An LBT image coder and decoder was created in Matlab, in order to compare the rate distortion performance to the LOT. Visual effects were also examined. The setup is exactly the same as the previous LOT experimental results, except that the LBT replaced the LOT.
The image 'peppers512x512` was used for all experiments. Again, the test was run with a signal dependent Z matrix and a signal independent Z matrix. The difference was not noticeable in the rate-distortion plane, so only signal independent Z results will be reported.
The following figure shows the LBT synthesis functions along with the corresponding LOT basis functions.. Notice that only the top two functions differ significantly from the LOT basis functions.
Figure 5: LOT basis functions (+) and LBT synthesis functions (*)
The difference becomes apparent upon a closer look of the two functions:
Figure 6: First 2 LOT basis functions (+) and LBT synthesis functions (*)
Now the synthesis functions decay much closer to zero in a smoother fashion. This will assure less discontinuities in reconstruction errors due to quantization. The following example shows the alleviation of blocking effects due to the smoother decay. The LBT typically leads to gains of about 1 dB over the DCT.
The following images are 0.1 bpp reconstructions. The leftmost image uses DCT coding, while the rightmost image uses LBT coding. The DCT has a PSNR of 24.8 dB while the LBT has a PSNR of 25.9, a 1.1 dB improvement.
Figure 7: Reconstructed Images: DCT on the left, LBT on the right
The following rate distortion curves show that the LBT has approximately a 1 dB PSNR improvement over the DCT, for low bit rates.
Figure 8: Rate Distortion Curves