JPH01146486A - Picture information processor - Google Patents

Abstract

PURPOSE:To carry out the vector quantization of high quality and high compression at a comparatively small lookup table by carrying out integral quantization for one vector per block according to a category obtained with the classification of quantized vector data which are individually vector-quantized. CONSTITUTION:By checking the power, etc., of a prescribed sequency component in the block by a category classifying part 4, the said picture block is classified into one of 8 categories. Thus, immediately after orthogonal conversion, the category classification is carried out at every block by the magnitude of frequencies, further, the picture in the category is band-divided, each of them is vector- quantized, and then the number of the dimensions of the vector is reduced. Category classification code data are inputted from the category classification part 4 through a signal line 4-2 to a vector quantizing part (L) 6 and a vector quantizing part (H) 7, and the different lookup tables(LUT) prepared according to the categories are selected and used.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to an image information processing device, and in particular, to image information processing that orthogonally transforms digital image data block by block, vector quantizes the generated vector data, and encodes the image information. It relates to a processing device.

[Prior Art] In recent years, with the development of device technology typified by VLSIs and the like, the image data that is the subject of data transmission or data storage is shifting from conventional binary images to multi-value images. In addition, the degree of integration of image reading sensors has increased, and the resolution has also increased, resulting in an enormous amount of image data, and it is desired to realize a high-speed, highly efficient encoder. In the past, several reports have been made regarding the encoding of multivalued images. For example, predictive coding,
There are block coding, orthogonal transform coding, compression coding using vector quantization, etc. Of these, block encoding has recently attracted attention because of its ease of handling encoded data, and a method of vector quantization for each block has been devised because it can achieve particularly high compression rates. However, vector quantization has its problems.

The first problem is how to find the optimal reproduction vector for an image. Normally, in vector quantization, a training sequence is performed using various images with different frequencies in advance to obtain reproduction vectors. Then, a reproduction vector that minimizes distortion for the human image is found and encoded. Generally, a vector quantization algorithm called the LBG method is used to find the optimal solution for the reproduction vector. However, when vector quantization is performed using the LBG method on various image data having different frequencies, a reproduction vector that is extremely biased toward image data that appears frequently in a training sequence is obtained as the optimal solution for vector quantization. Therefore, when an image is vector quantized using such a method, there is a drawback that, although the degree of appearance is not so great, characteristic deterioration of the image becomes noticeable.

The second problem is that when converting to hardware, the scale of the hardware becomes extremely large. That is, in order to perform vector quantization on a block-by-block basis, a large-scale look-up table (LUT) is required that inputs all the image data in the block and outputs the code of the reproduction vector. Therefore, the block size for vector quantization is also limited. If the block size is small, it is difficult to utilize the correlation between images, and the compression rate cannot be increased.

[Problems to be Solved by the Invention] The present invention eliminates the above-mentioned drawbacks of the prior art, and its purpose is to perform high-quality, highly compressed vector quantization with a relatively small lookup table. An object of the present invention is to provide an image information processing device that performs the following steps.

[Means for Solving the Problems] In order to achieve the above object, the image information processing apparatus of the present invention includes a blocking means for dividing digital image data into MXN pixel units, and a blocking means for dividing the digital image data into MXN pixel units, and a method for dividing the divided block image data. a transformation means that generates vector data of a predetermined number of dimensions subjected to frequency analysis by performing orthogonal transformation; a classification means for classifying the vector data into bands according to the classified category; a band dividing means for dividing the vector data in the block into bands according to the classified category; and a band dividing means for dividing the vector data into bands according to the classified category. vector quantization means for performing vector quantization; and integrated quantization means for integrally quantizing the quantized vector data that has been vector quantized independently into one vector per block by vector quantization according to the classified category. The outline is to prepare.

Preferably, one aspect of the band division means is to perform band division in such a manner that the number of dimensions of the vector data in the low frequency band is reduced and the number of dimensions of the vector data in the high frequency band is increased.

