Deep Latent Low-Rank Representation for Face Sketch Synthesis

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Face sketch synthesis is crucial for digital entertainment.
• Existing methods struggle with insufficient training data, losing identity-specific details like accessories and hairstyles.
• This limitation hinders the accurate representation of unique facial features in synthesized sketches.

Purpose of the Study:

• To propose a novel face sketch synthesis framework addressing data insufficiency.
• To enhance identity-specific information in synthesized sketches.
• To improve the quality and vividness of generated face sketches.

Main Methods:

• Developed a Deep Latent Low-Rank Representation (DLLRR) framework.
• DLLRR induces hidden training sketches with identity-specific information to supplement insufficient original data.
• Utilized a coupled autoencoder for its representational capability to reveal hidden data.

Main Results:

• The proposed DLLRR method successfully generates sufficient training data with identity-specific information.
• Experiment results demonstrate that DLLRR synthesizes cleaner and more vivid face sketches compared to state-of-the-art methods.
• The framework effectively preserves details such as glasses, earrings, and hairstyles.

Conclusions:

• The DLLRR framework provides a robust solution for face sketch synthesis, overcoming data limitations.
• This approach significantly enhances the quality and accuracy of synthesized sketches.
• The method holds promise for applications in digital entertainment and beyond.