A novel quantization parameter estimation model based on neural network
Zhu Jianying, Zhelei Xia, Haibing Yin, Qiang Hua · 2012
In the video lossy compression, quantitative parameter (QP) has a great influence on compression efficiency and image quality. This paper proposes a QP prediction scheme, which use artificial neural network (ANN) model combining with H.264 rate-distortion mode. Experiments pick five main parameters that affect QP most, and then establish a multilayer error feed-forward neural network to output QP on frame layer immediately. The high prediction ability and robustness of neural network improve QP forecast in H.246 encoder. Experimental results show less Peak Signal-to-Noise Ratio (PSNR) fluctuation and same Rate-Distortion (R-D) performance, using proposed scheme instead of quantization model in JM14.2 reference software.