JPH01157168A - Image encoding system - Google Patents

Abstract

PURPOSE:To efficiently perform the encoding of data without generating remarkable deterioration on multivalue data by classifying a block to several defined categories, and performing quantization at every category. CONSTITUTION:Image data with an image element of 8 bits is segmented as signals X(x11-x44) whose one block is constituted of (4X4) image elements, and signals Y(y11-y44) can be obtained by performing the Hadamard transformation of the image data in the signal X. At the time of performing preliminary scan, frequency classified at every category classified at a segmentation part 102 is measured by an image analyzer 112, and a decided image type is outputted to the segmentation part 102. Coefficients y12-y12 are classified to several kinds of categories corresponding to the size of the coefficient and the feature of an original image. Selectors 103 and 111 select scalar quantizers 105-107 and vector quantizers 108-110 corresponding to each category, and the data are outputted as codes of 9 bits.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an image encoding method, and particularly to an image encoding method for encoding halftone image information.

[Prior Art] Conventionally, when transmitting and storing images, it has been common to reduce redundancy through encoding (generally referred to as compression) in consideration of efficiency. In recent years, with the development of digital image processing technology and device technology, images to be transmitted and stored have shifted from binary to multivalued images, and the resolution has also increased. As a result, the amount of data has become enormous, and highly efficient encoding technology has become necessary.

However, most of the encoding techniques to date have been aimed at binary images, for example, the representative ones are MH and MR, which were devised for facsimile.

MMR, etc., but these methods are essentially not suitable for encoding multivalued images. In addition, several encoding techniques for multivalued images have been reported, including block encoding, predictive encoding, orthogonal transform encoding, and the like. However, most of these methods are aimed at television images, and not general document images.

It is not suitable for encoding halftone images.

Particularly in orthogonal exchange coding, the conventional property that power is concentrated in small parts of the sequence has been utilized. However, although this method is effective for continuous tone images such as photographs, it has the disadvantage of causing significant deterioration when applied to images with unique edge structures such as document images or halftone dot images. .

[Problems to be Solved by the Invention] The present invention eliminates the drawbacks of the above-mentioned conventional examples and efficiently handles multivalued data including document images, halftone images, and photographic images without significant deterioration. An image encoding method for encoding is provided.

[Means for Solving the Problem] As a means for solving this problem, the image encoding method of the present invention divides an image into blocks of a predetermined size and performs orthogonal exchange on each of the blocks. The blocks are classified into several predetermined categories based on the coefficients of the transformed matrix and the characteristics of the input image detected or set in advance, and quantization is performed for each category.

[Operation] With this configuration, image information is pre-scanned to determine the nature of the image, such as a document image or a halftone image.

It classifies continuous tone images such as photographs, and uses this classification information to solve inappropriate encoding caused by extracting features only in blocks.

[Embodiment] FIG. 1 shows the configuration of an image encoding device according to an embodiment that implements the image encoding method of the present invention.

Here, 101 is an Agemard transform unit that performs Hadamard transform, 102 is a segmentation unit that classifies images into categories that will be explained in detail later, and 103 is a selector that selects a quantizer corresponding to the category based on the output from the segmentation unit 102. It is. 104 is a scalar quantizer (5calar Q
uantizationH (hereinafter referred to as SQ), 105,
106, 107 are scalar quantizers (SQ) that perform scalar quantization suitable for each category, 108, 109,
110 is a vector quantizer that performs vector quantization suitable for each category.
on: hereinafter VQ). 111 is a segmentation unit 10 from the output of the vector quantizer of each category.
This is a selector that selects a block corresponding to the block to be processed using an output line output from 2, and functions similarly to the selector 103 at the previous stage.

Reference numeral 112 denotes an image analyzer that performs image analysis through preliminary scanning performed prior to image encoding, which identifies the type of image to be encoded: 1. halftone image, 2. character image, 3. photographic image (continuous). It is classified into three types (tone). Preliminary scanning and actual encoding are the same operation, but during preliminary scanning, segmentation unit 1
Signal line 1 indicates the frequency of each category classified in 02.
13, the image analyzer 112 determines the type of image, and no operation after the selector 103 is performed.

At the time of encoding, the image type (the above-mentioned three types) determined during preliminary scanning by the image analyzer 112 is output to the segmentation unit 102 via the 2-bit signal line 114, and is fed back to category classification.

