to each signal value. However, the level shift is not applied when the image signal is a signed integer such as Cb or Cr in the YCbCr signal.

Equation (1) used for forward and reverse conversion of JPEG2000 DC level shift is as follows.

I (x, y) ← I (x, y) -2 ^{Ssiz (i)} ... forward conversion
(1)
I (x, y) ← I (x, y) +2 ^{Ssiz (i)} ... Inverse transformation Here, Ssiz (i) is an image component i (i = 0, 1, 2 for RGB image). This is the bit depth minus one. I (x, y) is a signal value (pixel value) at coordinates (x, y).

Color conversion is performed in order to efficiently compress a color image. The purpose of this is the same as that in JPEG, if the RGB image is converted into the luminance image Y and the color difference images Cr and Cb and then compressed, the compression rate is improved and as a result the reproduction image quality is improved. Two methods of reversible and irreversible are defined for color conversion.

The reversible transformation is called RCT (Reversible multiple component transformation) and is characterized in that the coefficient of the transformation formula is an integer value. This conversion formula is shown in Formula (2).

Y0 (x, y) = floor ((I0 (x, y) + 2 × (I1 (x, y) + I2 (x, y)) / 4)
Y1 (x, y) = I2 (x, y) -I1 (x, y) Forward conversion Y2 (x, y) = I0 (x, y) -I1 (x, y)
(2)
I1 (x, y) = Y0 (x, y) -floor ((Y2 (x, y) + Y1 (x, y)) / 4)
I0 (x, y) = Y2 (x, y) + I1 {x, y) Inverse transformation I2 (x, y) = Y1 (x, y) + I1 [x, y)
Note that I indicates an original signal, and Y indicates a signal after conversion. In addition, 0 to 2 following I and Y are suffixes. Specifically, when reversibly converting an RGB signal, the I signal is represented as I0 = R, I1 = G, I2 = B, and the Y signal is represented as Y0 = Y, Y1 = Cb, Y2 = Cr. The Further, floor (x) represents a floor function of x (a function that replaces the real number x with an integer that does not exceed x and is closest to x).

The irreversible transformation is called ICT (Irreversible multiple component transformation), and is characterized in that the coefficient of the transformation formula is a real value unlike the RCT. This conversion formula is shown in Formula (3).

Y0 (x, y) = 0.299 × I0 (x, y) + 0.587 × I1 (x, y) + 0.144 × I2 (x, y)
Y1 (x, y) = − 0.16875 × I0 (x, y) −0.33126 × 1 (x, y) + 0.5 * I2 (x, y) Forward conversion Y2 (x, y) = 0. 5 × I0 (x, y) −0.41869 × I1 (x, y) −0.0811 × I2 (x, y)
(3)
I0 (x, y) = Y0 (x, y) + 1.402 × Y2 (x, y)
I1 (x, y) = Y0 (x, y) −0.34413 × Y1 (x, y) −0.71414 × Y2 (x, y) Inverse transformation I2 (x, y) = Y0 (x, y) + 1.772 × Y1 (x, y)
Note that I indicates an original signal, and Y indicates a signal after conversion. In addition, 0 to 2 following I and Y are suffixes. Specifically, when irreversibly converting an RGB signal, the I signal is expressed as I0 = Y, I1 = G, I2 = B, and the Y signal is expressed as I0 = Y, I1 = Cb, I2 = Cr. expressed.
<About the frequency converter 8>
Next, processing of the frequency conversion unit 8 will be described. The frequency conversion unit 8 performs frequency conversion on the pattern portion image R, the non-pattern portion color image P, and the specific image C. After the frequency conversion, the frequency component f _R of the pattern part image R and the frequency component f P of the non-pattern part color image _P are output. Details of the processing of the frequency converter 8 will be described.

  For each tile, the component after color conversion (referred to as a tile component) is frequency-converted by the frequency converter 8. In this embodiment, an example in which wavelet transformation is used as a frequency transformation technique will be described. By wavelet transform, it is divided into four subbands abbreviated as LL, HL, LH, and HH. When the wavelet transform (hereinafter also referred to as “decomposition”) is repeated recursively on the LL subband, one LL subband and a plurality of HL, LH, and HH subbands are finally generated. The Here, the two-dimensional wavelet transform adopted in JPEG2000 will be described.

  In the two-dimensional wavelet transform to be performed after the color conversion, a 5 * 3 wavelet transform for applying a 5 * 3 filter and a 9 * 7 wavelet transform for applying a 9 * 7 filter can be selected. The 5 * 3 wavelet transform is a transform that obtains one low-pass filter output (low-pass coefficient) using five pixels and obtains one high-pass filter output (high-pass coefficient) using three pixels.

  Similarly, the 9 * 7 wavelet transform is a conversion in which 9 pixels are used to obtain one low-pass filter output (low-pass coefficient), and 7 pixels are used to obtain one high-pass filter output (high-pass coefficient). is there. The main difference is the difference in the filter range. The same applies to the low pass filter at the center of the even position and the high pass filter at the center of the odd position. The same applies to 9 * 7 filters.

  Equation (4) shows the 5 * 3 wavelet transform.

Forward conversion C (2i + 1) = P (2i + 1) -floor ((P (2i) + P (2i + 2)) / 2) [step1]
C (2i) = P (2i) + floor ((C (2i-1) + C (2i + 1) +2) / 4) [step2]
Inverse transformation (4)
P (2i) = C (2i) -floor ((C (2i-1) + C (2i + 1) +2) / 4) [step3]
P (2i + 1) = C (2i + 1) + floor ((P (2i) + P (2i + 2)) / 2) [step4]
The conversion formula of the 9 * 7 wavelet transform is shown in Formula (5).

