Research On Anti-Noise Matching Criteria Of Motion Estimation

ME(motion estimation) is a key technic in video coding. It's used to eliminate temporal redundancy in video sequences. Encoders employ ME to get the most matched one from searching windows for target block prediction. The more matched, the less compressed the data is. So, matching criteria which plays the role of similarity estimator is determinant to ME.Noise brings challenges into ME. It leads to higher probability that motion vectors deviate from their optimized convergence. Noise increases residuals and then lowers the compression ratio. Moreover, deviation makes the field of motion vectors no longer reflects the true motion of objects which affects application such as video inspector, etc. As mentioned above, noise must be taken into consideration for motion estimation.This paper studies the characteristics of noise and proposes two novel matching criteria to eliminate noise.Residues are firstly transferred into frequency domain by DCT, then quantized and finally entropy coded. Obviously, a matching criterion with DCT"friendly"will bring better coding efficiency. Research shows, distribution of DCT coefficients is related to the variance of spatial domain: the smaller the variance is, the more compact the coefficients are. Some people proposed smooth constraint matching criteria based on this conclusion. However, the smooth estimator MME they proposed takes only 2 points and sensitive to noise, so we further modified the smooth estimator by absolute mean difference of multi-points to enhance the anti-noise feature. Simulation results show that, advanced algorithm improves about 0.2 dB PSNR gain for videos with variance 8 gauss white adding noise.MAD is a balance of precision and complexity, it can be considered as arithmetic mean of absolute residual. In the case of equi-normal distribution, MAD equals to mathematic expectation. Random Process theory has proved that when the degree of confidence is fixed, confidence interval length of mathematic expectation is inversely proportional to the square roots of sample points. In this paper, we propose a novel motion splitting based adaptive OBME algorithm which uses motion vector as the base of motion splitting makes the extension of super-blocks more reasonable. Super-block increases the samples and provides a more accurate confidence interval. We also validate the 3 hypothesis of our algorithm. Simulation shows the adaptive OBME improves 0.1~1.2dB PSNR gain under the same bit-rate comparing with the reference algorithms.