The operation will be explained here. Image data of 8 bits per pixel is cut out as 1 block of 4×4 pixels. This is designated as X (X++ to x44) shown in FIG. This one
When the image data of the block is subjected to Hadamard transformation, it becomes Y (y++ to y44) in Fig. 2, where Y++ is expressed as a value of O to 1020, that is, 10 bits, and YI2 to Y44 are each -
It can be expressed by a value of 510 to +510, that is, 10 bits.

Therefore, the scalar quantizer 104 quantizes YI) into 8-bit data, and uses this as the average brightness value within the block. Yll to Y44 are classified by the segmentation unit 102 into several categories depending on the magnitude of the values of these coefficients and the characteristics of the original image.

Selectors 103 and 111 each select a scalar quantizer and a vector quantizer according to this category, and
Output as a bit code (pattern code).

The Hadamard transform, segmentation, scalar quantization, and vector quantization will be explained below.

FIG. 2 shows a 4×4 Hadamard transformation, where X is a signal before transformation and Y is a signal after transformation.

Note that this conversion is expressed as . Here, Hts: 16th order Hadamard matrix, = [XII,
X12. XI3. XI-, X21. X
22°..., Xa8. X44]τ 1 or less margin 1 = [Y ++, Y r□, Y +3. Y 14. Y
211 Y 22°...+ Y431 Y44
] If T, then it becomes.

FIGS. 3(a) to 3(p) show examples of Hadamard transform, and when there is a vertical edge in the block as shown in FIG. 3(a), a large value appears in the coefficient Yrz. In addition, the sign is "0" in the left half and "2" in the right half depending on the slope of the brightness of X.
When the left half is "255" and the right half is "O", it is - (minus), and when the left half is "O", it is + (plus). However, the coefficient Y
Since z represents the brightness average, it will not be particularly explained here.

Similarly, FIG. 3(b) shows a case where there is a horizontal edge in the block, and a large value appears in the coefficient Y21.

Other patterns in FIGS. 3(C) to 3(p) show patterns such as vertical and horizontal lines, key shapes, diagonal edges, and diagonal lines, as shown in the figures.

FIG. 4 explains parameters when performing segmentation. The illustrated parameters are determined by focusing on the characteristics of the pattern shown in FIG. 3. Here, UFE is a vertical edge component parameter, OLE is a vertical line component parameter, 1 (EE is a horizontal edge component parameter,
) ILE is a horizontal line component parameter, OTH is a component parameter such as edges of other families, and EF is a parameter representing the strength of an edge, and when this is large, it is understood that the edge is strong. Similarly, LF is a parameter indicating the strength of a line segment.

FIG. 5 shows a flowchart of a procedure for classifying categories into 16 types using the five parameters shown in FIG. 4, and each branch and each category has the following meaning. This is called segmentation.

Category 20 = diagonal edge, line 21: diagonal edge, line + complex pattern 22: horizontal edge + complex pattern 23: vertical edge + complex pattern 24: horizontal line + complex pattern 25: vertical line + complex pattern Category 30: monotonous flat part 31 : Flat edge 32: Horizontal edge with small gradation difference 33: Vertical edge with small gradation difference Category IJ-40: Horizontal edge with large gradation difference 41: Vertical edge with large gradation difference IJ-50: Large gradation difference Weft edge pattern 51: Weft pattern with large gradation difference 52: Weft pattern with large gradation difference 53: Warp pattern with large gradation difference In addition, in FIG. 5, each branch has the following meanings. Yes.

<Branch 1> Separation of strong diagonal line (large OTH) Branch 2> Separation of edge pattern (large VEE, HEE) and linear pattern (large VLE, HLE) <Branch 3> Separation of strong horizontal edge pattern ( Extracting branch 4> Vertical edge (large VEE) and horizontal edge (H
Branch 5: Extraction of complex patterns including diagonal edges (large OTH). Branch 6: Separation of weak vertical and horizontal patterns (large EF) from flat parts. Branch 7: Separation of vertical edges (VEE is large) and horizontal edge (F
Branch 8> Extract those with large 0T) I from among complex patterns. Branch 9> Extract those with large OLE from among complex patterns. Branch 10> Extract those with large OLE from among complex patterns. Extract large patterns (Branch 11) Extract those with large HEE from complex patterns Branch 12> Extract strong vertical and horizontal line patterns (large LF) Branch 13> Extract vertical and horizontal patterns (large VLE) and horizontal line patterns (HLE (branch 14) Extract patterns that include edge components (with a large EF) from the line patterns separated in branch 2 (branch 15) Separate vertical edge (large VEE) and horizontal edge (HEE) Figure 6 shows a hardware block diagram for implementing the category classification algorithm shown in Figure 5. Shown below. Here, 501 is a parameter calculation unit that calculates the parameters shown in FIG. 4, 502 to 511 are comparators that determine each branch 1 to 16 shown in FIG. 5, and 512 is 16 types depending on each branch information. This is a lookup table (hereinafter referred to as LUT) that generates a category pattern of 4 bits as a 4-bit pattern.