Forward conversion C (2n + 1) = P (2n + 1) + α × (P (2n) + P (2n + 2)) [step 1]
C (2n) = P (2n) + β × (C (2n−1) + C (2n + 1)) [step 2]
C (2n + 1) = C (2n + 1) + γ × (C (2n) + C (2n + 2)) [step 3]
C (2n) = C (2n) + δ × (C (2n−1) + C (2n + 1)) [step 4]
C (2n + 1) = K × C (2n + 1) [step5]
C (2n) = (1 / K) × C (2n) [step6]
Inverse transformation (5)
P (2n) = K × C (2n) [step1]
P (2n + 1) = (1 / K) × C (2n + 1) [step2]
P (2n) = X (2n) −δ × (P (2n−1) + P (2n + 1)) [step 3]
P (2n + 1) = P (2n + 1) −γ × (P (2n) + P (2n + 2)) [step 4]
P (2n) = P (2n) −β × (P (2n−1) + P (2n + 2)) [step 5]
P (2n) = P (2n + 1) −α × (P (2n) + P (2n + 2)) [step 6]
However, α = -1.586134342059924
β = -0.052980118572961
γ = 0.882911075530934
δ = 0.443506852043971
K = 1.230174104914001
Next, the procedure of wavelet transform and the definition of the decomposition level, resolution level, and subband will be described.

  FIG. 13 shows an original image and a coordinate system after DC level shift in the process of performing wavelet transform (5 × 3 transform) in two dimensions (vertical direction and horizontal direction) on the luminance component of a 16 × 16 image. This is an example of mirroring. FIG. 14 exemplifies the arrangement of coefficients after filtering in the vertical direction in the same process. FIG. 15 shows an example of the arrangement of coefficients after filtering in the horizontal direction in the same process. FIG. 16 shows an example of the rearranged coefficient array in the same process.

  Referring to FIG. 13, xy coordinates are taken, and for a certain x, the pixel value of a pixel whose y coordinate is y is expressed as P (y) (0 ≦ y ≦ 15). In JPEG2000, first, a high-pass filter is applied to a pixel whose y coordinate is odd (y = 2i + 1) in the vertical direction (Y coordinate direction) to obtain a coefficient C (2i + 1), and then the y coordinate is even (y = 2i). ) To obtain a coefficient C (2i) by applying a low-pass filter centering on the pixel of (), and such processing is performed for all x.

  Here, the high-pass filter and the low-pass filter are expressed by the relational expressions of Step 1 and Step 2 described above in order. When a filter is applied to the edge of an image, there may be insufficient adjacent pixels with respect to the central pixel. In such a case, the pixel value is appropriately compensated by a technique called mirroring (see FIG. 13). It will be. Mirroring assumes a virtual pixel outside the edge of the image, and the pixel value inside the edge is copied symmetrically with respect to the pixel outside the edge, with the pixel at the edge of the image as the axis of symmetry. This is a well-known technique. Mirroring is performed at all four edges of the image (tile).

For simplicity of explanation, if the coefficient obtained by the high-pass filter is denoted by H and the coefficient obtained by the low-pass filter is denoted by L, the image shown in FIG. 13 is converted to the L coefficient as shown in FIG. , Converted into an array of H coefficients (frequency coefficients).

  Next, with respect to the coefficient array as shown in FIG. 13, first, a high-pass filter is applied around the coefficient whose x coordinate is an odd number (y = 2i + 1) in the horizontal direction, and then the x coordinate is an even number (x = 2i). A low-pass filter is applied centering on the coefficient, and these processes are performed for all y. In this case, P (2i) and the like in the first step and the second step are read as representing coefficient values.

  Next, for simplicity of explanation, the coefficient obtained by applying the low-pass filter centered on the L coefficient described above is LL, the coefficient obtained by applying the high-pass filter centered on the L coefficient is HL, and the coefficient obtained by applying the high-pass filter is centered on the low-pass filter. If the coefficient obtained by applying the filter is denoted by LH, and the coefficient obtained by applying the high-pass filter centering on the H coefficient is denoted by HH, the coefficient array shown in FIG. 13 becomes a coefficient array as shown in FIG. Converted. The coefficient group given the same symbol here is called a subband, and is composed of four subbands in the form shown in FIG.

As described above, each wavelet transform (one decomposition (decomposition)) is completed for each of the vertical and horizontal directions, and only the above-described LL coefficients are collected for each subband as shown in FIG. If only the LL subband is taken out, the “image” with exactly half the resolution of the original image
Is obtained. In this connection, the classification for each subband is called deinterleaving, and the arrangement in the processing state as described with reference to FIG. 1 is called interleaving. The second wavelet transform uses LL subbands. It can be regarded as an original image and the same conversion as described above may be performed. In this case, when rearrangement is performed, the form shown in FIG. 17 is obtained. The coefficient prefixes 1 and 2 here indicate how many times the wavelet transform has been obtained with respect to horizontal and vertical, and are called decomposition levels.

  FIG. 18 illustrates the relationship between the decomposition level and the resolution level. Here, it is shown that the resolution level is almost opposite to the composition level.

In the above description, when only one-dimensional wavelet transformation is performed, processing in only one direction may be performed, and the number of times wavelet transformation has been performed in any one direction becomes the decomposition level.

When a 9 * 7 filter is selected as the wavelet filter, linear quantization of coefficients is performed for each subband after the above-described division, but when a 5 * 3 filter is selected,
The specification does not perform linear quantization for each subband.

As the next processing, each subband is divided into rectangles called precincts. The precinct is a collection of three subbands HL, LH, and HH obtained by dividing a subband into rectangles. In a broad sense, the precinct represents a position in an image. There are three precincts, which are one group, but 1 when the LL subband is divided.
One unit becomes one. The precinct can be the same size as the subband, and the code block is obtained by further dividing the precinct into rectangles.

