KR0178900B1 - A dct encoder using subsampling method - Google Patents

Abstract

The present invention relates to a discrete cosine transform encoder using subsampling, wherein an interpolator 18 is additionally installed in the subsampling unit 10 of the encoder portion, and the original input image signal, the subsampling unit 10 and the interpolator ( 18, i.e., the discrete pixel information signal that may occur during interpolation, that is, the discrete pixel cosine transform (DCT), is quantized, encoded, and transmitted separately from the subsampled myth, thereby interpolating in the decoder part. In this case, there is an advantage in restoring an image close to the original image by using the lost pixel information signal.

Description

Discrete Cosine Transform Coder Using Subsampling

1 is a diagram for a sub-sampling scheme of HD-MAC,

1a is a subsampling diagram for each HD-MAC luminance signal mode;

Figure 1b is a two-dimensional filter pass band display for overlapping prevention for each mode according to the HD-MAC luminance signal sample pattern.

2 is a diagram of a sub-sampling scheme of MUSE,

2a is a display diagram showing a field offset subsampling method of MUSE;

2b is a flowchart of a luminance signal subsampling system of MUSE;

Figure 2c is a subsampling schematic of MUSE.

3A is a diagram showing all pixel signals of an original image;

FIG. 3B is a diagram of primary sub-sampling of the image signal of FIG. 3A;

FIG. 3C is a second sub-sample of the image signal of FIG. 3B;

3d is a diagram showing the characteristics of a lozenge filter for subsampling.

4a is a block diagram showing transform coding;

4b illustrates the principle of discrete cosine transform.

5 is a partial block diagram of an encoder according to the present invention.

* Explanation of symbols for main parts of the drawings

10: sub-sampler 12: first discrete cosine converter

14: first quantizer 16: the first encoder

18: interpolator 20: subtractor

22: second discrete cosine converter 24: second quantizer

26: second encoder

The present invention relates to a discrete cosine transform coder using subsampling, and in particular, an interpolator is additionally installed in a subsampling part of an encoder part so that a difference between an original input image signal and a signal passing through a subsampling part and an interpolator, i.e., occurs during interpolation. By performing discrete cosine transform (DCT) processing and then quantizing, encoding, and transmitting the missing pixel information signal separately from the sub-sampled signal, the original image using the missing pixel information signal transmitted when interpolation is performed in the decoder part. An encoder for restoring an image close to the

High-definition television (HDTV) has more than twice the vertical and horizontal resolution of current televisions, improves the problems of current televisions such as cross color, and has an aspect ratio of 16: 9 and CD sound quality. By having digital audio performance, we say next-generation TV which provides clear picture and clean sound quality more than current TV.

HDTV systems can be classified into two types based on the transmission method.

The first is a sampled value transmission method in which digital signals are regarded as analog signals and are transmitted by multiple sub-nyquist sampling encoding (MUSE) in Japan and high definition-multiplex analog components (HD-MAC) in Europe. Yes. The second is a full-digital system that digitalizes from signal processing to transmission, and is rapidly reaching the stage of practical application, starting with the United States.

Japan's MUSE and Europe's HD-MAC compress video bands by subsampling and time division multiplexing (TDM) and transmit them in the form of analog signals. On the other hand, in the full-digital method, the deterioration becomes large. On the receiving side, even if a channel error occurs as long as the signal-to-noise ratio is maintained above a certain threshold, even if a channel error occurs, the image quality of the transmitting side is almost completely transmitted to the receiving side.

Since HDTV signals have a huge amount of data, the need arises to compress the data for transmission. There are various compression methods, but the band coding method, which has recently been of great interest, is divided into a plurality of consecutive bands by using a plurality of filters in a frequency band, and subsampling for each signal series ( Subsampling), encoding, and transmission method.

The sub-sampling method is based on HD-MAC in Europe and MUSE in Japan among HDTV systems.

FIG. 1 is a diagram of a sub-sampling method of HD-MAC. There are three different spatial sub-sampling methods in HD-MAC. That is, there are subsampling methods used in the 80ms mode, the 40ms mode, and the 20ms mode, respectively. The refresh interval is 80ms in the stop mode (speed range: 0-0.5sample / frame), the refresh interval is 40ms in the slow movement mode (0.5-12sample / frame), and the refresh interval in the move mode (12 or more samples / frame). 20 ms.

FIG. 1A is a subsampled luminance signal in a high quality sample grid, and the displayed digitized points are sampled values of the digitized fields. (The size of the block is 16 * 16, but here only 8 * 8 is shown.)

