JPH02137475A - Image data compression and recovery circuit - Google Patents

Abstract

PURPOSE:To compress image data with high efficiency as keeping superior picture quality by performing the encoding of the position of a contour line at a 0th hierarchy where the maximum encoding quantity can be obtained expressing in a binary value in pyramid encoding. CONSTITUTION:On a transmission side, data Lo at the 0th hierarchy is quantized at a quantizer 101(1), and after that, it is reformed to a diagram at a binarization means 15, and is compression-encoded at an encoder means 16. A compression code at each hierarchy is transmitted to a reception side. On the reception side, a received compression code is decoded by a decoder 110 at every hierarchy. At the 0th hierarchy, the compression code(binarized one) is decoded by a decoder means 23, and is reproduced to the diagram. Next, a quantization level is recovered by a quantization level recovery means 26 based on a result in which the change direction of gradation of a high-order hierarchy is detected from the diagram by a change direction detecting means 27.

Description

[Detailed Description of the Invention] [Summary] Regarding an image data compression/restoration circuit that compresses and restores grayscale image data with high efficiency, when compressing multilevel halftone image data, the compressed image data is restored. In addition to obtaining high compression ratio when
In order to maintain the high quality of the restored image, the multilevel image data is divided into multiple spatial frequency bands.
When encoding an image in the highest frequency band in the spatial domain, a binarizing means detects a point where the gradation level crosses the 0 level and expresses the position as a binary image; an encoding means for encoding an output binary image on the transmission side; a change direction detection means for detecting a direction in which the gradation level changes from an image in the next highest frequency band when restoring the binary image; a decoding means for decoding a binary image; and a quantization level for arranging tone changes according to the detected tone change direction around the point where the tone level of the binary image decoded by the decoding means crosses the 0 level. A restoring means is provided on each receiving side.

[Field of Industrial Application] The present invention relates to an image data compression/restoration circuit that compresses and restores grayscale image data with high efficiency.

[Prior art] Pyramid coding is a highly efficient compression method for multi-level halftone images + IEEE 1981 Pattern
Recognition & Image Proc
essing, P218-P223 (1983)'
Image data compression
with the Laplattan Py
ramid'' (E, H, Adelson, P, J
, Burt) l. The principle of this pyramid encoding will be explained below.

FIG. 8 is an explanatory diagram of how image data is encoded by pyramid encoding, and FIG. 9 is an explanatory diagram of how image data is restored by pyramid encoding. The correlation between the pixels of the image is
This is achieved by subtracting a low-pass filtered image from that image (bandpass). At this time, the low-pass filtered image, that is, the low-pass image, is expressed with a lower resolution than the original image, and the band-pass processed image (band-pass image) has a smaller variance than the original image.

Therefore, by sequentially applying a low-pass filter with a low cutoff frequency to the image to reduce its resolution and encoding each band-pass image, it is possible to sequentially remove everything from local correlation to global correlation.

In this way, the low-pass image and the band-pass image become data of a pyramid-shaped hierarchical structure (0th floor to Nth floor) as shown in FIGS. 8 and 9 in order to sequentially reduce the resolution.

In pyramid encoding, the pyramid-shaped hierarchical data of a band-pass image is encoded, and this encoded data is used to sequentially restore the pyramid-shaped hierarchical data of a low-pass image from the top to the bottom, and finally the original image of the bottom is It is intended to restore. The method in the above-mentioned literature for such processing converts the pyramidal hierarchical data of the low-pass image and the pyramidal hierarchical data of the bandpass image into a Gaussian pyramid (Fig. 8 (a)) and a Laplacian pyramid (Fig. 8 (a)), respectively. It is called b)).

The operation of creating the first layer Li of the Gaussian pyramid and the first layer Li of the Laplacian pyramid is expressed as follows.

G,,,-REDUCE [Gil (1) L+
+H-G i EXPAND [C+++ 1-Gi
-EXPAND [REDUCE [Gil] Here, REDUCE is an operation that applies a low-pass filter to reduce the resolution and creates an upper layer of the Gaussian pyramid, and EXPAND is an operation that expands the Gaussian pyramid to the same size as the lower layer. refers to the operation to do. REDU
The operations of CE and EXPAND are shown by the following equations, respectively.

(1) REDUCE operation G# (i, j) 1 Σ Σ W (m, n) ・Gs−+ (2j+m+ 2j+n)
(3) Here, W is a weight function, and spatial convolution (a method of Convolution H filter calculation) is performed over mask sizes m and n. This low-pass filter, resolution reduction convolution is
5 in Figure 10 (B) as shown in Figure 10 as one-dimensional
x5 (odd number type) or 4 x 4 (even number type) as shown in FIG. 10(a). Note that a, b, and c in FIG. 10 are convoluted coefficients, which determine the strength of the low-pass filter.

