Graph-Regularized Locality-Constrained Joint Dictionary and Residual Learning for Face Sketch Synthesis

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Face sketch synthesis is vital for digital entertainment and law enforcement.
• Existing methods struggle to generate rare, individual-specific facial characteristics.
• Bridging the texture discrepancy between face photos and sketches remains a challenge.

Purpose of the Study:

• To propose a novel face sketch synthesis approach using residual learning.
• To improve the generation of individual-specific facial features (rare characteristics).
• To enhance the optimization process for face sketch synthesis.

Main Methods:

• Utilized residual learning to predict the difference (residual) between face photos and sketches.
• Developed a joint dictionary learning algorithm preserving local geometry.
• Transformed face sketch synthesis into a compact space using learned dictionary atoms.

Main Results:

• The proposed residual learning method demonstrated impressive performance on three public datasets.
• Achieved superior face sketch synthesis compared to state-of-the-art and deep learning methods.
• Effectively generated rare characteristic information crucial for individual identification.

Conclusions:

• Residual learning offers a more effective approach to face sketch synthesis than direct mapping.
• The joint dictionary learning enhances the synthesis by preserving data structure.
• The method shows significant potential for real-world applications in law enforcement and digital entertainment.