Deinterlacing Based on BasicVSR

Abstract

Deinterlacing is a critical technique in the field of digital video processing, aimed at converting interlaced videos into a progressive format. Interlaced video, a legacy from the analog television era, was developed to optimize bandwidth and storage constraints in earlier broadcast systems. It displays alternating lines of a frame in two separate fields, which can result in visual artifacts such as flickering and blurring when viewed on modern progressive displays. Due to the demand for superior video quality, deinterlacing technology has evolved, incorporating sophisticated algorithms and methodologies to address the artifacts and deficiencies inherent in interlaced video. In recent years, methods based on deep learning have achieved significant advancements.

Video super-resolution is another video restoration technique that aims to enhance the resolution of video frames by generating high-quality, detailed outputs from low-resolution inputs. This process focused on improving the clarity and sharpness of video content. While both deinterlacing and video super-resolution utilize temporal information to improve video quality, the latter has more extensive research and broader applications in the deep learning field.

By leveraging the advanced architectures and techniques developed for video super-resolution, deinterlacing models can be equipped with a robust framework for aggregating multiple misaligned frames. Based on a streamlined video super-resolution (VSR) baseline, BasicVSR, we propose a novel reconstruction technique for deinterlacing, and redesign the backbone network to prevent misalignment between different fields. As a result, we present a concise, efficient and versatile video deinterlacing baseline that achieves state-of-the-art performance, surpassing the second-best method by 1.18 dB on the Vimeo90K dataset. Additionally, we extend the model to include super-resolution functionality, enabling it to perform both deinterlacing and super-resolution tasks.