JPH0998420A - Sub band encoding device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent a circuit scale from being enlarged, to simplify the constitution of a filter bank and to prevent the generation of block distortion and mosquito noise. SOLUTION: A/D converted picture data provided with three-dimensional information are stored in a 3-1 dimensional scanning conversion memory 12 in the order of being outputted from an A/D converter 11. Then, when the data are stored in the 3-1 dimensional scanning conversion memory 12 for four frames, a read address is supplied to the 3-1 dimensional scanning memory 12 so as to perform scanning for taking out all picture elements within a three- dimensional block as a continuous data string with the 64 picture elements of 4 horizontal direction picture elements × 4 lines × 4 frames as one three- dimensional block by an address generator 13. The image data converted into one-dimensional in this way are read by the 3-1 dimensional scanning conversion memory 12 are read and efficiently encoded by 16-18.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image data subband encoding apparatus for image transmission or recording.

[0002]

2. Description of the Related Art For example, conventional compression of image data having three-dimensional information is performed by motion compensation prediction and discrete cosine transform (DC).
T) was performed in combination. That is, the discrete cosine transform (DCT) is used to compress the two-dimensional spatial information amount in the horizontal and vertical directions of the image, and the motion compensation prediction is used to compress the temporal information amount. Data is compressed in the space.

A two-dimensional subband coding system can be used for compressing the spatial information amount. In subband encoding of an image signal having two-dimensional information, band division is performed independently in the horizontal direction and vertical direction of the image. When performing multi-stage division in each direction, it is necessary to perform alternate division one by one in the horizontal direction and the vertical direction. That is, after the division in the horizontal direction, the division in the vertical direction is performed, the division in the horizontal direction is performed in the low region in the vertical direction, and this division is repeated.

As an example, FIG. 9 shows a block diagram of a two-dimensional 10-band divided subband filter. As shown,
Each stage has the same configuration as each other, and has a high-pass filter (H
PF) 29, low-pass filter (LPF) 30, 2 to 1
It has a sub-sampler 31 and a scan conversion memory 32.

[0005]

When the spatial cosine transform is performed by using the discrete cosine transform, block distortion and mosquito noise-related distortion related to image quality are likely to occur.

When the two-dimensional sub-band coding method is used, the horizontal and vertical directions of the image are band-divided independently. Further, when performing multi-stage division in each direction, it is necessary to divide the band in the horizontal direction and the vertical direction alternately. Therefore, as shown in FIG. 9, a scan conversion memory 32 for performing scan conversion from horizontal to vertical or from vertical to horizontal is required between each filter of the filter bank and the filter of the next stage, and the configuration of the filter bank is complicated. Becomes

Further, both require a motion compensation prediction circuit, so that the circuit scale becomes large.

In view of the above problems, it is an object of the present invention to propose a subband coding apparatus in which the circuit scale does not increase, the filter bank configuration is simple, and there is no block distortion or mosquito noise. It is in.

[0009]

In order to solve the above problems, the present invention adopts subband coding in which distortion related to image quality such as block distortion and mosquito noise is less likely to occur. In addition, the coding efficiency of moving images is improved by adopting a three-dimensional block configuration without using a motion compensation circuit, which requires a large circuit scale. Further, in order to simplify the structure of the filter bank, the three-dimensional image signal is scan-converted to be converted into a one-dimensional signal.

That is, according to the first aspect of the present invention, there is provided a structure in which a predetermined number of frame images are formed and the depth of the time axis is 3
A three-dimensional image is divided into a plurality of three-dimensional blocks each having a fixed length in the horizontal direction, the vertical direction, and the time direction, and all the pixels in each of the three-dimensional blocks are scan-converted to make a continuous one-dimensional image signal. Scanning conversion means for converting the one-dimensional data string into a plurality of sub-band components by a one-dimensional multi-stage divided sub-band filter, and the plurality of divided sub-band components. Coding means for coding is provided.

According to a second aspect of the present invention, in the first aspect, in the one-dimensional multi-stage divided sub-band filter, each stage extracts the input one-dimensional data as a sub-band component divided into four. The filter is composed of four filters of a pass filter, a first band pass filter, a second band pass filter and a high pass filter, and the components extracted from the low pass filters of the respective stages except the final stage are added to the next stage. It is characterized in that it is supplied to a filter.

