JPH0686077A - Hierzrchical image compressing system - Google Patents

Abstract

PURPOSE:To transfer the images at a high speed with the quantization performed to compress the images at the high rate, to correct the error of the quantization, and to reproduce the images of high picture quality. CONSTITUTION:The code data on the partial images obtained by dividing an original image 1 undergo the discrete cosine transformation DCT by a discrete cosine transformer 3, and a DCT coefficient is outputted in response to the space frequency. Then the DCT coefficient is quantized by a quantizer 4 by the step width decided by a quantization table 7. Thus a quantization coefficient is obtained. The code data is divided by the quantization coefficient and the carry is done by counting fractions of.5 and over as a unit and cutting away the rest by a quantization error calculating means 6. Then the ratio set between the quantization value obtained by the carry and the value obtained by counting fractions of.5 or over as a unit and cutting away the rest is defined as the error data. An expanded image is corrected based on the error data and therefore the picture quality of the expanded image is obtained.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a hierarchical image compression method, and more specifically, it has an error generated when an image is compressed as another information hierarchically, and refers to the error information at the time of expansion. The present invention relates to a hierarchical image compression method for improving image quality.

[0002]

2. Description of the Related Art In recent years, there has been a rapid move toward computer multimedia and TV / VTR systems from analog to digital. This is because the image is digitally processed: -No deterioration of image quality occurs during processing. -Various image processing can be performed easily and freely. -Any place in the data can be instantly accessed. It is due to the reason such as.

However, the amount of information of a digital color image becomes very huge, and for example, in order to convert one image data of A4 size with an accuracy of 400 dpi (dots / inch), it is actually about 48 Mbytes (megabytes). It needs a lot of information. Also, regarding the transmission time, the above A4
A phone line with a transfer rate of 9600 baud for a size image would take about 14 hours, and even using Ethernet would take about a minute. Therefore,
Transmission and storage of digital color image information is very difficult due to current storage and digital line capacity.

Therefore, it is extremely important to perform compression / expansion processing for reducing the transmission band, transmission time and storage capacity for the transmission and storage.
By the way, when transmitting compressed image data between various systems and decompressing after accumulating, it is an important element to maintain data compatibility. For this reason, there has been an increasing demand for standardization of a system capable of encoding digital image data of high image at high speed and with high efficiency and performing transmission / storage processing. However, a large amount of calculation is required to perform compression / expansion at a high compression rate without deteriorating the accuracy of image quality.

Therefore, it is not possible to perform a sufficiently high-speed processing using a conventional general-purpose processor or the like. Furthermore, a system that uses discrete components such as a dedicated arithmetic unit and an encoder for high-speed processing becomes complicated and expensive, and it is difficult to reduce the size, cost, and development period. It was However, with the development of IC technology in recent years, such movements have become more and more intense.

[0006] Based on the current situation, ISO (Intern
ational Organization for Standardization / IEC (International Electrotechnical C)
ommission: International Electrotechnical Commission and CCITT (Interna)
tional Telegraph and Telephone Consultive Committe
e: International Telegraph and Telephone Advisory Committee)
The (Joint Picture Expert Group) Committee has been working since 1986 to set standards for compression / decompression of still / continuous tone (color including halftone, black and white) digital images. At present, the algorithm itself has already announced the final specifications and will become an international standard within this year.

The compression algorithm of the above-mentioned JPEG basic system can be roughly divided into the following three different stages. 1. Redundancy removal of YUV data by two-dimensional DCT (Discrete Cosine Transform). 2. Quantization of DCT coefficients adapted to human visual characteristics. 3. Entropy coding of quantized DCT coefficients by Huffman arithmetic coding. Here, the YUV signal is a color difference signal in which Y represents the degree of brightness of the target pixel based on the separation coding method defined in CCIR-601 and Y represents a luminance signal. It represents the hue and saturation of the target pixel. As described above, the YUV signal has a form in which the brightness signal representing the brightness and the color difference signal representing the color property are classified, so that if the target image is a monochrome image, only the brightness signal is required and the color difference signal is required. Therefore, the amount of information can be shortened as compared with the RGB signal.

Further, since the YUV signal is a multiplex of a luminance signal having a wide bandwidth and a color difference signal having a relatively narrow bandwidth due to its characteristic, the quantization pit rate is 4: 2: 2 when normally transmitted. Subsampling is performed so that the ratio is 4: 1: 1. The YUV signal has the following relationship with the RGB signal. Y = 0.299 × R + 0.587 × G + 0.114 × B U = −0.169 × R−0.3316 × G + 0.5 × B V = 0.5 × R + 0.4186 × G−0.01313 × B It should be noted that when decompressing, the reverse steps are included.