As mentioned above, the categories are determined according to the image information, and in the case of halftone images, the categories are concentrated in categories 21 to 25, in the case of continuous tone images such as photographs, the categories are concentrated in categories 30 to 33, and in the case of text images, the categories are concentrated in categories 21 to 25. There are many categories 30 in the background part, and categories 40, 41 and 50 in the text part.
~53 tends to occur.

However, for example, in the case of a halftone image, if only a small area within a block is categorized, the category classification will result in an area with edges or lines, as described above, and if this is combined, The edges of the halftone dots are emphasized, resulting in a grainy reproduced image. Therefore, in the case of halftone images, it is rather category 30 to 3.
The deterioration in image quality will be less if the number of classifications in 3 is increased and the image is encoded like a photograph.

Therefore, as a preliminary operation, the frequency of each category is calculated in the image analyzer 112 via the signal line 513 using the category pattern determined by the LtJT 512 as shown in FIG. If there are many categories 21-25, it is determined that it is a halftone image, and the number of branches that fall into categories 21-25 is reduced and the number of branches that fall into categories 30-33 is increased. Actually, the branching into categories is corrected by increasing the comparison value 32 at branches 5 and 16 in FIG.

In FIG. 6, during image encoding, the comparison value 32 of OTH≧32 is changed from the image analyzer 112 via the signal line 517 to output, for example, 64, and changed to the comparison value of OTH≧64.

In the case of a photographic image, in order to further increase the number of 31 to 33, branch 5 and branch 16 are similarly changed to 0T)I≧96. In this case as well, this is done via the signal line 517.

For text images, use category 30 in the background.
, category 40.41 and category 50-5 in the character part
In order to increase the number of 3, change EF≧96 to EF≧64 at branch 3,
At branch 6, EF≧16 becomes EF≧32, and at branch 13, LF≧
64 to LF≧32, EF≧ to EF≧32 in branch 14, and signal lines 516 to 518, 514.41, respectively.
This is realized by changing the data of the comparator using 4.

FIG. 7 is a simple explanation of quantization. In the figure, one pattern of category 40 is explained as an example. First, for patterns that have been classified by segmentation, the +/- sign and absolute value of each coefficient except Yll are separated and are called "phase component" and "amplitude component," respectively. For the amplitude component, each coefficient is weighted and nonlinear scalar quantization is performed.The coefficients in the shaded area in the figure are ignored, and the numbers are the results of rounding.

FIGS. 8(a) to 8(p) show the nonlinear scalar quantization results for all 16 categories, and the coefficients indicated by dots in the figure are the coefficients to be ignored as described above. The numbers also represent the number of bits allocated to each coefficient.

FIG. 9 is a hardware block diagram for realizing the Hadamard transform, in which all coefficients are computed in parallel to increase speed.

A data input line 801 receives 16 pieces of 8-bit data for each pixel. 802 is its block buffer, which is composed of 4×4=16 latches. 8
03 is an address generator that specifies pixels in the X block buffer 802 and the Hkl element generator 8.
Specify the elements in 04.

That is, it specifies hk+, X, and J in the calculation of the Hadamard transform Y=1/4XH16X (4X4 block size) shown above. 80
Reference numeral 4 generates the hIJ described above, and is composed of a ROM. Also, what is the output? 1++, h2□+h3
□,...＋? 1161, 16 lines are connected in parallel, and the 4 bits output from the address generator are input to the ROM address, and numbers 1 to 16 are specified.

In the figure, i, j=1-4, K, tan=1-16.

