GB2253113A - Rounding numbers in quantization method used for image compression - Google Patents

Description

2 2 5 3, 1 1 -- 1 NOVEL QUANTIZATION METHOD FOR USE IN IMAGE COMPRESSION

BACKGROUND OF THE INVENTION

The present invention relates to a novel quantization method for use in an image compression system.

The actual data compression in image compression is achieved at the quantization step of the overall process. The principle behind the forthcoming PJEG international standard is to transform the spatial dimension (pixel data) to a series of 2 dimensional discrete cosine transform coefficients. It is these coefficients that are quantized in software or hardware and subsequently encoded to achieve actual compression.

SUMMARY OF THE INVENTION

It is an object of the present invention to provide an improved quantization method for use in image compression techniques. The improved quantization process yields a 3 - 6% improvement in compression (given a certain error level in the compressed file). It is important to note that this is achieved with minimal additional hardware and leaves the system still 100% compatible with the proposed standard. By using a rounding technique, a 3 to 6% gain in compression can be achieved with no additional loss of quality.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments of the invention and, together with the description serve to explain the principles of the

2 invention. Fig. 1 shows the data flow for a typical image compression apparatus.

Figs. 2A and 2B show the general arrangement of transform coefficients as utilized in image compression techniques, Fig. 3 illustrates the calculation of quantization, Fig. 4 shows a representation of a block of pixels, Figs. 5 and 6 show the results of quantization of typical test images of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

Reference will now be made in detail to the preferred embodiments of the invention, examples of which are illustrated in the accompanying drawings. While the invention will be described in conjunction with the preferred embodiments, it will be understood that they are not intended to limit the invention to those embodiments. On the contrary, the invention is intended to cover alternatives, modifications and equivalents., which may be included within the spirit and scope of the invention as defined by the appended claims.

In Figure 1, the flow of data shows pixels being input to the front end 10 of the system. A typical representation of the pixel values is 0 to 255 or -128 to +127. These values require 8 bits of data to represent.

The second stage is the transform 20. Even though the transform involves multiplication and/or additions by factors other than integers, the end result of the 3 transform is 64 frequency coefficients whose range is typically -1024 to + 1023. These can be represented by 11 bit numbers. These numbers are still considered to b6 integers.

The third stage 24 (shown with darker boundary) is the area of real interest. It is the quantization that actually achieves the compression by reducing many of the high frequency components to zero. The quantization may vary depending upon the coefficient in question. Empirical experimentation has shown that some coefficients can be "suppressed" without an adverse effect on image quality. The actual numerical operation for the quantization is a division by a quantization factor Q. If the result is a number X (as shown in Fig. 3) then we would express X as:

X = C + Q where C is the coefficient value before quantization.

Although allowable values for C are in the range of -1024 to +1024, and for Q the range is 1 to 512, typical values for both will generally be smaller. Much of the time the division of C by Q will result in a small number in the range of -5 to +5. With these small numbers, the value to the right of the decimal point can have a critical effect on the integer value chosen for X. Simple truncation to the right of the decimal point would not be acceptable.

The proposed JPEG international standard proposes the following scheme for rounding of the value x.

-0.5 < X <+0.5 0. 5 < x <+1. 5 1. 5 < x <+2. 5 implies setting implies setting implies setting X 0 X 1 X 2 etc.

4 The present invention (referred to as IIZII quantization) uses the metric:

-0.625 < x <+0.625 implies setting 0.625 < x <+1.625 implies setting 1. 625 < x <+2. 625 implies setting etc.

X 0 X 1 2 X At first glance, the difference between the two approaches may seem very small. However, the net effect on compression ratio, for a given error level, is an improvement of 3 to 6%. A description of the encoder will help explain why.

Figure 2a shows the general arrangement for the 64 transform coefficients. The DC value resides in the top left-hand corner, and the remaining 63 AC coefficients are in order of increasing horizontal and vertical frequency as one moves to the bottom right hand corner.