The invention according to claim 3 is the same as claim 1.
Or 2, the scanning is performed by the scanning conversion means so that the ^n- th stage filter of the one-dimensional multi-stage divided sub-band filter contains only the components within 2 ^(n-1) frames. It is characterized by

The invention according to claim 4 is the method according to claim 1, 2 or 3, wherein the dividing means divides each sub-band component extracted from the one-dimensional multi-stage divided sub-band filter into a plurality of data blocks. , In a fixed order other than the order related to the frequency of the band of each subband,
A means for changing the order of the plurality of data blocks, wherein the encoding means quantizes each of the reordered subband components, and performs run-length encoding,
It is characterized by having means for Huffman coding.

Further, the invention according to claim 5 is claim 1
Or 2, in the scan conversion in each of the three-dimensional blocks, the scan conversion unit scans four different pixels in the horizontal and vertical planes, and the scan for each of the four pixels is set as one area to set the horizontal time. It is characterized in that there is provided means for performing scanning for 16 pixels in four areas of the plane, and for performing scanning for the 16 pixels as one area, scanning for 64 pixels in four areas of the vertical time plane.

Further, the invention according to claim 6 is the invention according to claim 2.
In the above, each filter of each stage of the one-dimensional multi-stage divided sub-band filter has means for sub-sampling the output data to 1/4.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION According to the present invention, a three-dimensional image having a depth on the time axis, which is composed of a predetermined number of frame images, repeats scan conversion for extracting all the pixels in the three-dimensional block as a continuous data string. Is converted into a one-dimensional data string. Note that, at this time, the scan conversion is such that when it is subsequently divided into frequency components by the one-dimensional stage division sub-band filter, the output of the ^n- th stage division filter includes only the components within 2 ^(n-1) frames. Can be configured to.

The one-dimensionally converted data string is input to the one-dimensional multistage divided subband filter. In the one-dimensional multi-stage divided sub-band filter, each stage can have its basic configuration with four filters, a low-pass filter, a first band-pass filter, a second band-pass filter, and a high-pass filter. . Then, the plurality of stages configured by the four filters are configured such that the output of the low pass filter of each stage is supplied in parallel to the four filters of the next stage. In this way, when multi-stage sub-band division is performed, the output concentrates on some specific band components depending on the design, for example.

In each band component, when the intra-frame correlation is small and the inter-frame correlation is large, the components having a large absolute value of the filter output are continuously transmitted together, and the intra-frame correlation is large. When it is present and the inter-frame correlation is small, the components having a large absolute value of the filter output are continuously collected and sent to the quantizer. Here, when the intra-frame correlation is small and the inter-frame correlation is large, it means that a still image is input. In a general still image, the same data is continuous between frames, so that the correlation is very large, the resolution is higher in the frame than the moving image, and the correlation between pixels is small. In addition, when the intra-frame correlation is large and the inter-frame correlation is small, it means that a moving image is input. , The correlation becomes smaller, and the faster the movement within the frame, the lower the resolution and the larger the correlation.

Therefore, data having a small absolute value has a high probability of becoming 0 data, that is, invalid data by quantization, and rearranged and quantized data does not scatter valid data and invalid data in a data string. , And each becomes continuous, and efficient data compression can be performed by the run length code and the Huffman code.

An example of a subband coding apparatus according to the present invention will be described in detail below with reference to the drawings. FIG. 1 shows a configuration diagram of a one-dimensional division filter in a three-stage configuration, and FIG. 2 shows a schematic configuration diagram of a subband encoding device of the present invention.

In this example, the three-dimensional block is composed of 4 horizontal pixels × 4 lines × 4 frames (that is, 4 horizontal pixels × 4 vertical pixels = 16 pixels in one frame, and this 16 pixels × 4 frames = 64), and a three-dimensional block of 64 pixels (see FIGS. 3 to 4 described later).
Further, the one-dimensional multi-stage division filter is composed of a 4-tap digital filter for each band. It is assumed that the division filter has three stages of four bands and is divided into a total of ten band components. Here, as shown in FIG. 1, the component data of the 10 bands is set to LLL, LLB1, LLB2, LLH, in order from the low band side, that is, the output of the third stage low band filter.
Let LB1, LB2, LH, B1, B2, H.

