Multi-stage Attention Convolutional Neural Networks for HEVC In-Loop Filtering
Peng-Ren Lai, Jia-Shung Wang · 2020
In High Efficiency Video Coding (HEVC), lossy compression techniques bring various visual artifacts such as blocking, ringing, and blurring, etc. Recently, to explore convolutional neural networks (CNN) potential of removing such artifacts, researchers presented several CNN models. In this paper, we proposed a Multi-stage Attention Convolutional Neural Network (MACNN) to replace in-loop filters or postprocessing. Our proposed method has three characteristics of effective training and inferencing: employing a 3-stage design fashion with various loss criteria, adopting the revised Inception module, and utilizing Self-attention block. Such a design has several advantages. The 3-stage design clarifies the function of each stage, effectively alleviating the burden of the network; inception module along with self-attention block benefit our model having the ability to acquire global information with fewer layers. Hence, MACNN could reach the performance gain competitive to deeper models. The experimental results show that our MACNN achieves an average 6.5% BD-rate reduction compared to HEVC in all-intra configuration, beating state-of the-art methods.