Professional Documents
Culture Documents
e are witnessing phenomenal ally in analog form. However, for pro- Lossy schemes are widely used since the
20 IEEE POTENTIALS
019_23 2/19/01 12:00 PM Page 21
entropy coding after the quantization the decoder is built into the encoder). the energy of the original data being con-
step, however, is lossless. The decoding The design of a DPCM coder centrated in only a few of the significant
is a similar but reverse process: a) involves optimizing the predictor and the transform coefficients. This is the basis
entropy decoding is applied to the com- quantizer. The inclusion of the quantizer of achieving the compression. Only
pressed data to get the quantized data, b) in the prediction loop results in a com- those few significant coefficients are
dequantization is applied to it, and then plex dependency between the prediction selected and the remaining are discarded.
c) the inverse transformation to get the error and the quantization error. The selected coefficients are considered
reconstructed image. (This is an approxi- However, the predictor and quantizer are for further quantization and entropy
mation of the original image.) usually optimized separately, since a encoding. DCT coding has been the most
Major performance considerations of joint optimization is usually complex. common approach to transform coding.
a lossy compression scheme are: a) the [Under mean-squared error (MSE) opti- It is also adopted in the JPEG image
compression ratio (CR), b) the signal-to- mization criterion, independent optimiza- compression standard. The broad outline
noise ratio (SNR) of the reconstructed tions of the predictor and quantizer are of transform coding of images is shown
image with respect to the original, and c) good approximations to the jointly opti- in Fig. 5.
the speed of encoding and decoding. The mized solution.] Subband coding. In this scheme, the
compression ratio is given by: Block truncation coding. In this image is analyzed to produce the compo-
size of uncompressed data scheme, the image is divided into non- nents containing frequencies in well-
CR = overlapping blocks of pixels. For each defined bands, the subbands.
size of compressed data
block, threshold and reconstruction val- Subsequently, quantization and coding is
The PSNR is given by: ues are determined. The threshold is applied to each of the bands. The advan-
PSNR = 20 log10(peak data value/RMSE) usually the mean of the pixel values in tage of this scheme is that the quantiza-
where RMSE is the root mean square the block. Then a bitmap of the block is tion and coding well suited for each of
error, given by: derived by replacing all pixels whose the subbands can be designed separately.
values are greater than or equal (less The broad outline of transform coding of
1 N M
[ ]
2
RMSE = ∑ ∑ Ii, j − Iˆi, j
NM i =1 j =1
than) to the threshold by a 1 (0). Then images is shown in Fig. 6.
for each segment (group of 1s and 0s) in Vector quantization. The basic idea in
where N × M is the image size, Ii,j and Ii,j the bitmap, the reconstruction value is this technique is to develop a dictionary
are values of pixels at (i,j) in the original determined. This is the average of the of fixed-size vectors, called code vectors.
and the reconstructed (compressed- values of the corresponding pixels in the A vector is usually a block of pixel val-
decompressed) images, respectively. original block. The broad outline of ues. A given image is then partitioned
Predictive coding. In most images, block truncation coding of images is into non-overlapping blocks (vectors)
there is a high correlation among neigh- shown in Fig. 4. called image vectors. Then for each
boring pixels. This fact is used in pre-
dictive coding. Differential Pulse Code Encoder
Modulation (DPCM) is a popular pre- dm Third-Order
Predictor
dictive coding technique. The lossy xm dm
Quantizer
qm Entropy Channel
Coder B C
DPCM is very similar to the lossless pm
version. The major difference is that in Reconstructor xm Original Value
A
lossy DPCM, the pixels are predicted sm
pm Predicted Value
based on the “reconstructed values” of dm Prediction Error
Predictor rm Reconstructed Value Predicted Value of this Pixel:
certain neighboring pixels. The differ- rm
sm a.A + b.B + c.C
Quantized, Reconstructed,
ence between the predicted value and Prediction Error
the actual value of the pixels is the dif- Fig. 3 Lossy DPCM coding scheme
ferential (residual) image. It is much
less correlated than the original image.