In 80ms mode, samples of positions 1, 2, 3 and 4 are written for 80 ms (4 fields). In 80ms mode, it is assumed that there is no motion for 80ms, so samples may only be taken in 2 fields (one frame) of 4 fields (two frames). However, when transmitting a video signal to D / D2-MAC, only 8 * 4 samples (32 samples) can be sent per block, so samples at one position are samples at positions 2, 3, and 4 respectively in the first field. Send on the second, third, and fourth fields.

Since 32 samples must belong to an 8 * 4 block, the samples of lines 3, 7, 11 and 15 are sandwiched between the samples of lines 1, 5, 9 and 13, respectively, and the samples of lines 4, 6, 10 and 14 The transmission is sandwiched between them, which is called shuffling. The first line is defined as line 1. In the decoder, samples (128 samples) sent during 4 fields are collected and interpolated into 16 * 16 blocks.

In 40ms mode, samples of positions 1 and 2 are written for 40ms (2 fields). In FIG. 1A, since the odd-numbered lines belong to the odd field and the even lines belong to the even field, only the samples of the odd field are transmitted in the 40 ms mode. However, when transmitting a video signal to D / D2-MAC, only 8 * 4 samples (32 samples) can be sent per block, so samples at one position are placed in the odd field and samples at the second position are placed in the even field. Send it. Here, the samples are sampled in an 8 * 4 structure, so there is no need for shuffling. In the decoder, samples (64 samples) sent during two fields are collected and interpolated into 16 * 8 blocks to form a 40ms mode block in an odd field, and a 40ms block in an even field is filled by a motion compensation method.

In 20ms mode, samples of one position are transmitted in the odd field and samples of two positions in the even field for 20ms (one field).

Here, samples are sampled in a 4 * 8 structure, so shuffling in 8 * 4 form is necessary. The decoder collects 32 samples sent during one field and interpolates 16x8 blocks to form 20ms mode blocks in odd or even fields.

The subsampling of each mode is spatial deduction, and two-dimensional low pass filtering should be performed to prevent aliasing.

Figure 1b shows the passband of the two-dimensional low-pass filter for each mode. Since all three modes of sub-sampling are basically diagonal sampling (Quincunx Sampling), the passband is shown in a lozenge, but in Figure 1b, the right side of the diamond Only the upper side is shown.

Even in the case of diagonal sampling (Quincunx Sampling), a rectangular low pass filter can realize a filter that prevents aliasing.However, since the human eye is sensitive to high frequency components in the vertical or horizontal direction rather than the diagonal direction, Use a lozenge filter with high frequency emphasis. However, when a rhombic filter is used, the vertical and horizontal components are not separated into two one-dimensional filterings, and thus only two-dimensional filtering can be implemented.

The sub-sampling method of MUSE has the same sampling pattern for each mode of moving picture and still picture, which is different from the method used in HD-MAC. There is no need to transmit, and in particular, there is an advantage in that a serious image quality deterioration due to an error in mode information between the encoder and the decoder can be excluded.

FIG. 2A is a sub-sampling method of MUSE, in which the sample position is changed for each field, and the inter-field offset sub-sampling takes sampling pixels in four-field sequential. In addition, since the image quality deteriorates, the correlation between adjacent scan lines, fields, and frames is strong, and almost the same shape of a picture or pattern, so interpolation is performed using correlation to compensate for insufficient sample values.

Figure 2b shows a flow chart of the luminance (Y) signal subsampling system of the MUSE. The luminance (Y) signal divides the signal band into a still picture area and a moving picture area. When extracting the sample value for each pixel, one pixel is staggered in the horizontal direction between fields, that is, inter-field offset sampling is performed for sample values between two adjacent fields, and since the moving image area does not need a wide band, Is added to the still image area through the prefilter. The RY and BY signals, which are the chrominance signals C, are also extracted for each scan line, and the first line is the RY signal, the second line is the BY signal, and the third line is the sequential signal. Perform offset subsampling.

FIG. 3A shows the entire pixel signal of the original image, FIG. 3B shows the primary sub-sampling with respect to FIG. 3A, and FIG. 3C shows the secondary sub-sample with respect to FIG. 3B with the primary sub-sampled.