FIG. 19 illustrates the relationship among images (tiles), subbands, precincts, and code blocks. Here, as an order of physical size, it is shown that image ≧ tile> subband ≧ precinct ≧ code block.
[About the quantization unit 10]
With respect to the quantization of the quantization unit 10, in the present embodiment, quantization for normalization for frequency gain is excluded. Prior to the description of the quantizing unit 10, the reason for “excluding quantization for normalization for frequency gain” will be described.

Quantization in transform coding is divided into two types, that is, quantization for normalizing frequency conversion gain and quantization for reflecting human visual characteristics and image characteristics, and is usually performed as follows.
(I) When the frequency transformation by the frequency transformation unit 8 is an orthogonal transformation When the frequency transformation is an orthogonal transformation, an error occurring in the frequency coefficient of each band is a pixel value calculated by frequency inverse transformation according to the definition. Contributes at a rate equal to. For this reason, quantization (coefficient quantization or truncation or both) is performed in consideration of human visual characteristics, etc. without considering the contribution (no need for quantization for frequency conversion gain normalization). Do.
(II) When frequency transformation is non-orthogonal transformation (ex. Bi-orthogonal transformation) (II-i) Degree of contribution of error generated in frequency coefficient of each band to pixel value calculated by frequency inverse transformation (so-called Since the frequency conversion gain is different for each band, normalization is first performed in consideration of the contribution for each band. The normalization is typically performed by linear quantization of the coefficient for each band (ex. The coefficient of the band with a small contribution degree is quantized in advance to give an error). The intention is normalization.

  (II-ii) Next, quantization (quantization of coefficients or truncation or both) is performed in consideration of human visual characteristics and the like.

  However, in the encoding in which the coefficient cannot be quantized and only truncation can be performed, the normalization in (II-i) is reflected in the form including the truncation amount in (II-ii). As described above, “encoding omission or discard of n bit planes in (ii) is equivalent to dividing the coefficient by 2 to the nth power in (i), and the quantization in (iii) is It is equivalent to dividing the coefficient by S1 × (2 to the power of n) ”.

  Further, even in a method capable of normalization in (II-i), in order to avoid performing linear quantization twice, the quantization in (II-i) and the quantization in (II-ii) are combined together. This is a typical example.

  In color image encoding, color conversion is performed simultaneously with frequency conversion. A typical example of color conversion used in the present application is to convert RGB signals into luminance / color difference signals. As in the case of frequency conversion, when color conversion is not orthogonal conversion, the gain of the color conversion is reflected. Thus, normalization of errors for each luminance / color difference is performed at the same time.

  As described above, the quantization in the present embodiment excludes “quantization for normalization for frequency gain” or “quantization for normalization of luminance / chrominance conversion gain”. In other words, it refers to visual quantization excluding normalization.

  In addition, when transform coding is a method using subband transform, it is possible to define the visual importance of coefficients for each subband. For example, in the case of JPEG2000 that uses wavelet subband transformation, there is an amount named Visual Weight as an index representing the amount of visual quantization in the implementation of the present application. The Visual Weight indicates the visual importance for each subband (band), and is an index indicating that the importance is larger as the value is larger.

  Since Visual Weight is a value obtained by integrating human visual characteristics (visual sensitivity, experimental value) in the band section of the subband (in short, it quantifies the human visual sensitivity in the frequency band). It is not an index specific to the wavelet transform, but a value definable for any subband transform. However, if you want to define Visual Weight for both luminance and color difference, it is necessary to specify the luminance / color difference space, determine the visual sensitivity for the color difference component based on the experimental value in the specified color difference space, and integrate it. There is.

  20A to 20C show visual weight decomposition levels such as luminance signal Y, chrominance signal Cb, and HL of Cr visual distance 1000 [dot] (see FIG. 15 and the like). Further, FIGS. 21A to 21C show visual weight decomposition levels such as HL of the luminance signal Y, the color difference signal Cb, and the visual distance 3000 [dot] of Cr.

  The subband band of the wavelet transform is calculated from the band of the image itself. However, the closer the visual distance to the image is, the more high frequency errors are noticeable, and the far distance the visual distance is, the lower frequency errors are only noticeable, so the subband bandwidth for the observer's eyes is It is a function of the observer's visual distance. For this reason, Visual Weight is determined for each Visual distance. The higher the distance, the lower the importance of the high frequency than the low frequency, and the less important the color difference component than the luminance component (see, for example, FIGS. 20 and 21).

In JPEG2000, when the coefficient is linearly quantized to reflect this Visual Weight, the denominator (quantization step number) of the quantization processing by the quantization unit 10 is expressed as follows: quantization step number = (for normalization) Value for each subband used for division / Visual Weight of the subband) × n (n is a constant selected by the user according to the compression ratio, etc.) (6)
Is generally. Therefore, the smaller the Visual Weight, that is, the sub-band with a smaller visual importance, the larger the quantization.

  Then, when calculating the number of quantization steps according to Equation (6), the degree of quantization in the present embodiment “excluding quantization for frequency conversion gain normalization and color conversion gain normalization” Indicates the number of quantization steps, the number of sub-bit planes, etc.

(1 / Visual Weight of the subband) × n (7)
It becomes. Hereinafter, the value calculated by Expression (7) is referred to as “substantial quantization step number”. When n is common between the foreground P ′ and the background Q, the “substantial number of quantization steps” is the reciprocal of Visual Weight.