As shown in FIG. 2, the image signal is input to the input terminal 1
It is input from 0 and is converted into digital data by the A / D converter 11. At this time, a general image signal has three-dimensional information because it is scanned in the horizontal direction, the vertical direction, and the time direction of the screen. This data is stored in the 3-1D scan conversion memory 12 in the order output from the A / D converter 11. Next, when the data for four frames is stored in the 3-1D scan conversion memory 12, the address generator 13 causes 4 horizontal pixels × 4 lines ×
64 pixels of 4 frames as one 3D block,
The 3-1D scan conversion memory 1 is used so that all pixels in the 3D block are scanned as a continuous data string.
2, the read address is given. As a result, the one-dimensionally converted image data is read from the 3-1D scan conversion memory 12.

The one-dimensional data string read from the 3-1D scan conversion memory 12 is input to the divided filter bank 14. As shown in FIG. 1, each of the divided filter banks 14 includes a low pass filter (LPF) 6, a first filter, and a first filter.
Band pass filter (BPF1) 5, second band pass filter (BPF2) 4, high pass filter (HPF) 3
The four filters of (1) are basically configured, and these are cascaded in three stages to form a one-dimensional multi-stage divided filter. That is, the output of the low pass filter 6 of each stage is converted into the low pass filter (LPF) 6, the first band pass filter (BPF1) 5, and the second band pass filter (BPF2) of the next stage.
4. The high pass filter (HPF) 3 is configured to be supplied in parallel. In addition, here, Z of each filter
The transfer function represented by the transformation was set as follows.

The transfer function represented by the Z transformation of the low pass filter 6 is

[0025]

[Equation 1] H ₀ (Z) = 1 × Z ⁰ + 1 × Z ⁻¹ + 1 × Z ⁻² + 1 × Z ^−3, and the transfer function represented by the Z conversion of the first bandpass filter 5 is

[0026]

[Expression 2] H ₁ (Z) = 1 × Z ⁰ + 1 × Z ⁻¹ −1 × Z ⁻² −1 × Z ^−3, and the transfer function represented by the Z conversion of the second bandpass filter 4 is

[0027]

[Equation 3] H ₂ (Z) = − 1 × Z ⁰ + 1 × Z ⁻¹ + 1 × Z ⁻² −1 × Z ^−3, and the transfer function represented by the Z conversion of the high-pass filter 3 is

[0028]

## EQU00004 ## H ₃ (Z) =-1 × Z ⁰ + 1 × Z ^-1 -1 × Z ^-2 + 1 × Z ^-3 . As a result, component outputs for 10 bands can be obtained. The output of each filter is 4: 1.
The subsampler 7 performs 4: 1 subsampling.

At this time, among the input images, the components corresponding to the image having a fast motion (the intra-frame correlation is large and the inter-frame correlation is small) are LLB2, LB1,
Large output concentrates on B1. Then, the components corresponding to the still image (the intra-frame correlation is small and the inter-frame correlation is large) have large outputs concentrated on LLH, LB2, and B2.

Since the filter output components have the above characteristics, the band components in which these large outputs are concentrated are put together. Then, here, the components are rearranged with an emphasis on the visual sense of an image having a fast motion. That is, all band components output from the one-dimensional multi-stage split filter are components LLL that commonly correspond to each other, and an image having fast motion (the intra-frame correlation is large and the inter-frame correlation is small).
Components LLB2, LB1, B1, components LLH, LB2, B2 corresponding to a still image (with a small intra-frame correlation and a large inter-frame correlation), and a component LLB1, corresponding to an intermediate between the two. Rearrange in the order of LH and H. This rearrangement is performed in the multiplexer 8.

The band components thus rearranged and taken out from the multiplexer 8 are quantized by the quantizer 16. Here, with respect to the components other than the component in which a large output is concentrated due to the input image, most of the data becomes 0 data by the quantizer 16. Also, LLL and LLB2
Regarding the band component data of LB1 and LB1, the correlation between the data of the same component of the preceding and following three-dimensional blocks is large. Therefore, it is assumed that the DPCM 15 transmits the difference value from the output value of the same component of the preceding three-dimensional block without performing the direct quantization. However, at this time, the true value is transmitted every 8 block intervals.