Subsampling can be said to take out the pixels in the middle without using all the pixels of the original image in the shape of a checkerboard. For example, subtracting one pixel in the diagonal direction in FIG. It is called Quincunx Subsampling. Quincunx Subsampling is used a lot in image processing because of the least noticeable overlapping noise that appears when sampling. Even in the case of diagonal sampling (Quincunx Sampling), a rectangular low pass filter can realize a filter that prevents aliasing.However, since the human eye is sensitive to high frequency components in the vertical or horizontal direction rather than the diagonal direction, Use a lozenge filter with high frequency emphasis. However, when a rhombic filter is used, the vertical and horizontal components are not separated into two one-dimensional filterings, and thus only two-dimensional filtering can be implemented.

3d is a filter having a rhombic shape for subsampling, and after filtering the original image as shown in FIG. 3a, the pixel of the 3 pixel point position (X mark) must be extracted, and again, the 3 pixel point position (X). In order to fill the pixels in the display), the pixel at the one pixel point position (○ mark) is inserted into the empty space at the three pixel point position (X mark), and then the rhombic filtering as shown in FIG. This filling of the empty space at the 3 pixel point position (X mark) is called interpolation.

If the second subsampled image is the same as the first subsampled 3b, the second image can be obtained.

Since the human visual structure does not recognize the resolution of the subject well in the moving image area where the subject moves, and the subject's resolution is stingy in the still image area where the subject is still, the resolution of the still image area is increased. There is a need. Therefore, since the resolution must be increased in the still image area of the subject, only the pixels of one pixel point position (○ mark) are transmitted in the image of FIG. 3B, and the pixels of the two pixel point position (△ mark) are transmitted using the pixels of the previous frame. do.

That is, in the still image area, the pixels at the one book point position (○ mark) and the pixels at the two pixel point position (Δ mark) are alternately selected for each frame and transmitted.

In the moving image area of the subject, as shown in FIG. 3C, pixels of one pixel point position (○ mark) subsampled twice are transmitted.

The DCT (Discrete Cosine Transform) of the transform encoding method to be used in the present invention will be described. The high autocorrelation of the image means that there are many parts that change slowly, and in terms of frequency, the energy is concentrated in low frequency components.

The Fourier transform is a method of converting a signal of a time series into a signal of a frequency series. The DCT (Discrete Cosine Transform) also converts a pixel of a time series into a frequency signal according to a matrix operation.

Discrete Cosine Transform (DCT) is based on the large correlation of the video signal in the spatial direction, and concentrates the energy dispersed in all the pixels of the image in several transform coefficients with low frequencies including DC. It's a way to take advantage of the benefits.

4A is a block diagram for transform coding, in which DCT (Discrete Cosine Transform) belongs to one of transform coding. The transmitter converts the input signal, quantizes it, encodes it, and transmits it. The receiving end performs inverse transform by decoding the encoded signal.

4B is a diagram showing the principle of DCT (Discrete Cosine Transformation). When DCT (Discrete Cosine Transformation) is performed using 8 * 8 pixel values as one block, it is converted into a coefficient corresponding to a frequency component of 8 * 8.

Since energy is mainly concentrated in low frequency components, fine quantization of low frequency components and sparse quantization of high frequency components with less energy can reduce errors caused by quantization.

Such quantization is also suitable for the visual characteristic that visibility is reduced at high spatial frequencies. When the image information is transmitted by subsampling the input image signal, performing quantization, encoding, and transmitting the image information, the decoder reconstructs the original image through interpolation.

As described above, subsampling an input video signal can reduce pixel information in half. However, there is a problem that the reconstructed image is of lower quality than the original image because pixel information may be lost when interpolation is performed in the conventional decoder.

However, in order to solve the above-described problems, the present invention may further include an interpolator installed in the subsampling unit of the encoder, so that the difference between the original input image signal and the signal passing through the subsampling unit and the interpolator may occur. DCT processing is performed separately from the sub-sampled signal, and then quantized, encoded, and transmitted, thereby restoring an image close to the original image using the missing pixel information signal transmitted during interpolation in the decoder. The purpose is.

The encoder of the present invention for achieving the above object comprises a sub-sampling unit for performing sub-sampling to reduce the pixel information of the input image signal; A first discrete cosine transformer for converting the pixel values of the sub-sampled time series from the sub-sampling unit into a signal of a frequency series according to an operation of a matrix having one block of 8 * 8; A first quantizer configured to digitally quantize data bits to the image signals converted by the first discrete cosine converter; A first encoder for encoding an image signal quantized by the first quantizer; An interpolator for interpolating a subsampled image signal from the subsampling unit; A subtractor for generating a self signal of the input image signal and the sub-sampling signal interpolated in the interpolator; A second discrete cosine converting unit for converting a difference signal between the input image signal from the subtractor and the sub-sampling signal interpolated by the interpolator into a signal of a frequency sequence according to an operation of a matrix having one block of 8 * 8; A second quantizer configured to digitally quantize data bits to the difference signals converted by the second discrete cosine converter; And a second encoder for encoding the difference signal quantized by the second quantizer. Hereinafter, with reference to the accompanying drawings will be described in detail with respect to the present invention.