Preferably, one aspect of the band division means is to perform band division of different band patterns depending on the category of the classification means.

[Operation] In this configuration, the blocking means divides the digital image data into units of MXN pixels. The transform means performs orthogonal transform on the divided block image data to generate frequency-analyzed vector data of a predetermined number of dimensions. The classification means classifies the block image into one of a plurality of categories depending on the magnitude of frequency based on a predetermined vector element of the generated vector data. The band dividing means divides the vector data within the block into bands according to the classified category. Preferably, the band dividing means performs band division so that the number of dimensions of the vector data in the low frequency band is decreased and the number of dimensions of the vector data in the high frequency band is increased. Preferably, the band division means performs band division of different band patterns according to the categories of the classification means, and the vector quantization means independently vector quantizes the band-divided vector data according to the classified categories. .

The integrated quantization means integrates and quantizes the quantized vector data that has been vector quantized independently into one vector per block by vector quantization according to the classified category.

[Description of Embodiments] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

First, the present invention will be explained in detail. First, the present invention is based on the following idea. Highly correlated information sources such as images have the property that power concentrates in low frequencies when frequency analysis is performed. Utilizing this property, efficient encoding is performed by converting the image block by block into the frequency domain and particularly encoding the low frequency components.

However, in order to perform frequency analysis and obtain sufficient correlation, the transform block size must be large to some extent. Therefore, if the converted blocks are directly encoded using a technique such as vector quantization, the problems described above will occur, namely, reproduction vectors that are extremely biased toward image data that appears frequently will be obtained, and Hardware requires a huge amount of LUTs (lookup tables)
is necessary, and does not solve the above-mentioned problem.

Therefore, the present invention solves this problem based on the second idea described below. That is, using image correlation, after orthogonal transformation, each block is classified into categories based on the magnitude of frequency. Furthermore, the number of vector dimensions is reduced by dividing the images within the category into bands and vector quantizing each of them.

This method solves the above-mentioned playback vectors that are biased in appearance frequency, and at the same time reduces the number of LUTs in hardware by band division.

Furthermore, the present invention is based on a third idea.

When encoding an image that has been vector quantized for each category and band division, the obtained reproduction vector is encoded. However, when the image is divided into bands and the reproduction vectors determined independently are encoded, the image Combinations of reproduction vectors for each band that almost never appear when considering the correlation between them are also encoded. Therefore, in the present invention, by considering the correlation and converting independently encoded signals for each band into one code, encoding efficiency that cannot be obtained with the conventional method is obtained.

FIG. 1 is a block diagram of an image information processing apparatus according to an embodiment. In the figure, reference numeral 1 denotes an image input device, which in this embodiment inputs digital image data with a density of 8 bits and a resolution of 16 pel/mm using a raster scan method. Reference numeral 2 denotes a line buffer, which stores four lines of digital image data input along each reading scanning line, and sequentially outputs block image data of 4×4 pixels in parallel.

3 is an orthogonal transform unit, which performs Hadamard transform of a 4×4 matrix in this embodiment. FIG. 2(A) shows the orthogonal transform unit 3
Block image data X+ , X2 , =, X
+a (8 bits each), and FIG. 2(B) shows vector data (sequence components) Yz, Y12. after Hadamard transformation by the orthogonal transformation unit 3. −, Y44 (
FIG. 3 shows a two-dimensional Walsh type transformation pattern employed in the Hadamard transformation of this embodiment. Returning to FIG. 1, 4 is a category classification unit, which classifies the image block into one of eight categories (eight characteristic properties of the image block) by examining the power, etc. of a predetermined sequence component within the block. Classify. Further, the category classification unit 4 extracts the DC component Yll from among the intra-block sequence components Y11 to Y44 and outputs it as is.

