Wavelet-Based Topological Loss for Low-Light Image Denoising.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Current image denoising methods often assume Additive White Gaussian Noise (AWGN), which is inaccurate for real-world distortions.
• Supervised learning for denoising heavily relies on high-quality and diverse training datasets.

Purpose of the Study:

• To develop a novel image denoising framework that addresses real noise distortions by incorporating image structure and spatial information.
• To propose a new loss function that utilizes topological invariants and textural information for improved denoising performance.

Main Methods:

• A novel denoising loss function was developed, integrating topological invariants and textural information from the wavelet domain.
• State-of-the-art denoising models were trained using the BVI-Lowlight dataset, which contains diverse real noise patterns.
• The proposed loss function was evaluated by comparing its performance against common loss functions using the Learned Perceptual Image Patch Similarity (LPIPS) metric.

Main Results:

• The novel loss function significantly improved the LPIPS metric by up to 25% compared to conventional methods.
• Neural networks trained with the proposed loss function demonstrated a better ability to learn noise characteristics.
• The method successfully extracted topological features of noise-free images, enhancing contrast and preserving textural details.

Conclusions:

• The proposed loss function effectively handles real noise distortions in images, outperforming traditional approaches.
• Incorporating topological and textural information into the denoising process leads to superior image reconstruction quality.
• This framework offers a promising direction for developing more robust and accurate image denoising algorithms for real-world applications.