Knowledge Distillation for Face Photo-Sketch Synthesis.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks, especially generative adversarial networks (GANs), have advanced face photo-sketch synthesis.
• Existing methods are hindered by a scarcity of paired photo-sketch training data.

Purpose of the Study:

• To investigate the efficacy of knowledge distillation (KD) for training neural networks in face photo-sketch synthesis.
• To propose an effective KD model to enhance the performance and quality of synthetic facial images.

Main Methods:

• Utilized a teacher network trained on a large dataset to transfer knowledge of face photos and sketches to two student networks.
• Implemented mutual knowledge transfer between student networks for enhanced learning.
• Developed a KD+ model combining GANs with KD to improve texture realism and reduce noise in synthetic images.

Main Results:

• The proposed KD and KD+ models demonstrated superior performance in face photo-sketch synthesis.
• Extensive experiments and user studies confirmed the effectiveness of the developed models.
• The KD+ model, integrating GANs and KD, produced synthetic images with improved perceptual quality, realistic textures, and reduced noise.

Conclusions:

• Knowledge distillation is a viable approach to address data limitations in face photo-sketch synthesis.
• The proposed KD and KD+ models significantly outperform state-of-the-art methods.
• The integration of GANs with KD offers a promising direction for generating high-fidelity synthetic facial images.