Fast Intra Mode Decision Based on Edge Information and Reference Samples Similarity in HEVC

Abstract

Abstract High Efficiency Video Coding (HEVC) is a new video coding standard. The first version has been published on 2013. The second version has been published on 2015. HEVC includes many new features, such as format range extensions, scalable coding extensions and multi-view extensions. HEVC adopts many advanced tools and therefore achieves 2 times compression ratio compared to the previous standard, H.264. However, the encoding time becomes much longer due to the increased complexity. There are many related works that aim to reduce the encoding time. This thesis presents a novel algorithm which can accelerate the intra coding process. The proposed algorithm works before the rough mode decision (RMD) stage. This method utilizes the characteristics of both the original block and the reference samples to find the optimal intra prediction mode earlier. By calculating the dominant edge direction in the block, only the angular modes with the similar direction are selected and tested in the RMD stage. Afterward, by evaluating the similarity among the reference samples, only one of the intra modes, which have the similar reference samples, is remained to be tested in the RMD stage. Mean and variance of the reference samples are used to determine if two intra modes have similar reference samples. So, there are two thresholds in this algorithm, the mean threshold and the variance threshold. After ruling out the intra modes by the two methods, only a part of the intra modes are processed in the RMD stage. The performance of the proposed algorithm can be adjusted by two thresholds. Generally, bigger thresholds can provide more encoding time reduction, but the video quality will get worse and the bit rate will increase at the same time. Since the number of intra prediction modes which enter the RMD stage is reduced. The number of intra prediction modes which enter the Rate Distortion Optimization (RDO) stage also reduced. The encoding time is therefore further reduced. As shown in the experimental results, when the mean threshold equals 2 and the variance threshold equals 200, the proposed algorithm can reduce 28.83% encoding time with 3.66% BD rate increase and 0.19dB BD PSNR-Y decrease on average.