Reference numeral 5 denotes a scalar quantization unit, into which sequence components Y, □ to Y44, excluding the DC component Yll among the intra-block sequence components Yll to Y44, are input. According to the conventional method,
The 15 sequence components Y12 to Y44 are scalar quantized with an appropriate number of bits to form 15-dimensional vector elements, and these are vector quantized all at once. However, this requires a huge number of look-up techniques (LOT) and is not practical. Information sources with high correlation, such as general images, have the property that power is concentrated in low frequency components when frequency analysis is performed. Utilizing this property, efficient encoding can be achieved by converting an image block by block into the frequency domain and encoding with emphasis on low frequency components. However, in order to obtain sufficient correlation through frequency analysis, the block size of the orthogonal transform must be increased to a certain extent. Therefore, the number of dimensions of the vector data generated by the transform also increases, and vector quantization is performed as is. If you do this, the LUT will become extremely large. However, in order to perform frequency analysis and obtain sufficient correlation, the transform block size must be large to some extent. If the converted block is then encoded as is using a technique such as vector quantization, the problem described above will occur, namely, the reproduction vector t will be obtained which is extremely biased toward image data that is frequently applied. , and a huge amount of LUTs are required to implement it in hardware. Therefore, in this example, we utilize the correlation of images, and after orthogonal transformation, classify each block into categories based on the size of the frequency, and further divide the image within the category into bands, and vector quantize each to create a vector dimension. Reduce the number. With this method, it is possible to solve the aforementioned playback vector biased toward application frequency and at the same time reduce the number of LUTs in hardware by band division. Accordingly, the scalar quantization unit 5 of this embodiment divides the input sequence components Y12 to Y44 into a predetermined sequence component with a low frequency (low sequence component) and a predetermined sequence component with a high frequency (high sequence component) according to the input category classification data. Each band is divided into bands (components), each band is scalar quantized using a specific number of bits, and the quantized low sequence vector elements and high sequence vector elements are separately provided in a vector quantization unit (LUT). supplying.

Reference numeral 6 denotes a vector quantization unit (L), which vector quantizes low-sequence vector elements formed according to category classification data. Reference numeral 7 denotes a vector quantization unit (I), which vector quantizes high-sequence vector elements formed according to category classification data. A reproduction low sequence vector element and a reproduction high sequence vector element each of 256 fi are calculated for each category based on the training sequence performed on representative images. Therefore, the reproduction vectors of each of these 256 ffi are calculated in advance by each 8
Associated with bit vector quantization code data.

Therefore, the vector quantization unit (L) 6 and the vector quantization unit (H) 7 combine each of the input low sequence vector elements or high sequence vector elements with 256 types of reproduced low sequence data obtained in advance from the training sequence. A look-up table (LUT) is stored that selects 8-bit vector quantization code data that provides a small amount of distortion by comparing vector elements or reproduction height sequence vector elements. In addition, since the high sequence vector element or the low sequence vector element of this embodiment has different sequence components for each category as described later, the vector quantizer (L) 6 and the vector quantizer (H) 7 are Category classification code data is inputted from the category classification section 4 via the signal line 4-2, and different LOTs prepared for each category are thereby selectively used. In this way, a predetermined sequence vector within a block is generated by a vector quantizer (L) 6 and a vector quantizer (H).
7 into 8-bit vector quantization code data and output to signal lines 6-1 and 7-1.

Reference numeral 8 denotes an integrated vector quantization unit, which integrates each 8-bit vector quantized code data obtained by vector-quantizing the low-sequence vector element and the high-sequence vector element, respectively, and converts it into a total of 10 bits of integrated code data. .

9 is a code data storage unit which stores the upper 8 bits of the Yth sequence component (DC component) of the signal line 4-1 and the signal line 4-1.
The 3-bit category classification code of 2 and the 10-bit integrated quantized code data of the signal line 8-1 are stored.

FIG. 4 is a diagram showing the information structure of the block encoding code of the embodiment. In the figure, the block encoding code consists of a total of 21 bits, the 8 bits from the MSB are data for the sequence component Yz (average value), the next 3 bits are category classification code data, and the next 10 bits are data for a predetermined sequence other than Y11. This is integrated code data obtained by vector encoding the components. In this way, digital image data of (4x4) pixels x 8 bits = 128 bits per block is compressed into block encoded code data of fixed length 21 bits, and the data compression rate is approximately 1/6.