Figure 2b shows some typical values that may occur for an 8x8 pixel block. The international standard calls for the encoder to "zig-zag" scan through the AC coefficients. The reason for this approach is to lengthen runs of zeros where greatest compression is achieved. As shown in this figure, "stray 11 ones (highlighted in Fig. 2b) often break up otherwise long runs of zeros. Let's say that the value of the marked "stray" 1 was 0.609 before rounding. Using the JPEG metric, the value is rounded up to 1.

The present invention rounds the 0.609 down to zero. Under the JPEG system, the last 18 coefficients are coded as 10 zeros, a 1, and then 7 zeros. With the present invention, the last 18 coefficients are coded as A E a run of 18 zeros. The net result is longer runs of zeros, hence the name IIZII quantization.

One of the big advantages of the improved quantization method is that it remains 100% compatible with the proposed standard. At the decompression end of the system, the decompressor does not need to know whether JPEG or IIZII was used during quantization.

The choice of the value 0.625 for IIZII quantization seems optimal for several reasons. Even so, values other than 0.5 for the "ZI, quantization may be acceptable. A higher value such as 0.75 would probably hurt high bit rate compressions where accuracy is essential. A value of less than 0. 625 would probably only yield a gain of say to 1 to 2% and may not be worthwhile.

Choosing 0.625 yields 3 to 6% and is easy to implement in hardware or software. Consider the output of the x = C/Q calculation, as shown in Fig. 3.

The value x, after rounding, will be an 11 bit -integer. Rounding will involve examination of bits to the right of the decimal place (b- 1 to b_ 3). The JPEG algorithm needs to look at just bit b- 1 to do the rounding:

if b_ 1 = 1 and b_ 10 = 0, then add 1 to x The IIZII quantization approach according to the present invention requires only a slightly more complicated algorithm for the rounding:

if b_ 1 = 1 and (b- 2 or b_ 3 = 1) and b 10 = 0, then add 1 to x 6 Similar logic can be applied to negative numbers.

Results The discrete cosine transform introduces differing degrees of errors at various places in the 8x8 pixel block. For comparison purposes, we will consider the data at the corner pixels "A", and center piXels "B". Figure 4 shows a representation of the block of pixels.

The proposed "Z" quantization method reduces errors everywhere, but is particularly effective at the corner pixels "All.

Figures 5 and 6 show the results for a typical test image. At given bit rates, the "Z" quantization approach is marginally better than JPEG on the center pixels (see Fig. 5 and not that higher SNR equates to less error). On the corner pixels "B", the difference is much more significant. This is where most of the win will come, as seen in Fig. 6.

A very simple change to the quantization circuitry in a JPEG image compression system can yield significant gains in the compression ratio (3-6%) for very little added cost. The quantization method maintains full compatibility with the proposed standard. The choice of the value 0.625 seems close to optimal from the standpoint of both effectiveness and ease of implementation.

The foregoing descriptions of specific embodiments of the present invention have been presented for purposes of illustration and description. They are not intended to be exhaustive or to limit the invention to the precise forms disclosed, and it should be clear that

7

Claims (1)

1. many modifications and variations are possible in light of the above

   teaching. The embodiments were chosen and described in order to best explain the principles of the invention and its practical application, to thereby enable others skilled in the art to best utilize the invention and various embodiments with various modifications as are suited to the particular use contemplated. It is intended that the scope of the invention be defined by the Claims appended hereto and their equivalents.

   8 In the Claims:

   1. A quantization methodFor use in image compression, the method comprising the steps of receiving input pixels of N by N pixel blocks, transforming said pixel blocks to a frequency domain format resulting in a plurality of frequency transform coefficients including a DC coefficient and a plurality of AC coefficients, quantizing said frequency coefficients to form a multibit integer including a most significant bit and including at least three decimal bits, examining the first decimal bit and the second or third decimal bits, examining the most significant bit, and adding a logical one to said integer if said first decimal bit and said second or third decimal bits are a logical one and said significant bit is a logical zero.