(2) EXPAND operation j -4Σ ΣW (m, n) ・Gz-+ ((i+m)/2. (j+n)/2
) EXPAND expands the image size by interpolation.

In the Gaussian pyramid, the higher the layer, the lower the cutoff frequency of the image becomes, so the lower the layer of the Laplacian pyramid as a difference, the more high-frequency spatial frequencies are applied to the image. . From the perspective of visual characteristics, images with gradual gradation changes require coarse sampling but fine quantization; conversely, images containing many fine parts require coarse quantization. , it is important to sample in detail.

Pyramid encoding takes advantage of this visual characteristic. In other words, the number of samples is larger in the layer with more high frequency components, and the amount of data can be reduced by quantizing more coarsely in the lower layer.

The prior art in the above-mentioned literature separates the 0th layer and the 1st layer and above as follows, and uniformly (equally) quantizes at 3 levels and 5 levels, respectively.

If Cj (i, j) is the quantization result of LJ (11J), then Co (1+ J) Co ct, j) where 0-J≦N (N is the highest layer).

At this time, if the number of pixels in the entire image is N, the number of samples in the first layer and above is N/3, so the amount of data is INI 0g23+ (N/3) ・fl og25
1/N was approximately 2.36 bits/cell. The above document states that further compression can be achieved by performing variable length encoding after quantization.

FIG. 11 is a block diagram showing an example of the configuration of an image data compression/decompression circuit using the prior art described above. The figure shows the sending side (a) that compresses the input image data and sends it out.
It consists of a block on the receiving side (b) that decompresses the compressed image data.

Image data is input from the input terminal 1 and stored in the memory 1 (1) in the hierarchical data creation means 100 (1) for one screen. This original image becomes the 0th layer of the Gaussian pyramid. Images stored in memory (1) are REDUCE
The above-mentioned REDUCE operation is performed in the unit 2(1) and stored in the memory 1(2) in the hierarchical data creation means 100(2). Images stored in memory 1 (2) are stored in memory 1 (
The size is 1/4 of the image stored in 1), and becomes the first layer of the Gaussian pyramid.

Next, the image in the memory 1 (2) is stored in the hierarchical data creation means 10.
0 In the EXPAND section 3 (1) in (1), the above EXP
An AND operation is performed and the result is input to the subtraction circuit 4(1).

The output of the memory 1 (1) is also input to this subtraction circuit 4 (1), and a bandpass image is obtained by subtracting the output of the EXPAND section 3 (1) from the output of the memory 1 (1). 5(1).

The image stored in the memory 5(1) becomes the 0th layer of the Laplacian pyramid. The image in the memory 5(1) is read out, quantized by a quantizer 101(1), encoded by an encoder 102(1), and output as a compressed code. Note that the layer data creation means 100 for the first layer or higher
(2) to 100(n) are also the 0th layer data creation means 10
0 (1) has approximately the same configuration and performs the same processing. In this case, memories 1(2) to 1(n) and memories 5(2) to
The size of 5(n) is sequentially reduced to 1/4. In the above method,
Hierarchical data creation means 100(2) to 10 for the first layer or higher
0 (n), pyramid data for each layer is created, quantized and encoded, and a compressed code is output.

Next, the operation on the receiving side (b) will be explained.

The compressed data of the 0th layer is input to the decoder 110(1) and stored in the memory 6(1) in the layered data restoring means 111(1) as pyramid data of the 0th layer. Next, the data read from the memory 6(1) is added to the adder circuit 8(1).
) is sent. The restored Gaussian pyramid data of the first layer is stored in the memory 9(2) in the adder circuit 8(1).
From there, an EXPAND operation is performed and input via the EXPAND section 7(1). As a result, the adder circuit 8(1)
Gaussian pyramid data of the 0th layer (
A restored image) is obtained, stored in the memory 9(1), and then output to the output terminal 2.

Here, the hierarchical data restoration means 111 (2
) to 111(n) have almost the same configuration, and by adding the Gaussian pyramid data (restored image) obtained in the upper layer to the decoded pyramid layer data, the Gaussian pyramid data of that layer is (Restored image)

[Problems to be Solved by the Invention] In pyramid coding as described above, data compression is performed by quantization and variable length coding, but the amount of data in the Laplacian pyramid is approximately 1.5 times that of the original image. Therefore, there was a problem that a high compression ratio could not be obtained. For this reason, methods have recently been adopted to improve the compression rate by combining pyramid coding and vector quantization (for example, Saito et al., "Laplacian P
yramld Coding Performance Improvement”, 1986
Annual Image Coding Symposium 3.6). however,
Since the lower layer of the Laplacian pyramid represents the outline of an image, if vector quantization is performed on a plurality of pixels at once, there is a problem that the outlines will not match between blocks.