FIG. 9 is a block diagram showing the configuration of a conventional DCT basic system. For example, an original image 21 divided into 8 × 8 blocks is a DCT-based encoder 22.
Are converted into parameters and code data 28 by. DCT
The encoder 22 based on is composed of a DCT 23, a quantizer 24, and an entropy encoder 25,
The 8 × 8 block original image is subjected to DCT conversion in block units by the DCT 23 and converted into DCT coefficients corresponding to spatial frequencies. The converted DCT coefficient is converted into D by the quantizer 24 based on the quantization table 26.
The CT coefficient is quantized and the quantized DCT coefficient is output. The quantized DCT coefficient is further entropy-encoded by the entropy encoder 25 according to the encoding table 27. The entropy-encoded parameters and the coded data 28 are transmitted through the transmission line 29, and the DCT operates in the reverse step of the encoder 22.
The decoded image data is obtained by decompressing it by the decoder 31 based on. That is, the decoder 31 based on DCT can obtain reproduced image data of 8 × 8 blocks. Decoding is performed by the entropy decoder 32 according to the encoding table 27, the quantized DCT coefficient of the quantizer 33 is inversely DCT-transformed by the IDCT 34, and 8 × 8 reproduced image data 35 is obtained.

[0010]

However, not only in the above-mentioned conventional JPEG method, but also in the method using other vector quantization and various orthogonal transformations, when an image, particularly a color image is targeted, Generally, the amount of information is very large, so it takes a long time to perform communication and storage, and in order to obtain a practically usable compression rate, encoding such as entropy coding, DPCM, or run length conversion is performed. The lossy coding that can obtain a large compression rate is often used, though there are some errors like the vector quantization and the orthogonal transform described above, rather than the lossless coding that does not cause an error before and after . Therefore, until now, how to make the resulting error inconspicuous by utilizing human visual characteristics has been a major factor in determining the quality of the restored image.

For example, considering the case of the JPEG system described above, it is the DCT transform and quantization portions that may cause errors. Here, the error that occurs during the DCT conversion is an error that occurs when the DCT coefficient is cut off after the decimal point and carried when carrying out the calculation of the DCT conversion. Therefore, it is sufficiently ignored except for special applications. It is a good value. Therefore, the problem is the error that occurs in the quantization part. As described above, the quantization is represented by a small number of bits by dividing the code data by the value of the quantization coefficient. In other words, in other words, since the gradation is changed, the error caused by this change cannot be ignored in most cases. Even if the optimal bit allocation is adaptively performed by utilizing the human visual characteristics and the general spatial frequency characteristics of the image, in order to obtain a sufficient compression rate, some coarse quantization is required. This is not an error that can be ignored because it must be done.

[0012]

In order to solve the above-mentioned problems, the present invention provides: (1) a discrete cosine converter for converting an image input divided into a plurality of blocks into code data corresponding to a spatial frequency ( DCT) and the coded data coded by the discrete cosine transformer are divided by a predetermined number of quantization steps, and a quantizing means for outputting image compressed data, and the division by the quantizing means are generated. From a quantization error calculating unit that calculates a quantization error and outputs the quantization error data, and an entropy encoding unit that inputs the image compression data and the quantization error data and encodes them according to the appearance probability of the data. Image coding means, entropy decoding means for cascading the data transmitted from the image coding means, a quantizer, and an inverse dissociative cosine transform (IDCT). Input to the reconstructed image encoding means, the quantizer signal of the reconstructed image encoding means is corrected based on the quantized error data to expand the reproduced image, and (2) quantized error calculation Requantization of the quantization error calculated by the means, requantization error calculating means for calculating the error generated by the requantization, and the error of the image compression data based on the quantization error and the requantization error And image reproducing means for reproducing the original image by decompressing the compressed image data by correcting and compressing the error-compressed image data, and (3) compressing the image signal to form compressed image data. Image compression means, error calculation means for calculating an error between the compressed image data compressed by the image compression means and the current image, and the error calculated by the error calculation means based on the error calculation means. Corrects the image errors, is characterized in extending the image.