The data output from the quantizer 16 and the DPCM 15 is run-length encoded by the run-length encoder 17. The run-length coding is coding in which valid data and data of the number of 0 data are made into one set of pair data. Therefore, if the number of 0-data increases, the 0-data itself is not transmitted and the data compression efficiency increases. Since the band component data of LLL and LLB2 and LB1 are quantized by taking the difference value between the data having a large correlation, a large number of 0 data can be obtained. Therefore, the compression rate is increased here.

The run-length encoded pair data is
Huffman coding is performed by the Huffman encoder 18.
In the Huffman coding, a code having a shorter code length is sequentially assigned from data having a higher appearance probability in the entire data, and a code having a longer code length is assigned to the data having a lower appearance probability. As a result, a large amount of data having a short code length occupies a large amount of the whole data, so that the code amount becomes small as a whole and data compression can be performed.

The Huffman-encoded data is transmitted or recorded after being modulated according to the transmission line recording device 19 or the like. At the time of decoding, the Huffman decoder 20, the run-length decoder 21, the inverse quantizer 22, the inverse DPCM 23, the synthesis filter bank 24, and the 1-3D scan conversion memory 2
5, the address generator 26 and the D / A converter 27 perform the processings which are almost the reverse of the above, and the video signal is output from the output terminal 28.

An example of scanning for extracting all the pixels in the three-dimensional block in the 3-1D scan conversion memory 12 as a continuous data string will be described in detail with reference to the drawings. 3 to 4 show an example of scanning a three-dimensional block.

FIG. 3 shows the horizontal and vertical directions of the image signal.
An example of scanning in a three-dimensional block having a length of 4 pixels in the time direction is shown. The number of pixels in one block is 4
There are 64 pixels in the horizontal direction pixel × 4 lines × 4 frames. The numbers in the circle are sequentially scanned from the 0th pixel to the 63rd pixel.

FIG. 4A shows the divided filter bank 14
An example of scanning regarding the output of the first-stage division filter of the one-dimensional multi-stage division filter is shown. First, in the horizontal and vertical plane, four adjacent pixels of a square are made into one block, and scanning is performed so that each pixel passes only once. In this example, the lower left pixel is first scanned in the right direction, then in the upper direction, and finally in the left direction to perform U-shaped scanning for four pixels. As a result, the component output of the division filter of the first stage is obtained by 4 pixels (eg, 0 to 3 in FIG. 3) in the same frame.

FIG. 4B shows the divided filter bank 14
2 is a scan related to the filter output of the second stage of the one-dimensional multi-stage division filter, and each block is composed of the above-mentioned block composed of 4 pixels as one data, and the adjacent data in the horizontal time plane is compared with each data. Scan so that it only passes once. In this example, the scanning is first performed in the horizontal direction, then in the time direction, and finally in the horizontal direction. Ie 4 each
U-shaped scanning is performed on four data composed of pixels. As a result, 16 pixels in 2 frames (example: 0 to 0 in FIG. 3)
The component output of the division filter in the second stage is obtained from the block configured in 15).

FIG. 4C shows the division filter bank 14
Is a scan related to the filter output of the third stage of the one-dimensional multi-stage division filter of (1), and each block is composed of the above-mentioned block composed of 16 pixels as one data, and the adjacent data in the vertical time plane is compared with each data. Scan so that it only passes once. In this example, first, the scanning is performed in the vertical direction at the same time as the scanning in the time axis direction, then in the vertical direction, and finally in the vertical direction, and at the same time, the scanning is performed in the time axis direction. As a result, the component output of the division filter in the third stage can be obtained from the block composed of 64 pixels (for example, 0 to 63 in FIG. 3) in four frames. FIG. 5 shows the relationship between the three-dimensional block shown in FIGS. 3 and 4 and the component output of the division filter.

As described above, the divided filter bank 1
10 band component outputs (LLL to H:
(See FIG. 1)
This will be described in detail with reference to Tables 1 to 3 shown in FIGS.

