Real-World Image Denoising with Deep Boosting.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Traditional image denoising methods struggle with complex, real-world noise.
• Boosting algorithms offer a framework for iterative improvement but often rely on handcrafted features.

Purpose of the Study:

• To develop an advanced image denoising framework integrating deep learning with boosting.
• To address the limitations of conventional boosting units and improve denoising performance and efficiency.

Main Methods:

• A Deep Boosting Framework (DBF) was proposed, replacing handcrafted units with convolutional neural networks.
• A lightweight Dense Dilated Fusion Network (DDFN) was designed to mitigate gradient vanishing and optimize parameters.
• A new Real-world Image Denoising (RID) dataset was created for training and evaluation.
• A one-shot domain transfer scheme was developed to handle domain shift issues.

Main Results:

• The DBF demonstrated superior performance in simulation tasks like Gaussian denoising and JPEG deblocking.
• Experiments on the RID dataset showed significant improvements in real-world image denoising.
• The proposed method outperformed existing techniques on widely used benchmarks.

Conclusions:

• The Deep Boosting Framework offers a powerful and efficient solution for real-world image denoising.
• The integration of deep learning into boosting significantly advances the state-of-the-art in image restoration.
• The developed dataset and domain transfer scheme facilitate future research in learning-based denoising.