MASc seminar - Yi Shan

Tuesday, September 13, 2016 2:00 pm - 2:00 pm EDT (GMT -04:00)

Candidate

Yi Shan

Title

Fast Intra-frame Coding Algorithm for HEVC Based on TCM and Machine Learning

Supervisor

En-Hui Yang

Abstract

High Efficiency Video Coding (HEVC) is the latest video coding standard. Compared with the previous standard H.264/AVC, it can reduce the bit-rate by around 50% while maintaining the same perceptual quality. This performance gain on compression is achieved mainly by supporting larger Coding Unit (CU) size and more prediction modes. However, since the encoder needs to traverse all possible choices to mine out the best way to encode the data, this large flexibility on block size and prediction modes has tremendously jacked up the encoding time. In HEVC, intra-frame coding is an important basis, and it is widely used in all configurations. Therefore, fast algorithms are required to alleviate the encoding time of HEVC intra coding.

In our research, a machine learning based fast algorithm is proposed to predict the CU decisions for the HEVC encoder. Hence the computational complexity can be significantly reduced with negligible loss in the coding efficiency. Machine learning models like Bayes decision, Support Vector Machine (SVM) are used as decision makers while the Laplacian transparent composite model (LPTCM) is utilized as a feature extraction tool. In the main version of the proposed algorithm, features named with Summation of Binarized Outlier Coefficients (SBOC) are extracted to train SVM models. An online training structure and a performance control method have been proposed to enhance the robustness of decision makers.

When applied on All Intra Main (AIM) full test, the main version of the proposed algorithm can achieve, on average, 48% time reduction with 0.78% BD-rate increase. Also, through adjusting the setting, the algorithm can change the trade-off between encoding time and BD-rate, which can generate a performance curve to meet different requirements. By testing different methods on the same machine, the performance of proposed method has outperformed all CU decision based fast intra algorithm in the benchmarks.