An Efficient Online Learning Approach for Super-Resolution in Dynamic Environments

Rongqing Liu, Zhiqiang Li, Feng Tian, Jie Zheng, Rui Cao, Chuanyuan Zhao, Jie Ren

Abstract: In the context of the growing demand for an enhanced viewing experience, driven by the popularity of 4K live video streaming, the necessity for a reliable and high-quality network environment and efficient encoding algorithm becomes paramount to facilitate smooth 4K streaming experiences. Pre-vious work explores the integration of Super-Resolution (SR) techniques within high-resolution video streaming frameworks, aiming to adaptively enhance streaming quality based on the video content and current network conditions. However, due to the complexity and diversity of environmental scenes, the SR model cannot consistently provide users with high-quality high-resolution images. Moreover, transfer learning for a single scene still incurs significant latency, making it unsuitable for live scenarios. Our approach employs an SR online learning algorithm to fine-tune the pre-trained SR model in response to variations in the video scene. We evaluate our approach on over 3000 4K video clips. Experimental results indicate that our approach outperforms state-of-the-art methods in terms of user perceptual Quality.