  The description of the processing of the frequency conversion unit 8 will be returned. Next, the frequency conversion unit 8 performs frequency conversion (wavelet conversion) on the background R. FIG. 22 shows an example of the square root of the frequency gain of 9 × 7 wavelet transform.

  The sum of squares of errors generated in each frequency conversion coefficient is multiplied by the gain shown in FIG. 22 at the time of frequency reverse conversion, and further multiplied by the gain shown in FIG. 23 at the time of reverse color conversion. Therefore, as known, if the coefficient for each frequency is divided by the reciprocal of the square root of “frequency conversion gain × color conversion gain” (shown in FIG. 24), the influence of the quantization error generated in each coefficient is It can be made uniform between frequencies and between components (each luminance signal Y, color difference signals Cb, Cr).

  As described above, in the present embodiment, as described in [Reason for excluding quantization for frequency conversion gain normalization in the quantization unit 10], quantization for normalization as shown in FIG. In this embodiment, the degree of quantization is defined by the above equation (7) (substantial number of quantization steps).

  25A, 25B, and 25C are examples of the substantial number of quantization steps for the foreground P ′ for the Y signal, the Cb signal, and the Cr signal, respectively. All nunts have a value of 64. In Equation (7), Visual Weight is set to 1 for each frequency / component and n = 64. In FIGS. 25A to 25C, the number of quantization steps of the foreground P ′ is the same from high frequency to low frequency.

  On the other hand, (A), (B), and (C) in FIG. 26 are examples of the substantial number of quantization steps for the background R in this embodiment. This is obtained by applying Visual Weight 1000 of Visual distance 1000 to each frequency / component in equation (7) and setting n = 64. 26A, 26B, and 26C, the substantial number of quantization steps is larger at the high frequency HH than at the low frequency.

Here, in the first embodiment, as shown in FIG. 25 (quantization step of the foreground P ′) and FIG. 26 (quantization step of the background R ), the high-frequency component HH of the non-pattern part color image (foreground P ′). Is made weaker (smaller) than the degree of quantization of the high-frequency component HH of the pattern portion image (background R 1 ). Here, the degree of quantization means, for example, a quantization step, the number of sub bit planes, and the like. Further, the quantization unit 10 performs at least one of quantization of the frequency coefficient and discarding the lower bits of the frequency coefficient.

  As described above, in this embodiment, the second method (method for generating a specific image C (see FIG. 4) with the smallest code amount in the foreground P ′) is used. The specific image C is often a color extremely different from the colors S and T. Therefore, as a first problem, when the high frequency component of the foreground P ′ is not left, the color S, the color T, and the color V are mixed, resulting in a problem.

  Therefore, as in this embodiment, the degree of quantization of the high-frequency component HH of the non-pattern part color image (foreground P ′) is weakened (that is, the high-frequency component is left), so that the colors S, T, and V There is an effect that there is no inconvenience of mixing and.

  As a second problem, in the foreground P ′ of MRC, the reproducibility of the colors S and T which are the colors of the characters A1 and A2 is important, while the reproducibility of the color V which is the specific image C is almost the same. There are circumstances that are not questioned. In particular, in a form in which one of the foreground P ′ or the background Q is selected by the mask, the reproducibility of the color V of the non-character pixel C is not questioned at all.

  Therefore, when the foreground P ′ of MRC is quantized, it is not limited to the first method and the second method, but it is preferable that the quantization error is originally concentrated on the specific image C and no error is generated in the character pixel. .

  However, in the foreground P according to the first method, it is difficult to perform the above-described control “concentrate quantization errors on non-character pixels and do not cause errors on character pixels”. This is because the first method is to create a foreground P by blurring in a gradation between the color S of the character A1 and the color T of the character A2. Therefore, the character pixel and the non-character pixel belong to the frequency of the same band (almost low frequency) in terms of frequency (the accurate expression is that the frequency calculated from the character pixel and the non-character pixel is the same low frequency band. The above expression is used for convenience here).

  In the encoding with frequency conversion, the frequency coefficient is quantized, but the frequency coefficient in the same band is usually subjected to almost the same degree of quantization. When the quantized coefficient is returned to the pixel value by performing inverse quantization and frequency inverse transform, the quantization error generated in the frequency coefficient is reflected in the error of the pixel value used to calculate the frequency coefficient. Therefore, almost the same error occurs in the coefficients belonging to the same band.

  That is, in the first method that does not cause an edge, an error cannot be concentrated on a non-character pixel because an edge is not generated.

  On the other hand, in the foreground P ′ according to the second method, the two pixels adjacent to the edge, that is, the character pixel A1 and the non-character pixel (specific image C) adjacent thereto belong to the high frequency band. However, other non-character pixels belong to the low frequency band. When character pixels of similar colors are adjacent to each other, these character pixels also belong to the low frequency band, but in general, considering that the ratio of the character pixels occupying the foreground P ′ is small, Most are non-character pixels.

  In other words, according to the second method, most of the non-character pixels can be made to be pixels that generate a low frequency, and if the low frequency is quantized to a large extent, the non-character pixels are made to be non-character pixels as compared with the case where it does not. The error can be concentrated.

  In encoding an image, it is not usually performed to quantize a large amount of low frequencies. This is because humans are sensitive to low frequency errors. However, under the special condition of MRC encoding using the second method for generating the foreground P, the low frequency should be quantized more, and the quantization error is in a reconstructed state (background In the image in a state where the foreground and the foreground are combined), the image is not visible to the observer.

  From the above, even when considering the function of the layer itself called “foreground color”, the foreground image creation method (this is often determined in advance by the application) is further taken into consideration, and this creation method is It should be reflected in the quantization of.

  In this way, the background R and the foreground P are quantized (steps S18 and S24). Thereafter, the encoding unit 12 encodes the background R and the foreground P by a known technique (steps S20 and S26).

