Lossless Compression Handbook

CHAPTER 15 /JPEG-LS LOSSLESS AND NEAR LOSSLESS IMAGE COMPRESSION

309

indicate this hit. At this point the value of

RUNidx

is incremented by 1 and the process continues.

If an end of line is reached during a run that has not reached rm in length, a 1 is also transmitted;

however, the variable

RUNidx

is not changed. If the run ends in a length less than rm, then a miss

is indicated by the transmission of a 0 bit. After this bit

rk

bits are transmitted to indicate the

length of the run (this must be <

2 rk

- 1). The value of

RUNidx

after a miss is decreased by 1. The

coding of the run interruption sample is similar to regular mode coding except that two separate

contexts are reserved for determining the coding used on this sample. In addition, the prediction

uses either the pixel

Ra

that started the run or its neighbor

Rb

as the prediction without further

correction.

15.2.3 JPEG-LS Decoding

Since the JPEG-LS encoding algorithm uses information available to the decoder, the decoding

process recapitulates many of the basic encoder functions. After decoding the entropy data, the

decoder either completes the run replication (in run mode) or reconstructs the error values (in

regular mode) and adds these to the predicted values (with the same corrections provided at the

encoder) to generate the image. Note that the context definitions and updates in the decoder are

also identical to those produced in the encoder.

15.3 SUMMARY

Figure 15.10 illustrates near lossless coding for the

sensin

test image (8-bit grayscale). The

original image is shown on the top left (NEAR = 0). The reconstructed images for NEAR =

3, 7, and 15 are shown on the top fight, bottom left, and bottom fight, respectively. Note the

degradation in image qualitymparticularly in fiat regions such as the blackboardmas the value

of NEAR becomes larger.

The compression attained on the

sensin

image for varying values of NEAR is illustrated in

Table 15.1. By removing the restriction of lossless compression and replacing it with a near

lossless constraint that the maximum difference allowed in defining a run is 1, the compressed

file size drops nearly 40%, a savings of about 1.4 bits per pixel. By increasing NEAR to 3, an

additional 0.75 bits per pixel of compression are obtained. As was seen in Fig. 15.10, for larger

values of NEAR the additional compression probably is not worth the artifacts introduced by the

lossy codingmi.e., it is not near lossless compression, rather it is lossy compression. At these

lower rates, better image compression strategies exist [4].

Table 15.1 Compression in Bits per Pixel and Percentage of Original as NEAR Is Varied

from 0 (Lossless) to 15 for the 8-Bit Grayscale Image

sensin

NEAR Bits/Pixel % of Original

0 3.704 46.3%

1 2.281 28.5%

3 1.528 19.1%

7 1.093 13.7%

15 0.741 9.3%