SVCNet: Scribble-Based Video Colorization Network With Temporal Aggregation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Scribble-based video colorization aims to add color to monochrome videos using user-provided scribbles.
• Existing methods often struggle with colorization vividness, temporal consistency, and color bleeding artifacts.

Purpose of the Study:

• To propose SVCNet, a novel scribble-based video colorization network with temporal aggregation.
• To address and improve upon common issues in video colorization, specifically vividness, temporal consistency, and color bleeding.

Main Methods:

• SVCNet employs two sequential sub-networks for precise colorization and temporal smoothing.
• The first stage uses pyramid and semantic feature encoders; the second stage aggregates temporal information from neighboring and the first frames.
• Simultaneous learning of video colorization and segmentation minimizes color bleeding, with a Super-resolution Module handling varied resolutions.

Main Results:

• SVCNet demonstrates superior performance on DAVIS and Videvo benchmarks compared to existing methods.
• Experimental results confirm higher-quality and more temporally consistent video colorization.
• The network effectively alleviates color bleeding and maintains stability across different video sequences.

Conclusions:

• SVCNet offers a robust solution for scribble-based video colorization, achieving state-of-the-art results.
• The proposed architecture effectively balances colorization quality and temporal coherence.
• The method's adaptability to different resolutions and its ability to mitigate common artifacts make it a valuable contribution to the field.