Deep RED Unfolding Network for Image Restoration

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Deep Unfolding Networks (DUNs) are efficient for image restoration, typically using Deep Convolutional Neural Networks (DCNNs) as regularization modules.
• Existing DUN models often perform data fitting before regularization in each iteration, limiting adaptability.

Purpose of the Study:

• To introduce a novel DUN framework (DRED-DUN) with an improved regularization module placed before the data fitting module.
• To enhance image restoration performance and detail recovery compared to existing methods.

Main Methods:

• Developed a new regularization module by combining Regularization by Denoising (RED) with a newly designed DCNN.
• Employed a closed-form solution with Faster Fourier Transform (FFT) for the data fitting module.
• Designed an end-to-end trainable DRED-DUN model for joint optimization of parameters.

Main Results:

• The DRED-DUN model demonstrated superior performance over state-of-the-art model-based and learning-based methods in terms of Peak Signal-to-Noise Ratio (PSNR).
• Achieved significant improvements in visual quality and the recovery of salient image components like edges and textures.
• The regularization module offers learned image-adaptability and interpretability inherited from RED.

Conclusions:

• The proposed DRED-DUN model represents a significant advancement in image restoration frameworks.
• End-to-end trainability and the novel regularization strategy lead to superior performance and detail preservation.
• This method offers a more effective approach for recovering fine image structures.