5 is a partial block diagram of an encoder according to the present invention, in which a subsampling unit 10 which performs subsampling on an input image signal, and a matrix in which one block of pixel values of the sampled time series of the input image signal are 8 * 8. A first discrete cosine converting unit 12 for converting into a signal of a frequency series in accordance with the operation of; A first quantizer (14) for allocating data bits to digital digitization of the image signal converted by the first discrete cosine converter (12); A first encoder 16 for encoding an image signal quantized by the first quantizer 14; An interpolator (18) for interpolating the subsampled image signal from the subsampling unit (10); A subtractor (20) for generating a difference signal between the input image signal and the subsampling signal interpolated by the interpolator (18); A second discrete cosine converter for converting a difference signal between the input image signal from the subtractor 20 and the sub-sampling signal interpolated by the interpolator into a signal of frequency series according to a matrix of 8 * 8 blocks (22); A second quantizer (24) for digitally quantizing data bits for the difference signals converted by the second discrete cosine transform unit (22); And a second encoder 26 for encoding the difference signal quantized by the second quantizer 24.

Referring to Figure 5 looks at the operation of the present invention.

In order to reduce pixel information of the input image signal, the sub-sampling section 10 extracts the pixels in the middle without using all the pixels of the original image which are like a checkerboard.

The first discrete cosine transforming unit 12 converts the pixel values of the sub-sampled time series from the sub-sampling unit 10 into signals of a frequency series in accordance with a matrix operation in which one block is 8 * 8.

The first quantizer 14 assigns data bits to the image signal converted by the first discrete cosine transforming unit 12 to digitally digitize them. The values between the respective quantization intervals are collectively assigned to one code, and in reconstruction, a quantization representative value representing this range is used. The first encoder 16 functions to encode an image signal quantized by the first quantizer 14.

The interpolator 18 serves to interpolate the subsampled image signal from the subsampling unit 10. The subtractor 20 generates a difference signal by using a difference between the input image signal and the subsampling signal interpolated by the interpolator 18.

The second discrete cosine transformer 22 converts the difference signal subtracted by the subtractor 20 into a signal of a frequency series according to a matrix operation of one block of 8 * 8.

The second quantizer 24 performs digital digitization by allocating data bits to the difference signal converted by the second discrete cosine transform unit 22.

The second encoder 26 functions to encode the difference signal quantized by the second quantizer 24.

As described above, according to the present invention, the interpolator 18 is additionally installed in the sub-sampling unit 10 of the encoder portion, so that the original input image signal and the signal passing through the sub-sampling unit 10 and the interpolator 18 are separated. In other words, the lost pixel information signal that may occur during interpolation is subjected to discrete cosine transform (DCT) processing separately from the sub-sampled signal, and then quantized, encoded, and transmitted, so that the lost pixel information transmitted when interpolation is performed in the decoder part. It is effective to restore an image close to the original image using the signal.

Claims (1)

A subsampling unit 10 for performing subsampling to reduce pixel information of an input image signal, and a pixel value of a subsampled time series pixel value supplied from the subsampling unit 10 A first quantizer for digitally quantizing by allocating data bits to an image signal converted by the first discrete cosine transforming unit 12 for converting into a signal of a frequency series according to the operation (14) and a discrete cosine transform encoder using subsampling having a first encoder (16) for encoding an image signal quantized by the first quantizer (14), supplied by the subsampling unit (10). An interpolator 18 for interpolating a subsampled image signal to be output; A subtractor 20 generating a difference signal between the input image signal and the subsampling signal interpolated by the interpolator 18 and outputting a lost pixel information signal during subsampling; A second discrete cosine transforming unit (22) for converting the lost pixel information signal outputted from the subtractor (20) into a signal of a frequency sequence according to an operation of a matrix having one block of 8 * 8; A second quantizer (24) for allocating data bits to digitally digitized the lost pixel information signal during subsampling converted by the second discrete cosine transforming unit (22); And a second encoder (26) for encoding the lost pixel information signal during the subsampling quantized by the second quantizer (24).