Download (direct link):
It should be remembered that the achievable compression ratio using a lossless compression scheme is totally input data dependent. If the same algorithm is applied in a number of distinct data files, the algorithm will yield a different compression ratio in different files. The maximum compression ratio and hence the bits per sample that can be achieved losslessly is restricted by the entropy of the data file according to the noiseless source coding theorem by Shannon. Sources with less redundancy have more entropy and hence are more difficult to achieve compression. For example, it is very difficult to achieve any compression in a file consisting of mainly random data.
1.6.2 Quality Metrics
This metric is not relevant for lossless compression algorithms. The quality or fidelity metric is particularly important for lossy compression algorithms for video, image, voice, etc., because the reconstructed data differ from the original ones and the human perceptual system is the ultimate judge of the reconstructed quality. For example, if there is no perceivable difference between the reconstructed data and the original ones, the compression algorithm can be claimed to achieve very high quality or high fidelity. The difference of the reconstructed data from the original ones is called the distortion. One expects to have higher quality of the reconstructed data, if the distortion is lower. Quality measures could be very subjective based on human perception or can be objectively defined using mathematical or statistical evaluation. Although there is no single universally accepted measure of the quality met-
14 INTRODUCTION TO DATA COMPRESSION
rics, there are different objective and subjective quality metrics in practice to evaluate the quality of the compression algorithms.
22.214.171.124 Subjective Quality Metric Often the subjective quality metric is defined as the mean observers score (MOS). Sometimes, it is also called mean opinion score. There are different statistical ways to compute MOS. In one of the simplest ways, a statistically significant number of observers are randomly chosen to evaluate visual quality of the reconstructed images. All the images are compressed and decompressed by the same algorithm. Each observer assigns a numeric score to each reconstructed image based on his or her perception of quality of the image, say within a range 1-5 to describe the quality of the image—5 being the highest quality and 1 being the worst quality. The average of the scores assigned by all the observers to the reconstructed images is called the mean observer score (MOS) and it can be considered as a viable subjective metric if all the observers evaluate the images under the same viewing condition. There are different variations of this approach to calculate MOS—absolute comparison, paired comparison, blind evaluation, etc.
The techniques of measurement of the MOS could well be different for different perceptual data. The methodology to evaluate the subjective quality of a still image could be entirely different for video or voice data. But MOS is computed based on the perceived quality of the reconstructed data by a statistically significant number of human observers.
126.96.36.199 Objective Quality Metric There is no universally accepted measure for objective quality of the data compression algorithms. For objective measure, the most widely used objective quality metrics are root-mean-squared error (RMSE), signal-to-noise ratio (SNR), and peak signal-to-noise ratio (PSNR). If I is an M x N image and I is the corresponding reconstructed image after compression and decompression, RMSE is calculated by
1=1 J = 1
where i,j refer to the pixel position in the image. The SNR in decibel unit (dB) is expressed as SNR =
In case of an 8-bit image, the corresponding PSNR in dB is computed as
PSNR = m»S.. (^),
OVERVIEW OF IMAGE COMPRESSION
where 255 is the maximum possible pixel value in 8 bits.
It should be noted that a lower RMSE (or equivalently, higher SNR or PSNR) does not necessarily always indicate a higher subjective quality. These objective error metrics do not always correlate well with the subjective quality metrics. There are many cases where the PSNR of a reconstructed image can be reasonably high, but the subjective quality is really bad when visualized by human eyes. Hence the choice of the objective or subjective metrics to evaluate a compression and decompression algorithm often depends on the application criteria.
Similar objective quality metrics are used for audio and speech signals as well.
1.6.3 Coding Delay
Coding delay is another performance measure of the compression algorithms where interactive encoding and decoding is the requirement (e.g., interactive videoteleconferencing, on-line image browsing, real-time voice communication, etc.). The complex compression algorithm might provide a better amount of compression, but it could lead to increased coding delay, prohibiting the interactive real-time applications. The constraint to the coding delay often forces the compression system designer to use a less sophisticated algorithm for the compression system.