Unsupervised Eyeglasses Removal in the Wild.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Existing eyeglass removal methods lack scalability and require specific systems for different eyeglass types.
• Conventional approaches often fail in uncontrolled environments ('in the wild') and with diverse eyeglasses (rimless, full-rim, sunglasses).
• Previous work typically focuses on frontal face images under controlled laboratory conditions.

Purpose of the Study:

• To propose a unified and scalable model for eyeglasses removal that works with various eyeglass types in uncontrolled environments.
• To develop a method that does not rely on dense eyeglass location annotations, utilizing large-scale face images with weak annotations.
• To simultaneously address eyeglass removal and wearing tasks by learning to swap eye regions between faces with and without glasses.

Main Methods:

• Introduction of the eyeglasses removal generative adversarial network (ERGAN), a unified model for handling diverse eyeglasses.
• Simultaneous learning of eyeglass removal and wearing tasks by focusing on swapping the eye area between paired images.
• Leveraging large-scale face datasets with weak annotations, avoiding the need for precise eyeglass localization.

Main Results:

• ERGAN achieves competitive eyeglass removal quality, demonstrating high realism and diversity in generated eye regions.
• The model successfully handles various types of eyeglasses, including rimless, full-rim, and sunglasses, in real-world images.
• Experimental evaluation shows ERGAN's effectiveness as a preprocessing step for subsequent tasks like face verification and facial expression recognition.

Conclusions:

• The proposed ERGAN model offers a scalable and unified solution for eyeglass removal in diverse, uncontrolled environments.
• ERGAN's ability to handle different eyeglass types and its utility in downstream tasks highlight its practical applicability.
• The method advances the field by enabling robust eyeglass removal without requiring dense annotations, paving the way for broader applications.