Region Assisted Sketch Colorization

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Automatic sketch colorization is challenging due to its ill-posed nature.
• Existing methods suffer from limited color patterns and artifacts like color bleeding.
• Traditional convolutional structures inadequately utilize regional sketch information.

Purpose of the Study:

• To propose a novel method, Region-Assisted Sketch Coloring (RASC), for improved sketch colorization.
• To address limitations of existing methods by enhancing regional feature utilization.
• To reduce artifacts and expand color patterns in generated images.

Main Methods:

• Introduced an intermediate 'Region Map' representation to capture sketch's regional information.
• Employed a sketch encoder to extract hierarchical feature maps.
• Utilized a coarse-to-fine decoder with Region-based Modulation (RM) blocks and a Region Formulation module.

Main Results:

• The RASC method effectively enhances perception of region-wise features beyond pixel-wise features.
• Accurately captures inner-region local style and inter-region global context dependency.
• Achieved superior performance over state-of-the-art methods on synthetic and real sketch datasets, with fewer artifacts and diverse color patterns.

Conclusions:

• The proposed RASC method significantly advances automatic sketch colorization.
• Explicitly formulating regional information is key to overcoming existing challenges.
• RASC offers a promising direction for generating more realistic and artifact-free colorized sketches.