Perceptual Adversarial Networks for Image-to-Image Transformation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image-to-image transformation is crucial for various applications.
• Existing methods are often application-specific.
• A generic framework for learning image mappings is needed.

Purpose of the Study:

• To introduce Perceptual Adversarial Networks (PAN) as a generic framework for image-to-image transformations.
• To develop a novel perceptual adversarial loss for improved image translation.
• To demonstrate the effectiveness of PAN across diverse image transformation tasks.

Main Methods:

• Proposed PAN framework comprising two feed-forward convolutional neural networks (CNNs): a transformation network (T) and a discriminative network (D).
• Integrated generative adversarial loss with a novel perceptual adversarial loss.
• Employed an adversarial training process between network T and the hidden layers of network D.

Main Results:

• PAN effectively handles diverse image-to-image tasks, including de-raining, edge-to-photo, and semantic label to scene synthesis.
• The perceptual adversarial loss enhances the discrepancy detection between transformed and ground-truth images.
• PAN demonstrates advantages over existing image-to-image transformation methods.

Conclusions:

• PAN provides a powerful and flexible framework for learning image-to-image mappings.
• The proposed perceptual adversarial loss significantly improves transformation quality.
• PAN offers a promising solution for a wide range of computer vision applications.