805 is an arithmetic unit for the Yll coefficient of the Y matrix element, and as can be seen from the calculation formula shown above, it is composed of an adder. 806 and 807 are arithmetic units for coefficients other than Yll, each of which is composed of an adder/subtracter as shown in the calculation formula.

808, 809, 810 are 1/4 dividers, and the lower 2
It truncates bits. 811,812,
813 is the result of these calculations, Yll is a 10-bit positive number, and the rest are flag bits indicating +/- signs and 9-bit complement numbers are output.

As mentioned above, the circuit operates on 16 coefficients in parallel. The calculation details are as shown in the formula. That is,
Data is read from the block buffer 802 pixel by pixel, and addition or subtraction is performed depending on the +/- sign output from the Hkl element generator rake 804.

In this example, the Hadamard transform was used for ease of calculation and hardening, but the same idea may be applied using similar orthogonal transforms (discrete CO8 transform, slant transform, etc.) .

In this example, VEE is used for segmentation.

Five parameters were used: VLE, HEE, HLε, and OTH, but basically they are vertical, horizontal edge, vertical and horizontal line,
Any other method that performs categorization focusing on diagonal edges or the like can be used instead. For example, V.E.
E = (Y+-), VLE = (Y14), H
EE = (Y21), HLE = (Y41), 0T
) I ” (Y22), etc.

In this embodiment, scalar quantization and vector quantization are used for quantization, but the present invention is not limited to this, and scalar quantization or vector quantization alone may be used. In addition, “phase component”,
It uses a method of separating into “amplitude components” and quantizing them, but
I'm not limited to that.

Furthermore, in this embodiment, preliminary scanning of the image was used to determine the characteristics of the image, such as a halftone image, a photo image, a 3-character image, etc., and the results were fed back to the category classification at the time of image encoding. , the following other embodiments are also possible.

If the user wants to set the image type, the configuration is such that the comparison in the segmentation section can be manually set without preliminary scanning.

FIG. 10 shows the segmentation unit in that case. In Fig. 10, 520 is a controller, 521.52
Reference numeral 2,523 is a manual switch located in the controller for setting a halftone image, a photograph, and a single character image, respectively. The same operation as described above can be realized by outputting from the controller 520 the same outputs as the signal lines 514 to 518 in FIG. 6 during image encoding according to the state of the switch.

In FIG. 1o, 501 to 512 are the same elements as in FIG. 6.

As explained above, in this embodiment, the type of image is determined by performing preliminary scanning prior to encoding, classifying the image into categories, and measuring the frequency of each category. It feeds back the information and improves image quality deterioration due to inappropriate coding caused by conventional coding for each small block.

Further, by using category classification and using scalar quantization and vector quantization, significant deterioration of edge shapes within a block can be prevented.

[Effects of the Invention] According to the present invention, it is possible to provide an image encoding method that efficiently encodes multivalued data including document images, halftone dot images, and photographic images without significant deterioration.

[Brief explanation of the drawing]

Figure 1 is a configuration diagram of an image encoding device that implements the image encoding method of this embodiment, Figure 2 is an explanatory diagram of Hadamard transform, and Figures 3 (a) to (p) are diagrams showing examples of Hadamard transform. , Figure 4 is an explanatory diagram of parameters for performing segmentation, Figure 5 is a diagram showing a specific example of segmentation, Figure 6 is a hardware block diagram of segmentation and image analyzer, Figure 7 is a diagram of phase components and amplitudes. Figure 8 (a) to (p) are diagrams showing the results of scalar quantization, Figure 9 is a hardware block diagram for realizing the Hadamard transform, and Figure 10 is a diagram showing the separation from the components. FIG. 2 is a hardware block diagram of an example segmentation and image analyzer. In the figure, 101... Hadamard transform unit, 102... segmentation unit, 103... selector, 104...
...Scalar quantizer for brightness average, 105°106,1
07-, I, color quantizer, 108°109.110
...Vector quantizer, 111...Selector, 11
2...Image analyzer, 113°114...Signal line.

Claims (2)

[Claims] (1) Divide the image into blocks of a predetermined size, perform orthogonal exchange on each block, and divide the block into several blocks based on the coefficients of the transformed matrix and the properties of the input image detected or set in advance. An image encoding method characterized by classifying images into predetermined categories and performing quantization for each category. (2) The image encoding method according to claim 1, wherein the properties of the input image are determined based on the frequency of appearance of each category during preliminary scanning.