[0013]

[Operation] 1. In an image compression algorithm that includes quantization in image compression, the value of the error due to the quantization that occurs in the image compression data after quantization is stored as separate information, and, in some cases, the value obtained by requantizing the quantization error. When decompressing the compressed data of the image by decompressing the compressed data of the image by storing as a new quantization error value as separate information,
By referring to the value of the quantization error stored in the separate information, the quality of the expanded image is significantly improved. 2. Furthermore, when decompressing with the data encoded as in the above method and the error value as separate information for the compression method that may cause an error without being limited only to quantization, The original image is theoretically reproduced as it is by referring to it.

[0014]

1 is a block diagram of a basic system for explaining an example of a hierarchical image compression system according to the present invention, in which 1 is an original image, 2 is a DCT-based encoder, 3 is a DCT, 4 is a quantizer, 5 is an entropy encoder, 6 is a quantization error calculating means, 7 is a quantization table, 8 is an encoding table, 9 is a parameter and code data transmitter, and 10 is a transmission line. , 11 is a parameter and code data receiver, 12 is a DCT-based decoder, 13 is an entropy decoder, 14 is a quantizer, and 15 is an inverse DC.
T and 16 are reproduced image data. Hereinafter, the constituent blocks of the basic system of the hierarchical image compression method shown in FIG. 1 will be described with reference to the drawings.

FIG. 3 is a diagram showing an example of block interleaving. In the JPEG system, the original image 1 is divided into several blocks, for example, blocks 1, 2, 3, ..., Blocks i, i + 1 ,. Divided into meshes,
DCT conversion is applied to each block. Considering the error accuracy and the processing speed, the block size is 6 (normally 8 × 8).
It is 4 pixels.

DCT conversion The DCT conversion formula is shown in the following equation 1.

[0017]

[Equation 1]

FIGS. 4A and 4B are diagrams showing a method of converting color image data into DCT coefficient values obtained by DCT conversion. FIG. 4A is color image data, and FIG.
The figure shows an array of DCT coefficient values after DCT conversion. DCT
Is a kind of orthogonal transform, and if DCT transform is applied to each block of the color image data in FIG.
8) constants, as shown in FIG.
Convert to CT coefficient. At this time, the blocks are independent or uncorrelated due to the property of orthogonal transformation. This DCT
As a result of the conversion, the input image data is converted into an output value corresponding to the spatial frequency. Here, the value at the upper left of each block is the average of all pixels, which is called the DC value, and other coefficients are represented by the correlation with this DC value. These values are called DC values for other blocks with respect to DC values,
The arrow u direction (horizontal direction) of the block represents the horizontal spatial frequency direction, and the arrow v direction (vertical direction) represents the vertical spatial frequency direction, which indicates that the higher the spatial frequency is, the further to the right and downward. Here, the spatial frequency represents the distribution of the density of the image in the block, that is, the degree of variation when the histogram is taken.

In the DCT3, 64 pixels of each block are converted into 64 DCT coefficients, and the image data is not compressed as it is. Therefore, the quantizer 4 uses the redundancy of the DCT coefficient to perform linear quantization.

5 (a), (b), (c) and (d) are D
FIG. 3 is a diagram for explaining a method of scalar quantization from CT coefficients, which is a histogram of DCT coefficient values by zigzag scanning (FIG. (B)) the DCT coefficient values of 8 × 8 blocks (FIG. (A)) according to the arrows. (Fig. (C)) is obtained. A quantized coefficient value (FIG. (D)) of a small amount of data having a step width determined by the threshold value of the quantization table 7 is obtained. Generally, in an image, there is not much change in the density difference between adjacent pixels, so the DCT coefficient becomes smaller as the frequency becomes higher. Therefore, as shown in the figure, the higher the frequency component is, the coarser the quantization is performed, that is, the lower bit rate is assigned to reduce the data amount. Furthermore, the human visual characteristics can be used to quantize the chrominance component somewhat more coarsely than the luminance component.

The DCT coefficient quantized with the step width of the quantization table has a relatively large DC value, and the higher the AC value, the more likely it is to be "0". By utilizing this fact, entropy coding is performed by the entropy coding (variable length coding) device 5, so that the amount of information can be further reduced. First, the entropy encoder 5 performs zigzag scanning on the DCT coefficient to convert it into one-dimensional data.