The present invention has been made in view of the above-mentioned problems, and when compressing multilevel halftone image data, it is possible to obtain a high compression rate when restoring the compressed image data, and also to obtain a high compression ratio when restoring the compressed image data, and to The purpose of this invention is to provide an image data compression/restoration circuit that can maintain high quality.

[Means for Solving the Problems] FIG. 1 is a block diagram of the principle of the present invention. Components that are the same as those in FIG. 11 are designated by the same reference numerals.

In the figure, 15 is the point at which the gradation level crosses the 0 level when multilevel image data is divided into multiple spatial frequency bands and the image of the highest frequency band is encoded in the spatial domain. The output of the quantizer 101 (1) is received and binarized by the binarizing means which detects the position and expresses the position as a binary image.

Reference numeral 16 denotes an encoding means for compressing and encoding the output of the binarizing means 15. The structure of the layers above the first layer on the transmitting side is the same as in FIG. 16 is an encoding means for encoding the binary image output from the binarizing means 15;

27 is change direction detection means for detecting the direction in which the gradation level changes from the image of the next highest frequency band when restoring the 21ii image; 23 is a decoding means for decoding the binary image; 26 is decoding by the decoding means 23. This is a quantization level restoring means that arranges gradation changes around the point where the gradation level of the binary image crosses the 0 level according to the detected gradation change direction. The layers above the first layer on the receiving side are the same as in FIG.

[Operation] The 0th layer has the largest amount of data in the Laplacian pyramid, and is equivalent to the amount of data in the original image. The 0th layer is the layer containing the highest frequency component, and represents the outline of the image. FIG. 2 is an explanatory diagram of the operation of the present invention,
It is a figure which shows the gradation change of an outline part. Figure (a) is the original picture G. (b) shows the gradation change of the low frequency component (Gaussian image G+ of the first layer). The gradation change in the 0th layer of the Laplacian pyramid swings from negative to positive in the contour area as the gradation rises around the edge of the contour area, as shown in FIG. 2(c). this,
Figure 2 (d
), it appears only on the edges, and has the characteristic that the edges look like line figures. Here, the quantization levels when the decision levels are -T, O, +T are -A, 0, . +A.

The present invention utilizes the feature that edges at a lower level of hierarchy become almost line-like. In the present invention, zero-crossing points of three quantized levels of the 0th layer Laplacian image L0 are detected, the positions thereof are expressed as a two-level image (binary image), and a binary image encoding method is applied. Accordingly, a high compression rate can be obtained at the 0th layer. When decrypting,
The direction of tone change is detected with reference to the upper layer, and quantization levels of +A and -A are applied along the zero-crossing points of the decoded binary image to restore the contour.

In FIG. 1, 0th layer data L is transmitted on the transmitting side. After being quantized by the quantizer 101 (1), it is converted into a line figure by the binarization means 15, and compressed and encoded by the encoding means 16. The compression code of each layer is transmitted to the receiving side. Although not shown in the drawings, it is common to switch multiplexers and sequentially (for example, from the nth layer to the 0th layer) transmit the signal to the receiving side, and then use a distributor to distribute the signal to each layer on the receiving side.

On the receiving side, the received compressed code is sent to the decoder 110 of each layer.
decrypted with . At the 0th layer, the compressed code (binarized) is decoded by the decoding means 23 to reproduce the line figure.

Next, the quantization level is restored by the quantization level restoring means 26 based on the result of detecting the direction of change in the gradation of the upper layer from this line figure by the change direction detecting means 27. Thereafter, similarly to the conventional example, the restored data of each layer is sequentially added to create a restored image.

[Example] Hereinafter, an example of the present invention will be described in detail with reference to the drawings. Here, as an example, the present invention shown in FIG. 1 is different from the prior art shown in FIG. 11 in 21ii! quantization means 15. quantization level restoration means 26. The configuration and operation of the change direction detection means 27 will be described. The encoding means 16.degree. and the decoding means 23 are binary image data compression and restoration circuits, and can be realized by known techniques such as predictive encoding, so a description thereof will be omitted.

FIG. 3 is a block diagram showing a specific example of the configuration of the binarization means 15. The terminal 15-0 receives signals quantized into three levels (-1, 0,
The Laplacian image of the 0th layer (indicating each level with +1) is input. 15-1゜15-2.15-4.15-5
is a register and is combined with line memory 15-3 to create a 2×2
A window for pixels is formed. Register 15-1 becomes the pixel of interest.