  In addition, the encoding unit 12 performs entropy coding of coefficients (MQ coding in bit plane units) for each code block and in bit plane order in precinct units. In the embodiment of the present application, MQ coding is performed from the most significant bit plane to the lower bit plane defined in advance in the table.

  Further, since the encoding unit 12 regards “from the most significant bit plane to the lower bit plane defined in advance in the quantization table” as a necessary code, there is no unnecessary code described in parentheses. For these codes “from the most significant bit plane to the lower bit plane defined in advance by the quantization table”, a header is added and called a packet. The packet header includes information about codes included in the packet, and each packet can be handled independently. In other words, a packet is a unit of code.

  Further, the mask Q and the second mask V are encoded by a known technique MMR (Modified Modified READ) (step S28). Further, the second background W is encoded by JPM (step S30). Then, the packet generation unit 14 combines all the encodings to generate a packet. The packet generator 14 arranges the packets to form a code (see FIGS. 39 and 40). Further, the packet generator 14 may perform bit-plane encoding from the upper bit plane to the (necessary) lower bit plane for each subband. The packet generator 14 collects necessary codes and generates a packet. Then, the packet generation unit 14 generates codes by arranging packets for all the prints (all code blocks = all subbands).

Thus, in the first embodiment, the second method is used by making the degree of quantization of the high-frequency component of the non-picture part color image weaker than the degree of quantization of the high-frequency component of the picture part image. Even if it decodes, when it decodes, a specific image will not be mixed with a character image (character pixel).
[Embodiment 2]
In the second embodiment, the luminance / color difference conversion unit in the DC level shift unit 4 performs luminance / color difference conversion on the pattern portion image R and the non-pattern portion color image P ′, and the luminance Y _R and the color difference C _{R of the} pattern portion image. , and it generates a luminance Y _P and the chrominance C _P of the non-picture portion color images P.

In normal encoding, luminance / color difference conversion is performed, and quantization is performed for each luminance and each color difference. In normal natural images, reflecting human visual characteristics that are more sensitive to brightness than color,
The degree of luminance quantization is generally set to the degree of color difference quantization.

However, the degree of luminance quantization <the degree of color difference quantization or the degree of luminance quantization ≒ the degree of color difference quantization, even if it is a natural image, the luminance and color difference of the MRC foreground P ' It is desirable to change the quantization ratio (the degree of quantization of the color difference / the degree of quantization of the luminance) from that of the background R.

Therefore, in the second embodiment, the degree of quantization of the color difference / the degree of luminance quantization of the non-picture part color image P ′ is determined by the degree of the quantization of the color difference / the degree of luminance quantization of the picture part image. Also make it smaller. By doing in this way, the effect similar to Embodiment 1 can be acquired.

  27, the number of steps of quantization of the color difference Cb of the foreground P ′ shown in FIG. 25 / the number of steps of quantization of the luminance Y (shown by a solid line, hereinafter “ratio of the color difference Cb of the foreground P ′”). 26) and the number of steps of quantization of the color difference Cb of the background R / the number of steps of quantization of the luminance Y (indicated by a broken line, hereinafter referred to as “ratio of the color difference Cb of the background R”) shown in FIG. Indicates. The same applies to the broken and solid lines. Further, “the ratio of the degree of quantization of the color difference (number of quantization steps) to the degree of quantization of the luminance (number of quantization steps)” is expressed as “the number of steps of quantization of the color difference Cb (or Cr) / quantum of the luminance Y”. The number of steps of conversion.

  Also, the vertical axis of FIG. 28 shows the number of steps of quantization of the color difference Cr of the foreground P ′ / the number of steps of quantization of the luminance Y shown in FIG. 25 (indicated by a solid line, hereinafter “ratio of the color difference Cr of the foreground P ′”). )) And the number of steps of quantization of the color difference Cr of the background R shown in FIG. 26 / the number of steps of quantization of the luminance Y (indicated by a broken line, hereinafter referred to as “ratio of the color difference Cr of the background R”))) It shows. In both FIG. 27 and FIG. 28, the horizontal axis represents the frequency band. In the frequency band, 2LL is the lowest frequency body in all ratios, 2HL and 2LH are the same frequency band, and 1HH is the highest frequency band.

  As described above, in the second embodiment, the degree of quantization of the color difference / the degree of luminance quantization of the non-picture part color image R is set as the degree of quantization of the color difference / the degree of luminance quantization of the picture part image. Than (except 2LL).

Here, “except 2LL” is used because the number of substantial quantization steps of 2LL is often 1 for both luminance and color difference. That is, there are many cases where 2LL should be excluded in the comparison of magnitude relations.
[Embodiment 3]
In the encoding apparatus of the third embodiment, the high-frequency component of luminance Y of the non-picture part color image P ′ (foreground P ′) is most quantized, and the high-frequency component of the color difference of the picture part image is most quantized. Thereby, it is possible to further encode with a small code amount. Further, in the encoding apparatus of the third embodiment, either the first method or the second method may be used.

  First, the reason for “quantizing the high frequency component of the luminance Y of the non-pattern part color image P ′ (foreground P ′) most” will be described. The non-pattern part color image P ′ is an image representing “color”. Therefore, the luminance Y is most quantized in order to leave the color difference of the non-pattern part color image P ′. Further, it is known from experience that a large amount of color difference remains by quantizing the high-frequency component of luminance Y most.

In addition, the reason for “quantizing the highest frequency component of the color difference of the pattern portion image most” will be described. As mentioned above, humans are more sensitive to brightness than color. Therefore, the high-frequency component of the color difference of the pattern portion image is most quantized.
[Specific Example 1]
A specific example for realizing the first and second embodiments will be described. As a specific example 1, the weighting unit 16 is further provided in the encoding apparatus shown in FIG. In specific example 1, a weighting coefficient is used. In this specific example 1, Visual Distance is used. The Visual Distance will be described later.