FIG. 5(A) is a diagram explaining the basic concept of the category classification method of the embodiment. Categorization in this embodiment is performed based on a predetermined sequence component with a low frequency and based on the magnitude of its power. When we examine the frequency components of various block image data of a general image by Hadamard transform, we find that the power at the edges of the image is concentrated in low sequence components, and its absolute value is large.

In this example, using this property, the edge amount ED is calculated as follows: ED= l Y121 + l Y131 + l Y
It is defined as 21++IY2□1+1Y31+.

FIG. 6(A) is a diagram specifically showing the category classification method of the embodiment. Three threshold values Tl<T2<
Set T3, and in descending order of ED, if ED≦T1, the flat part, if Tl<ED≦T2, the halftone flat part, and if T2<
When ED≦T3, it is classified as a weak edge portion, and when T3<ED, it is classified as a strong edge portion. Furthermore, since there are vertical, horizontal, and diagonal edge patterns in the weak edge portion and the strong edge portion, there are a total of eight classifications as shown in FIG. 6(A).

FIGS. 5(B) and 5(C) show a method of classifying weak edge portions and strong edge portions. For images with strong edges in the vertical direction,
The power in the shaded area in FIG. 5(B) increases, and in an image with strong edges in the horizontal direction, the power in the shaded area in FIG. 5(C) increases. In this embodiment, by utilizing such properties, the following vertical edge amount VE and horizontal edge amount HE are defined.

V E = IY, □l+1Y131+1YI41HE
= l'hl+1Ysrl÷1Y411Using these, the weak edge and strong edge portions are further classified into three categories. That is, IVEI-IHEI>El car)
IVEI ≧1HElr7 check, 1VE1-
IHEI > El KacIVEI < 1) IEI
If so, I would classify it as Koesshi, otherwise I would classify it as Nanametsuji. However, El is a constant. The constant at the weak edge is E
Letting El be a constant at a strong edge, parameters E1 and El are determined so that El<El.

The shaded portions in FIG. 6(A) are the sequence components remaining without being masked according to each category. These are the fifth
It is shown in positional correspondence with each sequence component in the figure. In FIG. 6A, the sequence components in the shaded area are further divided into bands into low frequency components (shaded area) and high sequence components (shaded area). The numbers in parentheses are the number of dimensions when these are considered as low sequence vector elements and high sequence vector elements. In the flat category, the low sequence vector element Y
There are only three dimensions: 12+Y2+, Y22, and this is encoded into any of 1024 types (10 bits) of vector quantization code data. In categories other than flat areas, there are low sequence vector elements and high sequence vector elements, each of which is encoded into one of 256 different types (8 bits) of vector quantization code data. As shown in the figure, in this embodiment, the stronger the edge component, the greater the number of dimensions of the sequence vector. This is to preserve resolution. On the other hand, in flat areas, the number of gradations is increased and saved.

FIG. 6(B) is a diagram showing the bit allocation of scalar quantization for low sequence components and high sequence components defined for each category in the embodiment.

The total of each of the low sequence vector and high sequence vector after quantization is fixed at 18 bits. The number of vector dimensions is smaller for low-sequence vectors than for high-sequence vectors, and by increasing the number of bits allocated to sequence components with lower frequencies, priority is given to the gradation of low-frequency components of the image. encoding is performed.

FIG. 7 is a block diagram showing extracted vector quantization means of the embodiment. In the figure, the low sequence vector and high sequence vector divided into bands for each category by the scalar quantization unit 5 are manually processed into a vector quantization unit (L) 6 and a vector quantization unit (H) 7, each with 18 bits. Here, it is encoded into 8-bit vector quantization code data of any one of 256 types. At the same time, 3-bit category classification code data is manually input, and each vector quantizer 6.7 can be configured with an LUT that outputs 21-bit 8-bit quantized code data manually.

