Scientists from the Universidad Carlos III de Madrid and the University of Southern California (USC) have developed a new method for the compression of complex signals. The new method was evaluated and showed advantages over DCT-based compression methods, including H.264.
For TV broadcasters, compression efficiency determines how many programs can be transmitted in a given bandwidth and the quality of those programs. Broadcasters have to transmit signals using compression methods compatible with existing receivers—MPEG-2 for conventional ATSC broadcasting and H.264/AVC for ATSC Mobile DTV—however, as new receivers are developed (perhaps as software-defined radios), particularly for Mobile DTV, there may opportunities to incorporate new compression techniques such as this one.
The researchers used an analogy to explain how it works.
"We can see each image that makes up a video as a handful of sand spread out on the floor: each pixel is like a grain of sand. The objective of a transform for compression is to reorganize the grains of sand in various levels above the floor. In this way, on the lowest levels a few grains of sand would be used to draw the softest forms, like a sketch of the image. Moving upward, the rest of the grains of sand would be used and more details would be added. And if the structure collapsed, the grains would fall and the original image of the video would be drawn again."
One of the principal researchers, Eduardo Martinez Enrique, explained, "Once the grains of sand are reorganized in this manner, compressing is simple: we would remove levels from the top down as needed."
The technical details are presented in the paper Video Encoder Based on Lifting Transforms on Graphs. One of the illustrations in that paper (figure 5) compares PSNR at various bit-rates based on using DCT and the proposed compression method. From the graph, it appears that for the "Mobile" sequence at QCIF resolution, the proposed method provides slightly better than a 30 dB PSNR at 600 kbps. The DCT compression needed over 850 kbps to provide a similar PSNR. Using the "Foreman" sequence, DCT had the advantage at rates below 300 kbps (32 dB PSNR for both methods) but the proposed method did better at higher rates and better PSNR. For this sequence, above 35 dB SNR the proposed method provided the same quality around 50 kbps less bandwidth.
Another paper, Lifting Transforms on Graphs for Video Coding provides additional data and a good explanation of how the method works. As this paper points out, work continues on improving the method. The researchers note, "We are currently investigating how to use the graph information in the decoder to group together nonzero coefficients. Another interesting future line of research would be to design a low-complexity version of the transform that works with sub-graphs formed from the original graph without loss of performance."