FIGS. 6A, 6B and 6C are zigzag scans of the quantized DCT coefficient values shown in FIG.
It is a figure which shows the method of converting into the one-dimensional data of a figure, and arranging in order of a one-dimensional difference PCM value (DPCM) like a figure (c).
The frequency components in the horizontal and vertical directions are rearranged in order from the lowest coefficient. Here, the DCT coefficient has a relatively large DC value, and “0” values tend to line up in the AC value, so D
Encoding is performed by dividing the C value and the AC value. First, with respect to the DC value, the value itself is large, but in a natural image, it is expected that there will be little abrupt color change between neighboring pixels, so the average value of pixels between adjacent blocks is relatively small. Can be thought of as something. Therefore, the amount of information can be reduced by encoding a value obtained by subtracting the DC value of the immediately preceding block from the current block. Regarding the AC value, first, the value is “0” from the coefficient in the low frequency range.
Is determined. After that, a one-dimensional run-length transform is performed to zero-pack a value that is not "0" and the number of "0s" that follow, and encoding is performed using a combination of a DCT coefficient and a run-length size. FIG. 7 shows a specific example of the flow up to this point.

7 (a), (b), (c), and (d) are diagrams showing an example of the run-length transform. For example, DCT coefficient values ((a )) Is subjected to a zigzag scan (Fig. (B)) diagonally in the arrow direction to obtain a one-dimensional coefficient value (Fig. (C)), and the one-dimensional coefficient value is encoded into a data stream ((d ) Figure) is converted. In the JPEG system, Huffman coding or arithmetic coding is used as coding. Here, only Huffman coding will be described.

FIG. 8 is a diagram for explaining the principle of Huffman coding. In Huffman coding, the code with the shortest code length is assigned in order from the one with the highest appearance probability according to the data appearance probability. This is one of the variable length coding methods. That is, in the illustrated example, the highest appearance probability is 0.3, 0.25, 0.2, 0.1, 0.08, 0.05.
, S6 are represented by symbols S1, S2, ..., S6. Those having a high appearance probability are low bits, those having a low appearance probability are high bits, and those having a high appearance probability are low-bit codes 00, 01, to 1110, 1111. Is to be assigned.

The quantization error calculating means 6 divides the image data read from the block of the original image 1 by dividing the code data converted by the DCT converter 3 by the quantization coefficient, and dividing the divided value. After rounding off after the decimal point of, the difference between the compressed data and the division value obtained by carrying, and the difference between the code data when dequantizing this compressed data is calculated as an error value, For example, if the coded data is 230 and the quantized value is 99, the value after quantized the coded data is 230 ÷ 99 = 2.32,
Rounding up 0.32 to the right of the decimal point and then carrying
Compressed data 2 is obtained. The error at this time is 2.32
Although it is a relatively small value of −2 = 0.32, when this compressed data is inversely quantized, it becomes 2 × 99 = 198, and when it is compared with code data 230, a large error of 230−198 = 32 is obtained. This error is used as error information.

If necessary, the error is requantized with a quantization coefficient that is fine to some extent, and the requantized value obtained by this requantization is encoded. The error information and the re-error information are entropy-encoded by the entropy encoder 5 as in the JPEG method shown in FIG. Further, while the Huffman code is used for the normal compressed data according to the code value and the length of the run length, the error data uses the interval between the error value and the code that generated the error one before. Huffman codes can also be used. Since this error data is hierarchical data of compressed data, it can be used as needed.
If only the vertical pressure data is used, compatibility with the conventional JPEG system can be maintained. As described above, the error value generated before and after the quantization is stored as the error information as the hierarchical information of the compressed data, so that the quality can be improved by referring to this as needed during the decompression. Can be increased.

FIG. 2 is a flowchart of the hierarchical image compression method according to the present invention. The flow will be described below. step1: Here, the original image divided into 8 × 8 blocks is read. step 2: DCT conversion is performed (DCT (I)) corresponding to the I-th block of the divided original image, and this is read as an input value for performing quantization. step 3: The read DCT (I) is divided by the value quantization table value Q (I) of the predetermined number of quantization steps to obtain the quantized value PIC (I). The obtained quantized value PIC (I) is digitized by carrying or rounding down to the right of the decimal point and quantizing the error value GOSA (I) generated with the I-th quantized value PIC (I). Store as an error value. This is done for all quantized values PIC. step4: I-th quantization error value GOSA if necessary
(I) is the I-th quantization table value for error data Q ′
Then, the quantized value is divided by (I) to obtain a requantized value GOSA '(I) having a finer number of quantization steps. This is done for all quantization errors GOSA. Note that st
ep4 may be omitted. Step 5: Check whether the requantization error value GOSA '(I) is "0". . Step 6: If the requantization error value GOSA '(I) is not "0" in step 5, entropy coding is performed by the combination of the quantization value error GOSA (I) and the run length size ("0" in this case). .