FIG. 8 is a diagram showing the operational concept of the integrated vector quantizer 8 of the embodiment. In the figure, 'L 1 .

L2. ..., L256 is vector quantization code data of a low sequence vector, Hl, H2゜...,
H256 is vector quantization code data of a high sequence vector. Therefore, the number of combinations of each vector quantization code data of these low sequence vectors and high sequence vectors is generally 256 ffl or more. However, considering the correlation of general images, there is a correlation between low sequence vectors and high sequence vectors, so it is normal that there are not all possible combinations, that is, 256x256=65536. Therefore, we perform training on sample images in advance to find possible combinations, such as those between the upper and middle rows of Figure 8, and under conditions such that the number of combinations falls within 1024 types (10 bits).
Further vector quantization is performed by grouping together those with similar distortions,
A new integrated (combination) vector quantization code (10 bits) from 01 to C1024 is created. The configuration that achieves this is to create a 10-bit integrated vector quantization code using a vector quantization method from each 8-bit low sequence vector quantization code data, high sequence vector quantization code data, and 3-bit category classification code data. It consists of LUTs that generate data.

In this way, when encoding an image that has been vector quantized for each category or band division, the obtained reproduction vector is encoded, but it is divided into bands, and each independently determined reproduction vector is encoded. In this case, when considering the correlation of images, combinations of reproduction vectors for each band that rarely appear will also be encoded. Therefore, in this embodiment, by considering the correlation and combining independently encoded signals for each band into one code, it is possible to obtain encoding efficiency that could not be obtained with the conventional method.

Next, a method of determining the reproduction vector in the embodiment will be explained.

In this embodiment, a plurality of sample images having different frequencies are scanned and read in advance, and an optimal reproduction vector is determined by the well-known LBG method for each category classification and each band division described in FIGS. 5 and 6. That is, a training sequence is performed in the same manner as described with reference to FIGS. 1 to 8 to obtain a predetermined code length (21 bits in this embodiment). The big difference is that there are no hardware restrictions on the training sequence. In reality, the optimal solution for the reproduction vector is determined by computer simulation. Therefore, after orthogonally transforming the sample image data, the image shown in Fig. 6 (
Figure 6 (
Sequence vector elements for each category as described in A) are extracted, divided into bands into low sequence vectors and high sequence vectors, and an optimal reproduction vector is determined by the LBP method. In this way, the scalar quantization error shown in FIG. 6(B) is prevented from affecting the optimum reproduction vector design.

In addition, when designing the optimal reproduction vector, the parameter TI for category classification based on the edge amount ED explained in FIG. 6(A)
, T2. T3 is determined as follows. T1 in FIG. 6(A). T2. Multiple sample images are used when determining the four classifications based on T3, and the number of 4×4 pixel blocks belonging to each is calculated as follows: When category flatness ED≦T1 B1 When category halftone dot flatness Tl<El)≦T2 When B2 When the category weak edge T2<ED≦T3 B3 When the category strong edge ED>T3 B4 Then, the parameters TI, T2. Determine T3. This method is used to prevent the Euclidean space of vectors from becoming abnormally large as edges become stronger when vector quantization is performed for each category. In other words, since the number of reproduction vectors for each category is fixed, it is possible to prevent the error between the input image and the reproduction vector from becoming too large as the edge becomes a category. This is for the purpose of preventing a reproduction vector that is extremely biased toward a flat area from being determined as the optimal solution because it appears frequently.

FIG. 9 is a flowchart showing how to obtain a reproduction vector in the embodiment. In the figure, in step Sl, the sample image is cut out into blocks of 4×4 pixels. In step S2, the block is orthogonally transformed (Hadamard transform in this embodiment).

