Improvement of feature preservation in high efficiency video coding

Abstract

Feature information is known as interest points (keypoints) in an image which is very useful information for video analytic functions such as object detection and tracking, video classification, etc. Due to the characteristic of video coding, it is exploited following human vision system where high frequency component could be removed for better compression. Thus, it can affect to the keypoints, which mostly are the edge information. As a result, with less number of keypoints left, this affects the accuracy of video analytics. To solve this problem, this thesis presents an algorithm to preserve feature information of reconstructed video in the high efficiency video coding (HEVC). Scale-Invariant Feature Transform (SIFT) is chosen to extract the keypoints from raw video sequence. We then consider keypoints as an indicator of the importance of the largest coding unit (LCU). Adaptive LCU selection is defined to determine LCU into two different groups, important LCU group (IMLCU) and non-important LCU group (Non-IMLCU). Moreover, two different bit allocations are generated in rate control to each group based on coding mode, Intra or Inter mode, to achieve the target bit rate and also to keep the feature information. The experimental results show that our proposed algorithm can maintain more keypoints compared to HEVC reference software at the same bitrate based on the peak signal-to-noise ratio (PSNR) and SIFT similarity computation.