Step 7: When it is determined in step 5 that the requantization error value GOSA '(I) is "0", the runlen size E is increased by 1 to E + 1, the next input value is read, and the process returns to step 2. , Quantization is repeated until the requantization error value GOSA '(I) is not "0". Then, this requantization error value GOS that is not "0"
Entropy coding of step 6 is performed by the combination of A '(I) and the run length size so far, and this is output. step8: Here, the input value I is 8 × 8 per block
Since there are = 64 of them, this is repeated until all of them are read and the result is output. If the last quantization error in one block is “0”, regardless of the run length size so far. , EOB (End of Block) entropy coding is determined and output to indicate that error correction of one block is completed. step9: To perform entropy coding on all input values for which entropy coding has not finished in step6 s
Return to tep2.

This method can also take the form of band division coding. For example, when quantizing normal compressed data, there is almost no error in the low frequency region, or fine quantization that can be ignored,
The high frequency region is quantized so as to be almost 0, and instead, the error information uses only the error in the high frequency region. Since the visual characteristics of the human eye are sensitive to low frequencies and relatively dull to high frequencies, it is usually compressed data,
That is, even if only the values in the low frequency region are used, a certain amount of image quality can be guaranteed. Then, the image quality can be further improved by adding error data, that is, a high-frequency region, if necessary.

[0030]

EFFECTS OF THE INVENTION In order to transmit an image having a very large amount of information in a short time to form a reproduced image, lossy encoding which has a large error, but which can obtain a large compression rate, has been conventionally performed. With respect to the image quality deterioration, the human visual characteristics are simply used to make it inconspicuous, but according to the present invention,
By storing the value of the error generated before and after the quantized part where the error occurs most often as the error information, the error information is referred to and the image is restored. High quality images can be transmitted in a short time.

[Brief description of drawings]

FIG. 1 is a configuration block diagram of a basic system for explaining an example of a hierarchical image compression method in the present invention.

FIG. 2 is a flowchart of a hierarchical image compression method according to the present invention.

FIG. 3 is a diagram showing an example of block interleaving.

FIG. 4 is a diagram showing a method of converting color image data into DCT coefficient values obtained by DCT conversion.

FIG. 5 is a diagram for explaining a method of scalar quantization from DCT coefficients.

6 (a) is a quantized DCT coefficient value is zigzag scanned to be converted into one-dimensional data in FIG. 6 (b),
It is a figure which shows the method of arranging in order of one-dimensional difference PCM value (DPCM) like the figure (c).

FIG. 7 is a diagram showing an example of run-length conversion.

FIG. 8 is a diagram for explaining the principle of Huffman coding.

FIG. 9 is a configuration block diagram of a conventional DCT basic system.

[Explanation of symbols]

1 ... Original image, 2 ... DCT-based encoder, 3 ...
DCT, 4 ... Quantizer, 5 ... Entropy encoder, 6
... Quantization error calculation means, 7 ... Quantization table, 8 is an encoding table, 9 ... Parameter and code data transmitter, 10
... transmission line, 11 ... parameter and code data receiver, 12
... Decoder based on DCT, 13 ... Entropy decoder, 14 ... Quantizer, 15 ... Inverse DCT, 16 ... Reproduced image data.

Claims (3)

[Claims] 1. A discrete cosine transformer (DCT) for converting an image input divided into a plurality of blocks into code data corresponding to a spatial frequency, and coded data coded by the discrete cosine converter. Quantization means that divides by a predetermined number of quantization steps and outputs image compression data, and quantization error calculation means that calculates a quantization error caused by the division of the quantization means and outputs quantization error data. And an image encoding means including the image compression data and the quantization error data and encoding the image compression data according to the appearance probability of the data, and the data transmitted from the image encoding means in cascade. Connected entropy decoding means, quantizer, and inverse dissociative cosine transformer (I
DCT) and a reconstructed image is expanded by correcting the quantizer signal of the reconstructed image coding means based on the quantized error data. Compression method. 2. A requantization error calculating unit for requantizing the quantization error calculated by the quantization error calculating unit and calculating an error generated by the requantization, and the quantization error and the requantization error. 2. The hierarchy according to claim 1, further comprising image reproducing means for correcting an error of the image compressed data based on the above, and expanding the image compressed data by the error-compressed image compressed data to reproduce the original image. Image compression method 3. An image compression unit for forming compressed image data by compressing and encoding an image signal, an error calculation unit for calculating an error between the compressed image data compressed by the image compression unit and the current image, and the error. A hierarchical image compression method, wherein an image error of the compressed image data is corrected on the basis of the error calculated by the calculation means to expand the image.