Recursive Copy and Paste GAN: Face Hallucination From Shaded Thumbnails.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Existing convolutional neural network (CNN) based face hallucination methods perform poorly on low-resolution (LR) faces under non-uniform illumination.
• Non-uniform illumination significantly degrades the performance of current face hallucination techniques.

Purpose of the Study:

• To propose a novel Recursive Copy and Paste Generative Adversarial Network (Re-CPGAN) for generating high-resolution (HR) face images from LR inputs, specifically addressing non-uniform illumination conditions.
• To develop a method that recovers authentic HR faces while simultaneously compensating for illumination variations.

Main Methods:

• The proposed Re-CPGAN incorporates two key components: internal and recursive external Copy and Paste networks (CPnets).
• The internal CPnet utilizes facial self-similarity for detail enhancement.
• The recursive external CPnet employs multiple external Copy and Paste (EX-CP) units for progressive illumination compensation and detail enhancement in a coarse-to-fine manner, using an external guided face.

Main Results:

• The Re-CPGAN method successfully generates authentic HR face images with a 16× magnification factor under uniform illumination.
• The method demonstrates superior performance compared to state-of-the-art techniques, both qualitatively and quantitatively.
• A novel illumination compensation loss function was developed and integrated effectively.

Conclusions:

• The Re-CPGAN effectively addresses the challenge of face hallucination under non-uniform illumination by progressively compensating for illumination and enhancing details.
• The proposed method achieves state-of-the-art results in generating high-quality, high-resolution face images from low-resolution inputs with challenging lighting conditions.