  The subband band of the wavelet transform described above is calculated from the band of the image itself. However, the closer the visual distance to the image is, the more high frequency errors are noticeable, and the far distance the visual distance is, the lower frequency errors are only noticeable, so the subband bandwidth for the observer's eyes is It becomes a function (for example, a linear function) of the observer's visual distance. For this reason, the weighting coefficient is shown for each of the above visual distances, and is set such that the farther the visual distance is, the less important the high frequency than the low frequency and the less important the color difference component than the luminance component. Has been.

  In addition, the frequency conversion unit 8 according to the first specific example divides the input signal for each subband by frequency conversion. Then, the weighting unit 16 calculates a weighting coefficient for the degree of quantization for each subband, and performs weighting with the weighting coefficient. In calculating the weighting coefficient, it is assumed that Visual Distance of the non-picture part color image <Visual Distance of the picture part image. The calculation of the weighting coefficient will be described in [Specific Example 2].

  Specific example 1 will be described using specific numerical values. FIG. 29 shows the substantial number of quantization steps of the non-picture part color image P ′ of the first modification. The values shown in FIG. 29 are obtained by applying Visual Weight 800 of Visual distance 800 to each frequency / component in the above equation (7) and setting n = 64. As described in the first embodiment, the values shown in FIGS. 26 and 29 are obtained by making the number of quantization steps of the high frequency component of the non-picture part color image weaker than the number of quantization steps of the high frequency component of the picture part image. Yes. FIG. 30 shows a comparison of the substantial number of quantization steps of the foreground P ′ and the background Q with respect to luminance. In FIG. 30, the background Q is indicated by a broken line, and the foreground P is indicated by a solid line.

  Similarly, the actual quantization step numbers of the foreground P ′ and the background Q shown in FIGS. 27 and 30 are expressed as a ratio of color difference Cb / luminance Y and color difference Cr / luminance Y for each frequency. 31 and 32. As described in the second embodiment, as is clear from FIG. 31, “substantial quantization step number of color difference / substantial quantization step number of luminance” is the value of foreground P ′ (except 2LL). Is smaller than the background Q. Also in FIG. 32, except for 1HH, the “substantial quantization step number of color difference / substantial quantization step number of luminance” is smaller for the foreground P ′ (except for LL) than for the background Q. In FIG. 32, the description is “except for 1HH”. However, as apparent from comparison at all frequencies, on average, “substantial quantization step number of color difference / substantial quantum of luminance”. The number of conversion steps ”(excluding LL) is smaller in the foreground 2 than in the foreground 1 (described in the second embodiment).

In addition, the Visual distance in FIG. 26 is 1000, and the Visual distance in FIG. 29 is 800. Therefore, it is the value of this modification 1.
[Specific Example 2]
In this second specific example, a second specific example using VisualWeight in order to realize the first and second embodiments will be described. The VisualWeight and the Visual Distance described in the specific example 1 will be described.
<1> Contrast Sensitivity Function (CSF) and its Formula Prior to the calculation of Visual Weight, the Contrast Sensitivity Function (CSF) and its formula will be described first. As will be described later, Visual Weight is obtained as an integral value of CSF in a predetermined band. CSF indicates the reciprocal of the difference / sum of contrast (= contrast) that an observer can discriminate when observing a line pair (stripes) of a predetermined spatial frequency in a predetermined visual distance, and is obtained by experiment. Is. In general, humans are less likely to discriminate contrast at higher frequencies.

When the contrast is considered as “the difference between the original image and the image after compression / expansion”, the CSF shows the limit that “the difference in density between the original image and the image after compression at each frequency cannot be determined”. Then, since the wavelet coefficient generally reflects the contrast of the original image, the CSF is “if the difference between the wavelet coefficient of the original image and the compressed image is less than this at each frequency (= each subband), the coefficient reflects. It is interpreted as an index indicating a limit that the difference between the original image and the image after expansion cannot be determined. Therefore, if the wavelet coefficient is quantized with “twice or less of the reciprocal of CSF”, the quantization error generated in each coefficient is less than or equal to the reciprocal of CSF, so that image quality degradation should not be recognized. .
<2> How to calculate VisualWeight,
Since CSF is “allowable quantization error amount of wavelet coefficient at each frequency”, for example, let us consider an allowable quantization amount of 1H subband obtained by decomposing once in the vertical direction. The 1H subband includes a wide range of frequencies higher than a certain frequency that have been passed by the high-pass wavelet filter. The frequency of the CSF to be used is simply considered as follows.

First, when the resolution of the original image is d [dpi] and this is observed from the distance v [inch], the approximate expression of the spatial frequency f is
f = v · d · tan (π / 360) (8)
This is the maximum frequency that the original image can contain.

  Here, v · d in the right side of the above equation (8) is Visual Distance (Visual distance in Japanese). The Visual Distance is originally calculated from the actual visual distance of the subject, and as described later, “To calculate a quantization parameter that gives the best image quality when an ordinary natural image is observed in the Visual distance” Is used.

  However, if a value other than the actual visual distance is used as the visual distance, “a quantization parameter that gives the best image quality when a normal natural image is observed at a distance that is not the visual distance” or “in the visual distance” When an image that is not a normal natural image (ex. Foreground of MRC) is observed, it is possible to obtain a “quantization parameter that gives the best image quality”. In the present application, in order to calculate the latter parameter, a value other than the actual visual distance may be used as the visual distance.

