Improved GOP-Based Temporal Filter for Low Illumination Scene Video

Modern hybrid video coding frameworks aim to compress video efficiently by exploiting the temporal and spatial correlations within the video signal. However, the presence of noise within the video signal can disrupt this correlation, leading to reduced compression efficiency. Noise is particularly noticeable in low-light video sequences, and most algorithms perform poorly in such scenarios. To mitigate the impact of noise on video compression, pre-processing filtering algorithms have been proposed by video encoders such as VP9, HEVC and AV1. Among these algorithms, the GOP-Based Temporal Filter (GBTF), which was first introduced in High Efficiency Video Coding (HEVC) and inherited by Versatile Video Coding (VVC), uses bilateral filtering as its core weight function to remove noise while preserving high-frequency information, ensuring clear filtered images. In this paper, we propose a mean-removal improvement scheme for the GBTF core motion search algorithm, specifically designed for low-light video sequences, which improves performance by an additional −0.46% to −1.88% on top of the original BD-BR. Furthermore, our improvement scheme almost does not increase the complexity of the original GBTF algorithm and is easy to modify.