Biphasic Face Photo-Sketch Synthesis via Semantic-Driven Generative Adversarial Network With Graph Representation...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Biphasic face photo-sketch synthesis is valuable for digital entertainment and law enforcement.
• Existing methods struggle with low-quality sketches and photograph variations, resulting in unnatural outputs.

Purpose of the Study:

• To develop a novel semantic-driven generative adversarial network (GAN) for high-fidelity biphasic face photo-sketch synthesis.
• To address limitations of previous global-view generation approaches.

Main Methods:

• Injecting class-wise semantic layouts into the generator for style-based spatial information.
• Constructing intraclass semantic graphs (IASG) and interclass structure graphs (IRSG) for enhanced facial detail and structural coordination.
• Employing a biphasic interactive cycle training strategy leveraging multilevel feature consistency.

Main Results:

• The proposed method significantly improves the quality and fidelity of synthesized face sketches.
• Graph representation learning enhances the realism of facial details and structural coherence.
• The approach outperforms state-of-the-art methods on the CUHK Face Sketch (CUFS) and CUFS FERET (CUFSF) datasets.

Conclusions:

• The semantic-driven GAN with graph representation learning offers a superior approach to biphasic face photo-sketch synthesis.
• The method effectively handles photograph variations and produces natural, high-fidelity results.