As described above, the maximum spatial frequency fmax that can be included in the original image is
f max = v · d · tan (π / 360) (9)
In the normal wavelet transform, since the sub-sampling is halved for each decomposition, the maximum value of the spatial frequency is halved for each decomposition level.

  Therefore, the wavelet filter is regarded as an ideal high-pass filter and low-pass filter, and the band of 0 ≦ f ≦ fmax is divided into 0 ≦ f <fmax / 2 and fmax / 2 ≦ f ≦ fmax in one decompression. Assume that That is, the band of 1H coefficient is regarded as fmax / 2 ≦ f ≦ fmax, and the band of 1L coefficient is regarded as 0 ≦ f <fmax / 2.

  Therefore, the average value of CSF in the band fmax / 2 ≦ f ≦ fmax is used as the csf of the 1H coefficient, and the average value of CSF in the band 0 ≦ f <fmax / 2 is used as the CSF of the 1L coefficient. Then, based on the fact that the characteristics of the linear filters connected sequentially are expressed by the product of the characteristics of the individual filters, the CSF of the 1LH coefficient has an average value in the band fmax / 2 ≦ f ≦ fmax and the band 0 ≦ f. The product of the average values at <fmax / 2 is used.

  Here, the average value is obtained by dividing the integration of CSF in each band by the width of the integration interval.

  Therefore, the value of CSF to be applied to each subband is calculated by the following equation, which is called Visual Weight (or CSFweight) in JPEG2000.

h is horizontal, v is vertical, Wh is the integration in the horizontal section, and Wv is the integration in the vertical section. W (f) is a CSF normalized to 1 at the maximum value, and w on the left side using one normalized to 1 for each component of Y, Cb, and Cr represents VisualWeight.

Then, VisualWeight is not applied to the non-picture part color image, and VisualWeight is applied to the picture part image to the picture part image. As described above, the first and second embodiments are realized by using VisualWeight.
[Modification 1]
Modification 1 will be described. A case where the wavelet transform of the background R and the foreground P ′ is performed using a 9 * 7 filter in step S16 and step S22 in FIG. 9 will be described. In this case, by using the values shown in FIGS. 24 and 26, the number of quantization steps including normalization is calculated by the above equation (7), and the frequency coefficient is quantized. However, in the case of the second modification, not all the bit planes of the quantized coefficients are encoded, but the lower sub-bit planes of the number shown in FIG. 33 are not encoded. Here, t in FIG. 33 is an arbitrary integer, and the code amount itself can be selected depending on the magnitude of t.

As mentioned above, the quantized coefficients are normalized for mathematical errors (per bit plane) and also reflect visual weights, so the same number of coefficients at all frequencies. If the number of sub-bit planes corresponding to the number of sheets is omitted, balanced subjective image quality can be obtained. However, in this example, all LL sub-bit planes are encoded.
[Modification 2]
Next, Modification 2 will be described. In the second modification, a case where the wavelet transform of the background R is performed using a 5 * 3 filter will be described. This is an example in which quantization is performed by omitting encoding of the lower bit plane.
FIG. 34 shows the square root of the frequency gain of the wavelet transform using a 5 * 3 filter. FIG. 35 shows the square root of the gain of reversible color conversion (RCT) in the second modification. As described above, if the coefficient for each frequency is divided by the reciprocal of the square root of “frequency conversion gain × color conversion gain” (shown in FIG. 36), the influence of the quantization error generated in each coefficient can be reduced between frequencies and Although it can be made uniform between components, in this example, quantization is performed by omitting encoding of the lower bit plane without using linear quantization. That is, using Equation (7), the Visual Weight is further reflected in the value of the constant multiple in FIG. 36, and this is converted into an exponent (a value approximated to an integer) when this is converted to a power of 2. Calculate the number of bit planes to be omitted.

  As described above, FIG. 37 shows the number of bit planes that are not encoded because of the foreground P ′ in the second modification. This is calculated as an exponent when converting Visual Weight to 1 for each frequency / component in equation (6) (= Not applying Visual Weight) and n = 64. It is.

  On the other hand, FIG. 38 shows the substantial number of quantization steps for the background Q in the third modification. This is calculated as an exponent when the Visual Weight of Visual distance 1000 is applied to each frequency / component in Equation (6) and converted to a power of 2 with n = 64. In addition, since the Visual Weight for RCT is not prescribed | regulated, the thing for ICT is diverted in a present Example.

  Note that the layout and layout information of each layout object with respect to the background in JPM. As shown in FIG. 39, the layout object ID, the stacking order, the number of vertical and horizontal pixels, and the layout offset with respect to the background are the layout object. Enter in the Header Box.

  In addition, as for the foreground and background arrangement and size information constituting each layout object, the arrangement offset with respect to the background is described in the Object Header Box as shown in FIG. The number of pixels of each object is described in the code of each object.

  In the above embodiment, JPM is used as the file format. However, any MRC type may be used, and MRC type PDF is also in that category.

In the above embodiment, the wavelet transform is used for the frequency transform. However, the present invention is not limited to this, and it is apparent that other frequency transforms represented by the discrete cosine transform can be applied. In addition, the term “high frequency component” in the claims of the present application refers to only the highest frequency band (coefficient), and at least the highest frequency band (coefficient) and the lowest frequency band (coefficient) When not included, two ways are conceivable.
[program]
FIG. 41 is a hardware configuration diagram showing the main configuration of the status notification device according to the present invention. The state notification device includes, as main components, a CPU 101, a ROM (Read Only Memory) 202, a RAM (Random Access Memory) 102, an auxiliary storage unit 204, an external storage device I / F unit 205, an input unit 206, a display unit 207, And the communication device 208.