The components are arranged in order from the low frequency side so that the components of 10 bands are normally output, and the result is as follows. Here, the underlined components represent components where large outputs are concentrated.

For images with fast motion LLL , LLB1, LLB2 , LLH, LB1 , LB2, LH, B1 , B2, H For still images LLL , LLB1, LLB2, LLH , LB1, LB2 , LH, B1 , B2 , H Next, rearranging the component outputs of the 10 bands is as follows. Here, similarly, the underlined components represent components in which a large output is concentrated.

For images with fast motion LLL, LLB2, LB1, B1 , LLH, LB2, B2, LLB1, LH, H For still images LLL , LLB2, LB1, B1, LLH, LB2, B2 , LLB1 , LH, H Quantizer 1 except for the underlined components
There is a high probability that 0 will result in 0 data. For this reason, if the rearranged data of 10 bands is treated as one data string, the probability of continuous 0 data increases.

Next, an example of the case of performing run-length coding in the above case for an image having a fast motion will be described. At this time, it is assumed that all the data except the underlined components has become 0 data by the quantizer 16.
The pair data obtained by run-length coding before and after rearrangement of component outputs is as follows.

Before rearrangement (0, LLL) (1, LLB2) (1, LB1) (0, LB1)… (0, LB1) (8, B1) (0, B
1)… (0, B1) (E0B) After rearrangement (0, LLL) (0, LLB2) (0, LB1) (0, LB1)… (0, LB1) (0, B1) (0, B
1)… (0, B1) (E0B) Comparing the paired data before and after rearrangement, there is no change in the number of paired data, but in the three paired data,
The number of 0 data is different (increasing). Next, the code lengths of the pair data when Huffman coded are compared from Tables 1 to 3 as follows.
When comparing code lengths, for the same pair data,
Since the code length does not change, the difference between the code lengths is obtained only for the data whose number of 0 data has changed.

Code length of (1, LLB2) is 2 to 6 bits longer than (0, LLB2) Code length of (1, LB1) is 2 to 6 bits longer than (0, LB1) Code of (8, B1) The length is 6 to 10 bits longer than (0, B1). Therefore, if the component output is quantized, run-length encoded, and Huffman encoded, the rearranged component outputs can be encoded 10 to 22 bits shorter. .

Similarly, an example of performing run-length coding in the above case on a still image will be described. At this time, it is assumed that all the data except the underlined components has become 0 data by the quantizer 16. The pair data obtained by run-length coding before and after rearrangement of component outputs is as follows.

Before sorting (0, LLL)(2, LLH) (4, LB2)(0, LB2) ... (0, LB2) (15,0)(4, B2)
(0, B2)… (0, B2) (E0B) After rearrangement (0, LLL) (15,0)(5, LLH) (0, LB2)(0, LB2) ... (0, LB2)(0, B2)
(0, B2)… (0, B2) (E0B) Comparing paired data before and after rearrangement,
There is no change in the body, but in the three pairs of data, 0 data
The number of data is changing. Next, each pair data
Table 1 to Table 3 show the code lengths when Huffman coding is performed.
The comparison is as follows. When comparing code lengths
Code length does not change for the same pair data
So, only for those that have changed in the number of 0 data,
Find the difference in code length.

The code length of (2, LLH) is 1 to 3 bits shorter than (5, LLH), and the code length of (4, LB2) is 4 to 8 bits longer than (0, LB2). The length is 4 to 8 bits longer than (0, B2). Therefore, if the component output is quantized, run-length encoded, and Huffman encoded, the rearranged component outputs can be encoded 5 to 15 bits shorter. .

For the above reasons, by rearranging the component outputs of the 10 bands, the code length in a block of 4 horizontal pixels × 4 lines × 4 frames can be shortened at the time of Huffman coding, and the compression efficiency can be reduced. Can be improved.

[0051]

According to the present invention, scanning conversion is performed to perform encoding including the time axis direction. Therefore, it is not necessary to select and perform suitable encoding for each of the still image portion and the moving image portion of the image. For example, it is not necessary to perform motion detection and perform inter-frame processing or coding based on inter-frame difference on a moving part. Also, since the properties of the intra-frame correlation and the inter-frame correlation of the input image appear in the encoded data, efficient encoding can be performed by using this to perform the rearrangement process. Further, by performing 1/4 sub-sampling using a 4-division filter, it is possible to realize an efficient encoding device that prevents an increase in encoded data.

