Prototypical Distribution Divergence Loss for Image Restoration.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Neural networks, particularly Convolutional Neural Networks (CNNs) and Transformers, have advanced image restoration.
• Loss functions are crucial for training image restoration networks but have received limited attention.
• Existing loss functions often rely on semantic or hand-crafted image representations.

Purpose of the Study:

• To explore the effectiveness of discrete representations as a loss function for image restoration.
• To propose a novel loss function based on discrete representations for improved image restoration.
• To enhance the performance of existing image restoration architectures using the proposed loss function.

Main Methods:

• Proposed a Local Residual Quantized Variational AutoEncoder (Local RQ-VAE) to learn discrete prototype vectors from high-quality images.
• Developed a Prototypical Distribution Divergence (PDD) loss to measure the difference between the distributions of restored and target images.
• Integrated the PDD loss with state-of-the-art CNNs and Transformers for various image restoration tasks.

Main Results:

• The PDD loss significantly improved image restoration quality across multiple tasks.
• Enhanced performance was observed in terms of both Peak Signal-to-Noise Ratio (PSNR) and visual fidelity.
• The proposed loss demonstrated effectiveness with both CNN and Transformer architectures.

Conclusions:

• Discrete representations offer a promising direction for designing effective loss functions in image restoration.
• The PDD loss provides a robust method for improving image restoration by leveraging discrete image representations.
• The proposed approach offers a valuable contribution to the field of image restoration, enhancing existing models.