Predicting split decisions of coding units in HEVC video compression using machine learning techniques

Abstract

In this work, we propose to reduce the complexity of HEVC video encoding by predicting the split decisions of coding units. We use a sequence-dependent approach in which a number of frames belonging to the video being encoded are used for generating a classification model. At each coding depth of the coding units, features representing the coding unit at that particular depth are extracted from both the present and previously encoded coding units. The feature vectors are then used for generating a dimensionality reduction model and a classification model. The generated models at each coding depth are then used to predict the split decisions of subsequent coding units. Stepwise regression, random forest reduction and principal component analysis are used for dimensionality reduction; whereas, polynomial networks and random forests are utilized for classification. The proposed solution is assessed in terms of classification accuracy, BD-rate, BD-PSNR and computational time complexity. Using seventeen video sequences with four different classes of resolution, an average classification accuracy of 86.5% is reported for the proposed classification system. In comparison to regular HEVC coding, the proposed solution resulted in a BD-rate loss of 0.55 and a BD-PSNR of −0.02 dB. The average reported computational complexity reduction is found to be 39.2%.