The differential image is then quantized Divide Original
Image Into Blocks
and encoded. Determine
Quantization Threshold
The schematic for lossy DPCM coder and Reconstruction
Bitmap Entropy Code:
Quantize of Block Bitmap and
is shown in Fig. 3, along with a third- Level for Each Block Each Block Reconstruction
order predictor. (In a third-order predic- Level
FEBRUARY/MARCH 2001 21
019_23 2/19/01 12:01 PM Page 22
Image Components
Group 4 Fax
at Different The Group 4 (G4) fax standard is a
Frequency Subbands
Original Image superset of the Group 3 standard and is
backwards compatible with it. The G4
Analysis Sub- standard is said to use a 2-dimensional
Quanti- Entropy
Filter
Sampling zation Coding
Bank coding scheme. This is because it uses
spatial redundancy in the vertical direction
also by making use of the previous line as
Fig. 6 Outline of subband coding of images a reference while coding the current line.
Most runs on a line usually lie nearly
ments by using standard image process- secutive pixels of the same value. The directly below a run of the same color in
ing techniques such as color separation, first run on each line is assumed to be the previous line. The differences in the
edge detection, and spectrum and texture white. Each line is considered to be made run boundaries between successive lines
analysis. Then each segment is looked up of 1728 pixels. Thus each line is are coded. The cases where a line may
up in a library of fractals. The library reduced to alternating runs of white and have fewer or more lines than the refer-
actually contains codes called iterated black pixels. The runs are then encoded. ence lines are suitably handled. The
function system (IFS) codes, which are Each end of a line is marked with an Group 4 standard generally provides
compact sets of numbers. Using a sys- EOL. Page breaks are denoted with two more efficient compression than Group 3.
tematic procedure, a set of codes for a successive EOLs.
given image are determined, such that Two types of encodings are used for JBIG
when the IFS codes are applied to a suit- run lengths—terminating codes and The Joint Bi-Level Image Group
able set of image blocks yield an image make-up codes. Terminating codes are (JBIG) standard was developed by the
that is a very close approximation of the
original. This scheme is highly effective
for compressing images that have good Scalar quantization
regularity and self-similarity. The broad Quantization is a process (function) that maps a very large (possibly infinite)
outline of fractal coding of images is set of values to a much smaller (finite) set of values. In scalar quantization, the
shown in Fig. 8. values that are mapped are scalars (numbers). In the context of image coding
and decoding, the range of pixel values say N, is divided into L non-overlapping
intervals, also known as quantization levels.
Image compression Each interval i is defined by its decision boundaries (di, di+1). During encod-
standards ing, the quantizer maps a given pixel value x to a quantization level
Image compression standards have l: l = Q(x), such that dl ≤ x < d l+1. Each quantization level i has its associated
been developed to facilitate the interop- reconstruction level ri.
erability of compression and decompres- During decoding, the (de)quantizer maps a given level l to a reconstruction
sion of schemes across several hardware pixel value rl = xˆ, xˆ = Q-1 (l). This introduces noise or error in the image (signal)
platforms, operating systems and appli- called quantization error. This is the root mean square value of the x - x. ˆ
cations. Most standards are hybrid sys- The essential difference among various types of quantizers is in terms of how
tems making use of a few of the basic the forward and inverse mappings are defined. These definitions are dictated
techniques already mentioned. The major according to the number of quantization levels, the decision boundaries and
image compression standards are Group the reconstruction values. The basic design objective of a quantizer is to minimize
3, Group 4, and JBIG (Joint Bi-level the quantization error, while being computationally simple. The quantizer has a
Image Group) for bi-tonal images, and large impact on the compression ratio and image quality of a lossy scheme.
JPEG (Joint Photographic Experts There are two broad types of scalar quantizers—uniform and non-uniform. In
Group) for continuous-tone images. The a uniform quantizer of k levels, the range of values is divided into k equally
most common application that uses com- spaced intervals. The reconstruction values are the mid-points of the intervals.
pression of bi-tonal images is digital fac- This is simple to implement but it does not attempt to minimize the quantization
simile (FAX). error. A quantizer that takes into account the probability distributions of the pix-
els in images performs better. Such a quantizer is a non-uniform quantizer,
Group 3 Fax where the intervals are non-uniform. The most common non-uniform quantizer
The image is scanned left-to-right and is the Lloyd-Max quantizer. For it, the decision boundaries and the reconstruc-
top-to-bottom and the runs of each tion levels are determined using the probability model of the image pixels such
color—black and white—are deter- that the quantization error is minimized.—SRS
mined. A run refers to a sequence of con-
22 IEEE POTENTIALS
019_23 2/19/01 12:01 PM Page 23
FEBRUARY/MARCH 2001 23