The values of these 2×2 pixel windows are determined by the contour detector 15-
6 will be man-powered. The contour detector 15-6 is
This is a circuit for detecting the point where the gradation level crosses zero in the contour shown in the figure, and when it detects any of the patterns shown in Figure 4, it outputs "1" to terminal 15-7. However, in other cases, "0" is output. Figure 4 (a
) shows an example of crossing zero in the horizontal direction, (b) in the vertical direction, and (c) in the diagonal direction.

FIG. 5 is a block diagram showing a specific example of the configuration of the changing direction detecting means 27. As shown in FIG. A restored first layer Gaussian image enlarged to the same size as the original image is input from the terminal 27-1. Register 27-2.27-3.27-4
．． 27-5 and line memory 27-6 form a 2×2 pixel window, and register 27-2 is the pixel of interest.

Two outputs from each register are input to each of horizontal, diagonal, and vertical detectors 27-7.27-8.27-9. Each detector 27-7.27-8.27-9 has two
The difference between the two registers is determined, and the determined difference is input to the direction determiner 27-10.

The direction determiner 27-10 compares the input difference with a predetermined threshold and checks whether the change is significant. If the input difference is greater than the threshold value, the direction of change is selected from among the largest difference values from ① to ② in Fig. 6, and its number is output. Further, if all the input differences are equal to or less than the threshold value, the number 0 is output. These outputs are output to the outside from the terminal 27-11.

FIG. 7 is a block diagram showing a specific example of the configuration of the quantization level restoring means 26. A restored binary image representing a contour line is input from the decoding means 23 of FIG. 1 through the terminal 26-1. Registers 26-3 to 26-6 and line memory 26-
7 constitutes a 2×2 window. The value of this window and the number representing the change direction detected by the change direction detection means 27 from the terminal 26-2 are determined by the quantization level write control circuit 2.
6-8.

The circuit in the lower half of the figure has three levels of quantization values (-1, 0
, +1). register 26
-9 to 26-12 are 2-bit registers that can be set to one of three levels (-1, O, +1), and are used together with 2-bit/word line memory 26-13.
A window of ×2 pixels is constructed. The quantization level writing control circuit 26-8 writes registers 26-9 to 26-12 based on the input binary image (contour line) and information on the direction of change in gradation.
A 3-level (-1, 0, +1) set command is issued to . The quantization level writing circuit 26-8 issues a set command to the registers 26-9 to 26-12 when the target pixel 26-3 of the binary image is '1'', that is, on the contour line, and writes data on both sides of the contour line. -1 and +1 are selected and placed.Then, the final result is the terminal 26-1
They are output sequentially from 4 onwards.

In the above embodiment, the binary image of the 0th layer Laplacian image is predictively encoded.
The present invention is not limited to this, and a chain code, which is a known technique, can also be used by thinning the binarized image into a one-dot line.

[Effects of the Invention] As described in detail above, according to the present invention, in pyramid coding, the position of the contour line of the 0th layer where the amount of code is the largest is represented in binary and encoded, so that good performance is achieved. Image data can be compressed with high efficiency while maintaining high image quality.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of the principle of the present invention, Fig. 2 is a diagram showing gradation changes in the contour portion, Fig. 3 is a block diagram showing a specific configuration example of the binarization means, and Fig. 4 is a block diagram of the input pattern. An explanatory diagram of binarization, FIG. 5 is a block diagram showing a specific configuration example of a change direction detection circuit, FIG. 6 is a diagram showing pattern classification of change direction, and FIG. 7 is a concrete diagram of a quantization level restoring means. A block diagram showing a configuration example. Figure 8 is an explanatory diagram of the encoding status of image data by pyramid encoding. Figure 9 is an explanatory diagram of the restoration status of image data by pyramid encoding. Figure 10 is the first order of REDUCE operation. FIG. 11 is a block diagram showing an example of a conventional circuit configuration. In FIG. 1, 15 is a binarization means, 16 is an encoding means, 23 is a decoding means, 26 is a quantization level restoration means, 27 is a change direction detection means, 0 is a hierarchical data creation means, 1 is a quantizer , 2 is an encoding means, 0 is a decoder, and 1 is a hierarchical data restoring means. .

Claims (1)

[Claims] Multilevel image data is divided into multiple spatial frequency bands,
Binarization means (15) for detecting a point where the gradation level crosses the 0 level when encoding an image of the highest frequency band in the spatial domain and representing the position as a binary image; The encoding means (16) for encoding the binary image outputted from the encoding means (15) is connected to the transmitting side, and the direction in which the gradation level changes from the image in the next band of the highest frequency is determined when the binary image is restored. A change direction detecting means (27) for detecting a change direction, a decoding means (23) for decoding a binary image, and a change direction detecting means (27) for decoding a binary image; and a change direction detecting means (27) for decoding a binary image; , and quantization level restoring means (26) for arranging gradation changes according to the detected gradation change direction, respectively, on the receiving side.