Fast Lossy Internet Image Transmission

John M. Danskin, Geoffrey M. Davis, Ziyong Song, Dartmouth College

One-line summary

Subband compression that dynamically trades redundancy bits for data bits, to deliver image data using unreliable (droppable) datagrams, like video. Sacrifices some quality to get much better performance (latency to get "entire" image) and eliminate multi-second TCP variances.

Main ideas

• Concentrate on "perceptually important" image bits. When loss rate increases, shift bits from data to redundancy but keep total # bits transmitted constant.
• Optimization: remember effective transfer rates across connections to eliminate slow start effects (since UDP is used anyway).
• Doesn't rely on network traffic prioritization or delivery classes.
• Main problem is how to allocate quantizer bins to get minimal distortion for given bit expenditure. Channel coder has access to image source data at application level, so it knows which bits can be "sacrificed" in lossy transmission.
• Assuming integrity of packets and that lost packets can be detected, a single parity block allows one lost block out of N to be reconstructed. Packets are sent in random order to avoid burst-error lossage wiping out a whole block.
• Similarly, each subband is distributed across many packets (pixel interleave).
• Experiments: knee of load/loss curve for network was determined offline and used to determine target bit rate; plan to incorporate rate control like that in TCP Vegas (remember connection bandwidth for various subnets).
• Results: performs very well even with high error rates and jitter. Encoding and decoding are linear time.
• "Network citizen": FLIIT doesn't slow down (as TCP does) when congestion happens, but TCP Reno forces dropped packets in this situation, so not clear who's the better network citizen.

Comments/Flaws

• Destruction of streamlike model may render some integrated-layer strategies unworkable, esp. to reduce extra copies.
• Nonstandard encoding format, based on wavelets (but authors claim their techique can be applied to DCT based encodings too).

Back to index