Subjective Coder Evaluation

Here we perform a comparison in subjective performance of the state of the art coder in progressive transmission SPIHT against RECON and REWIC, using a psychophysical experiment. To this aim, test image $\char93 24$ was firstly compressed to the same very low bit rates using the three compression methods. Fig.1 shows the respective reconstructions at 0.0156, 0.0312, 0.0625, and 0.08 bpp. Fifteen volunteers subjectively evaluated the reconstructed images  following an ITU-R Recommendation 500-10. Table I summarizes mean quality factors for reconstructions illustrated in Fig. 1. As can be seen from this table, quality factors predict a better visual fidelity using RECON than with the SPIHT reconstructions. We have that the visual quality of SPIHT decoded outputs is bad at 0.0625, 0.03125 and 0.015625 bpp. Whereas, the visual quality of RECON reconstructions is bad only at 0.015625 bpp.

Figure 1: Reconstructions of the test image # 24 using the SPIHT, REWIC, and RECON, at 0.08, 0.0625, 0.03125 and 0.015625 bit/pixel.
\includegraphics[height=22cm]{fig/Fig8.eps}


Table I:
$bit/pixel$ MEAN QUALITY FACTOR
 
     SPIHT           REWIC      RECON
0.015625 1.00 1.07 1.33
0.03125 1.27 2.07 2.07
0.0625 1.80 2.27 2.47
0.08 2.33 2.93 2.73
$MEAN$ 1.60 2.09 2.15


The RECON coder provides a method to prioritize, at very low bit rate, information adjacent to noise or insignificant detail. This is illustrated in Fig. 2. This figure shows the original image that contains regions of noise-like texture and a digitized mammographic section with clusters of fine, granular microcalcifications. It also shows reconstructions at 0.0625, 0.03125 and 0.015625 bpp using SPIHT. Table V provides the corresponding mean quality factors: SPIHT reconstructions exhibit bad visual fidelity at 0.0625, 0.03125 and 0.015625 bpp, since noise data was incorrectly reconstructed before key data (e.g., granular microcalcifications) from mammographic section. Again fifteen volunteers subjectively evaluated the decoded outputs following an ITU-R Recommendation 500-10.

Fig. 2 also shows the reconstructions using REWIC and RECON at the same bit rates. Table II summarizes mean quality factors for the decoded outputs. As can be seen from this table, quality factors predict a better visual fidelity using RECON and REWIC than with the SPIHT reconstructions. RECON output is bad only at 0.015625 bpp whereas it is poor at 0.03125 bpp and fair at 0.0625 bpp.




Figure 2: Reconstructions of the original image using SPIHT, REWIC and RECON, at 0.0625, 0.03125 and 0.015625 bit/pixel.
\includegraphics[height=22cm]{fig/Fig9.eps}


Table II:
$bit/pixel$ MEAN QUALITY FACTOR
 
SPIHT REWIC RECON
0.015625 1.00 1.53 1.73
0.03125 1.33 2.00 2.53
0.0625 1.40 2.40 3.00
$MEAN$ 1.24 1.98 2.42