Comparative Analysis Overview of CNN-Driven In-Loop Filtering Approaches in H.266/VVC

Video compression standards such as HEVC/H.265 and VVC/ H.266 have a hybrid coding framework consisting of various modules, including in-loop filters (ILFs), to reduce data size while maintaining video quality. However, the default ILFs in these standards have limitations and may not adapt well to complex natural videos. This review paper explores the use of Convolutional Neural Networks (CNNs) to enhance the performance of ILFs in video compression. CNN-based ILFs are categorized based on their position: replacing default ILFs, placed after existing ILFs, or inserted between default ILFs. They are also categorized by training setups: QP combined and separate models, frame combined and separated models, and channel combined and separated models. The paper provides a comparative analysis of recent CNN-based ILFs regarding their extra inputs, trained channels, Y/BD rate (PSNR), encoding/decoding time, and other metrics, including subjective quality analysis and complexity analysis. The reviewed CNN-based ILFs offer promising results in terms of video compression efficiency and quality enhancement, demonstrating the potential for further advancements in this field. However, there are still challenges to address, including the trade-off between compression efficiency and complexity, the integration of multiple deep tools, and more. This research highlights the importance of CNN-based ILFs in improving video compression standards and their potential for future developments in video coding.