  The CPU 201 is composed of a microprocessor and its peripheral circuits, and is a circuit that controls the entire state notification device. The ROM 202 is a memory that stores a predetermined control program (software component) executed by the CPU 2, and the RAM 203 executes various control operations by the CPU 201 executing a predetermined control program (software component) stored in the ROM 202. This is a memory used as a work area (work area) when performing.

  The auxiliary storage unit 204 is a device that stores various information including information related to a project such as a general-purpose OS (Operating System), a project management program according to the present invention, and task information, and is an HDD (Hard Disk) that is a nonvolatile storage device. Drive) is used. In addition to the auxiliary storage device 204, the various types of information may be stored in a storage medium such as a CD-ROM (Compact Disk-ROM) and a DVD (Digital Versatile Disk), and other media. Various information stored in the storage medium can be read via a drive device such as the storage medium reader 205. Therefore, various information can be obtained by setting the recording medium 515 in the storage medium reader 205 as necessary.

  The input device 206 is a device for the user to perform various input operations. The input device 206 includes a mouse, a keyboard, a touch panel switch provided so as to be superimposed on the display screen of the display device 207, and the like.

  The display unit 20 is a device that displays various data related to project management on a display screen. For example, it is composed of LCD (Liquid Crystal Display), CRT (Cathode Ray Tube) and the like.

2 ... Decomposition unit 4 ... DC level shift unit 8 ... Frequency conversion unit 10 ... Quantization unit 12 ... Encoding unit 14 ... Packet generation unit

H11-177777 No. 2007-005844

Claims (8)

From the original image, a pattern portion image in which the pattern of the original image is displayed, and a non-pattern portion color image in which a non-pattern color image of the original image and a specific image occupying a position other than the color image of the non-pattern are displayed A generating unit for generating a mask image on which the shape of the non-pattern is displayed ;
A frequency converter that converts the frequency of the pattern part image and the non-pattern part color image;
A quantization unit that quantizes the frequency coefficient of the picture part image and the frequency coefficient of the non-picture part color image obtained by the frequency conversion unit ;
An encoding unit that encodes the frequency coefficient of the quantized picture part image and the frequency coefficient of the non-picture part color image;
The quantization, except for normalization for frequency gain,
Wherein the quantization step number of the frequency coefficients of the high frequency band of the non-picture portion color images, the encoding apparatus characterized Reduces the than the quantization step number of the frequency coefficients of the high frequency band of the picture portion the image. From the original image, a pattern portion image in which the pattern of the original image is displayed, and a non-pattern portion color image in which a non-pattern color image of the original image and a specific image occupying a position other than the color image of the non-pattern are displayed A generating unit for generating a mask image on which the shape of the non-pattern is displayed ;
The pattern portion image, the non-picture portion color image by converting luminance and color difference, the luminance and chrominance of the picture portion image, and the luminance and color difference conversion unit that generates a luminance and color difference of the non-picture portion color images,
Luminance and chrominance of the picture portion image, and a frequency converter for frequency-converting the luminance and chrominance of the non-picture portion color images,
Luminance and chrominance frequency coefficients of the picture portion image obtained by the frequency conversion unit, and a quantization unit for quantizing the luminance and color difference frequency coefficients of the non-picture portion color images,
Luminance and chrominance frequency coefficients of the quantized picture unit image, and has an encoding unit for encoding the luminance and chrominance frequency coefficients of the quantized non-picture portion color images,
The quantization includes quantization for normalization for frequency gain and quantization for normalization of luminance / chrominance conversion gain,
Wherein the non-picture portion color images, the number of quantization steps of the quantization step number / luminance color difference, the picture part image, smaller remote by number of quantization steps of the quantization step number / luminance color difference An encoding device. From the original image, a pattern portion image in which the pattern of the original image is displayed, a non-pattern portion color image in which the non-pattern color image of the original image is displayed, and a mask image in which the shape of the non-pattern is displayed , a generating unit that generates a,
The pattern portion image, the non-picture portion color image by converting luminance and color difference, the luminance and chrominance of the picture portion image, and the luminance and color difference conversion unit that generates a luminance and color difference of the non-picture portion color images,
Luminance and chrominance of the picture portion image, and a frequency converter for frequency-converting the luminance and chrominance of the non-picture portion color images,
Luminance and chrominance frequency coefficients of the picture portion image after frequency conversion which is obtained by the frequency conversion unit, and a quantization unit for quantizing the luminance and color difference frequency coefficients of the non-picture portion color images,
Luminance and chrominance frequency coefficients of the quantized picture unit image, and has an encoding unit for encoding the luminance and chrominance frequency coefficients of the non-picture portion color images,
The quantization includes quantization for normalization for frequency gain and quantization for normalization of luminance / chrominance conversion gain,
An encoding apparatus characterized in that the frequency coefficient of the high-frequency band of luminance of the non-picture part color image is most quantized and the frequency coefficient of the high-frequency band of color difference of the picture part image is most quantized. The quantization unit, the encoding apparatus according to any one of claims 1 to 3, characterized in that the lower bits of discarded of frequency coefficients. The frequency conversion unit divides each subband by frequency conversion,
Calculating a weighting coefficient for the quantization for each subband, and having a weighting unit that performs weighting with the weighting coefficient,
5. The encoding device according to claim 1, wherein the weighting unit sets the visual distance of the non-picture part color image to the visual distance of the picture part image when calculating the weighting coefficient. The encoding unit employs the JPEG2000 system,
The encoding device according to claim 1, wherein VisualWeight is not applied to the non-picture part color image, and VisualWeight is applied to the picture part image.   The program for functioning a computer as an encoding apparatus in any one of Claims 1-6.   A computer-readable recording medium on which the program according to claim 7 is recorded.