Quality Metric Guided Portrait Line Drawing Generation From Unpaired Training Data.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Art

Background:

• Generating face portrait line drawings is challenging due to the need for paired training data, which is expensive and time-consuming.
• Existing unpaired image-to-image translation methods often fail to preserve crucial facial features in generated drawings.

Purpose of the Study:

• To develop a novel method for automatically transforming face photos into high-quality portrait drawings using unpaired training data.
• To enable the generation of drawings in multiple styles and even novel, unseen styles.
• To address the loss of facial features common in current unpaired translation techniques.

Main Methods:

• Proposed a novel quality metric for portrait drawings learned from human perception.
• Introduced a quality loss function to guide the generative network.
• Developed an asymmetric cycle mapping to ensure reconstruction information is visible and localized to relevant facial regions.
• Utilized localized discriminators for key facial areas to enhance feature preservation.

Main Results:

• The method successfully generates high-quality portrait drawings in various styles from a single network.
• It can create drawings in styles not present in the training data.
• Preserves all important facial features, overcoming limitations of existing unpaired methods.
• User studies confirm superior performance compared to state-of-the-art techniques.

Conclusions:

• The proposed method offers an effective and flexible approach for automatic face portrait drawing generation from photos.
• It significantly advances unpaired image-to-image translation for artistic style transfer, particularly for portraits.
• The integration of perceptual quality metrics and localized feature preservation is key to its success.