[Brief description of drawings]

FIG. 1 is a block diagram of a divided filter bank.

FIG. 2 is a block diagram showing an example of a subband coding device according to the present invention.

FIG. 3 is a diagram showing an example of scanning within a three-dimensional block of 4 horizontal pixels × 4 lines × 4 frames.

FIG. 4 is a diagram showing an example of scanning regarding a divided filter output corresponding to each stage of a one-dimensional multi-stage divided subband filter of a divided filter bank.

FIG. 5 is a diagram showing an example of a relationship between a 3D block of 4 horizontal pixels × 4 lines × 4 frames and division filter output.

FIG. 6 is a diagram including Table 1 showing a code length list.

FIG. 7 is a table 2 showing the difference in code length due to the difference in the number of 0 data.
FIG.

FIG. 8 is a table 3 showing the difference in code length due to the difference in the number of 0 data.
FIG.

FIG. 9 is a block diagram of a two-dimensional 10-band split subband filter according to the related art.

[Explanation of symbols]

 1,10,33 Input terminal Two division filter 3,29 High pass filter 4 Second band pass filter 5 First band pass filter 6,30 Low pass filter 7 4 to 1 subsampler 8 Multiplexer 9,28 Output terminal 11 A / D converter 12 3-1 dimensional scan conversion memory 13, 26 Address generator 14 Divided filter bank 15 DPCM 16 Quantizer 17 Run length encoder 18 Huffman encoder 19 Transmission line recording device 20 Huffman decoding Converter 21 run-length decoder 22 inverse quantizer 23 inverse DPCM 24 synthesis filter bank 25 1-3 dimensional scan conversion memory 27 D / A converter 31 2 to 1 subsampler 32 scan conversion memory

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Mitsuharu Sato 5-6-16 Ikegami, Ota-ku, Tokyo Inside Ikegami Tsushinoki Co., Ltd.

Claims (6)

[Claims] 1. A three-dimensional image composed of a predetermined number of frame images and having a depth on a time axis is divided into a plurality of three-dimensional blocks each having a constant length in the horizontal direction, the vertical direction, and the time direction, and Scan conversion means for scanning-converting all pixels in a three-dimensional block to convert the image signal into a continuous one-dimensional data string, and a plurality of one-dimensional data strings by a one-dimensional multi-stage divided sub-band filter. A subband coding apparatus comprising: a dividing unit that divides into subband components; and an encoding unit that encodes the divided plurality of subband components. 2. The low-pass filter, first band, according to claim 1, wherein each of the one-dimensional multi-stage divided sub-band filters extracts input one-dimensional data as four-divided sub-band components. A sub-band code that is composed of four filters, a pass filter, a second band pass filter, and a high pass filter, and that supplies the components extracted from the low pass filter of each stage except the final stage to each filter of the next stage. Device. 3. The scan in the scan conversion means according to claim 1, wherein the n-th stage filter of the one-dimensional multi-stage divided sub-band filter includes only components in 2 ^(n-1) frames. The sub-band coding device is characterized in that 4. The dividing means according to claim 1, 2 or 3, wherein each of the sub-band components extracted from the one-dimensional multi-stage sub-band filter is divided into a plurality of data blocks, and the sub-band of each sub-band is divided. A unit for changing the order of the plurality of data blocks in a fixed order other than the order related to the frequency, wherein the encoding unit quantizes each of the rearranged subband components to obtain a run length. A subband coding device having means for coding and Huffman coding. 5. The scan conversion means according to claim 1, wherein in the scan conversion in each of the three-dimensional blocks, scanning is performed on four different pixels in a horizontal and vertical plane, and scanning is performed on the four pixels. As a region for scanning 16 pixels in four regions in the horizontal time plane, and scanning for 16 pixels as one region, and scanning for 64 pixels in four regions in the vertical time plane. Subband encoder. 6. The subband coding apparatus according to claim 2, wherein each filter of each stage of the one-dimensional multistage divided subband filter has means for subsampling output data to ¼.