Face Sketch Synthesis by Multidomain Adversarial Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Face sketch synthesis aims to map facial photographs to sketches.
• Existing methods struggle with synthesizing high-quality sketches, particularly when dealing with blur and deformations.
• Generative adversarial training offers new possibilities for face sketch synthesis.

Purpose of the Study:

• To develop a novel face sketch synthesis method that overcomes limitations of existing approaches.
• To achieve high-quality sketch synthesis from photographs, effectively addressing blur and deformation issues.
• To introduce a multidomain adversarial learning (MDAL) framework for improved face sketch generation.

Main Methods:

• The proposed method, multidomain adversarial learning (MDAL), utilizes an "interpretation through synthesis" principle.
• It involves adversarial learning for self-reconstruction in both photo and sketch domains, defining a latent domain.
• Adversarial learning is employed to ensure indistinguishable latent variable distributions between photo and sketch reconstructions.
• A latent variable from a reconstructed photograph is used to synthesize the corresponding sketch.

Main Results:

• The MDAL method demonstrates superior performance compared to state-of-the-art techniques in quantitative comparisons.
• It effectively synthesizes high-quality face sketches, mitigating problems associated with blur and deformations.
• The approach successfully generates realistic sketches from input photographs.

Conclusions:

• The proposed multidomain adversarial learning (MDAL) method represents a significant advancement in face sketch synthesis.
• MDAL effectively overcomes the challenges of blur and deformation, leading to high-quality sketch generation.
• The method's superiority is validated through quantitative comparisons, highlighting its potential for practical applications.