NG-GAN: A Robust Noise-Generation Generative Adversarial Network for Generating Old-Image Noise

Area of Science:

• Digital Image Processing
• Computer Vision
• Artificial Intelligence

Background:

• Old images and videos suffer from unique noise patterns due to poor storage conditions.
• Supervised denoising methods require costly and difficult-to-obtain noisy-clean image pairs.
• Existing denoising techniques struggle with the diverse noise characteristics of historical media.

Purpose of the Study:

• To develop a method for effectively denoising old images and videos without requiring paired data.
• To address the challenge of acquiring paired datasets for training denoising models.
• To improve the quality and information content of degraded historical visual media.

Main Methods:

• Proposed a robust noise-generation generative adversarial network (NG-GAN) inspired by CycleGAN.
• Utilized unpaired datasets to learn and replicate the noise distribution of old images.
• Employed a perception-based image quality evaluator metric to control noise generation.
• Generated an unpaired dataset using clean images with features matching old images for model training.

Main Results:

• The dataset generated by NG-GAN effectively trains state-of-the-art denoising models for old videos.
• Denoising models trained on NG-GAN data showed significant improvements in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM).
• Average improvements of 0.37 dB in PSNR and 0.06 in SSIM were observed.

Conclusions:

• The proposed NG-GAN successfully generates realistic noise distributions for training denoising models on unpaired data.
• This approach overcomes the limitations of supervised learning for historical image and video restoration.
• The method offers a practical and effective solution for enhancing the quality of degraded visual archives.