In step S3, the sequence data Yll after orthogonal transformation is
~Y44 classifies each block into categories based on the edge amount ED. In step S4, each category is divided into bands as shown in FIG. 6(A), and low sequence vectors and high sequence vectors are extracted. Step S
In step 5, the steps from step 81 to step S4 are performed until all sample images (in this example, several representative images with different frequencies) have been scanned, and the low sequence vector and high sequence vector are scanned for each category. Create a table of all sequence vector sample vectors. Step S
6, for all the sample low sequence vectors and sample high sequence vectors obtained in steps 51 to S5, for each category, both the low sequence vector and the high sequence vector are 256 by the LBG method.
Find the optimal vector for the species. In step S7, for each category, the low sequence vector quantization code data and the high sequence vector quantization code data are integrated by the integrated vector quantization explained in FIG.
i's integrated vector quantization code data. The reproduction vector of this embodiment is determined by the above steps 31 to S7.

[Other Embodiments] In the above-mentioned embodiments, Hadamard transform is used as the frequency analysis means, but the present invention is not limited to this. Other orthogonal transformations such as cosine transformation and KL transformation may also be used.

[Effects of the Invention] As described above, according to the present invention, it is possible to eliminate the drawback of obtaining a reproduction vector whose appearance frequency is extremely biased in a training sequence.

Further, according to the present invention, when implementing hardware, LOT
This reduces the amount of data and facilitates hardware implementation.

Further, according to the present invention, efficient encoding can be performed using image correlation.

[Brief explanation of the drawing]

FIG. 1 is a block configuration diagram of an image information processing apparatus according to an embodiment. FIG. 2(A) is a diagram showing block image data XI, X2+, . . . , (B
) is the vector data (sequence component) after Hadamard transformation by the orthogonal transformation unit 3 Y+++Y12+ ..., Y
44, FIG. 3 is a diagram showing the two-dimensional Walsh type transformation pattern adopted in the Hadamard transformation of this embodiment, and FIG.
The figure shows the information structure of the block encoding code of the embodiment. Figure 5 (A) is a diagram explaining the basic concept of the category classification method of the embodiment. Figure 5 (B) and (C) are the weak edge FIG. 6(A) is a diagram specifically showing the category classification method of the embodiment, and FIG. 6(B) is a diagram showing the low sequence components and strong edge portions specified for each category of the embodiment. A diagram showing the bit allocation of scalar quantization for high sequence components, FIG. 7 is a block diagram extracting and showing the vector quantization means of the embodiment, and FIG. 8 is the operational concept of the integrated vector quantization unit 8 of the embodiment. FIG. 9 is a flowchart showing how to obtain a playback vector according to the embodiment. FIG. 10 is a diagram showing an example of band division by bands in conventional Schaming. In the figure, 1... Image input device, 2... Line buffer, 3... Orthogonal transform unit, 4... Category classification unit, 5...
... Scalar quantization section, 6... Vector quantization section (L)
,? . . . vector quantization section (H), 8 . . . integrated vector quantization section, 9 . . . code data storage section. Patent applicant: Canon Co., Ltd. (A) (B) Figure 2 Figure 3 Figure 5 Atsushi! i Archer/i1 Ashi (3,51 (3,6)
(3,51 Usee Ashi 4 Hikyu Goku Technique Te oblique Ushiki Technique Shizukoro No. 6
Figure (B)

Claims (3)

[Claims] (1) Blocking means for dividing digital image data into M×N pixel units; Transforming means for generating frequency-analyzed vector data of a predetermined number of dimensions by performing orthogonal transformation on the divided block image data; , a classification means for classifying the block image into one of a plurality of categories according to the magnitude of frequency based on a predetermined vector element of the generated vector data; Band division means for dividing the vector data into bands; Vector quantization means for independently vector quantizing the vector data divided into bands according to the classified categories; and the quantized vector data having been vector quantized independently. An image information processing device characterized by comprising an integrated quantization unit that performs integrated quantization into one vector per block by vector quantization according to the classified category. (2) The band dividing means performs band division so as to reduce the number of dimensions of the vector data of the low frequency band and increase the number of dimensions of the vector data of the high frequency band. image information processing device. (3) The image information processing apparatus according to claim 2, wherein the band division means performs band division of different band patterns depending on the category of the classification means.