Distortion Estimation Algorithm For H.264/AVC Multiple Reference Frames

The two basic component of rate-distortion model based on Lagrange multiplier in video coding are distortion and rate. The rate of a frame is the bits used to encode it, but the distortion can be evaluated only after the decoding process. Therefore, how to estimate precisely the distortion between the codec in error prone environments in order to choose the coding mode is the key problem of rate-distortion. The reference software model of H.264/AVC assumes there is no error when transmitting the video signals and thus ignore totally the distortion might be caused by transmission and decoding process, and use the distortion on encoder side as the total distortion. But the real transmission channel is never lossless and probably has a loss rate of as high as 20%, also the codes produced by prediction encoding is very sensitive to lossy channel. Hence how to estimate the distortion under packet-switch networks, so as to reduce the transmission errors as less as possible, is a big problem to be solved in video coding.Three algorithms estimating the distortion are compared in this thesis, experimental result shows that the recursive optimal per-pixel estimate(ROPE) algorithm is the best. Based on the ROPE algorithm, this thesis gives a detail analysis of error propagation under lossy channel and takes consideration on all three video communication process: encoding, transmission and decoding, and deduces a general distortion-estimating algorithm for H.264/AVC using multiple reference frames. The proposed algorithm separates the total end-to-end distortion into three parts: quantization distortion by encoder, propagation distortion caused by transmission channel and concealment distortion by decoder, where the quantization distortion can be computed after encoding, the concealment distortion can be computed by simulated decoding in encoder side, and the propagation distortion can be computed recursively after deduction and transform. The correlative coefficient is weighted against a motion vector term in the estimation model, which reduces the computational complexity of the algorithm in a certain degree. Experimental result based on reference software...