Synthesizing Camera Noise Using Generative Adversarial Networks

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Digital photography often suffers from noise, which can degrade image quality and hinder subsequent analysis.
• Controlling noise levels, particularly matching specific ISO settings, is crucial for consistent image quality and performance in various applications.

Purpose of the Study:

• To develop a novel technique for synthesizing realistic noise in digital photographs.
• To enable adjustment of noise levels in images to match a target ISO setting, either increasing or decreasing existing noise.

Main Methods:

• Utilized generative adversarial networks (GANs) to learn mappings between different ISO levels.
• Trained the model on unpaired image data, allowing flexibility in noise synthesis.
• Employed Kullback-Leibler divergence and Kolmogorov-Smirnov tests for quantitative evaluation.

Main Results:

• Successfully synthesized realistic noise, allowing precise control over image noise levels to match target ISO values.
• Demonstrated significant improvement in the performance of a state-of-the-art trainable denoising method when using the synthesized noise.
• Validated effectiveness through both quantitative metrics and qualitative assessments on numerous examples.

Conclusions:

• The proposed technique offers a robust method for realistic noise synthesis and ISO level matching in digital images.
• This approach has practical applicability in enhancing computer vision systems that require robustness to varying noise conditions.