Deep-Masking Generative Network: A Unified Framework for Background Restoration From Superimposed Images

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image restoration tasks like reflection removal, deraining, and dehazing commonly involve removing superimposed noisy layers.
• These tasks are often addressed individually due to the complex and varied nature of noise patterns.

Purpose of the Study:

• To present a unified framework, the Deep-Masking Generative Network (DMGN), for background restoration from superimposed images.
• To develop a method capable of handling diverse noise types within a single model.

Main Methods:

• The DMGN employs a coarse-to-fine generative process, generating a coarse background and noise image in parallel.
• A novel Residual Deep-Masking Cell is utilized as the core unit, learning a gating mask to control information flow for enhancing useful data and suppressing noise.
• The generated noise image is used as contrasting cues to refine the background image.

Main Results:

• The DMGN framework successfully generates high-quality background and noise images progressively.
• Extensive experiments on reflection removal, rain steak removal, and dehazing demonstrate superior performance compared to state-of-the-art methods.
• The unified approach shows consistent outperformance across different image background restoration tasks.

Conclusions:

• The Deep-Masking Generative Network (DMGN) offers a unified and effective solution for various image background restoration problems.
• The proposed Residual Deep-Masking Cell and coarse-to-fine strategy are key to the model's success in handling diverse noise layers.
• DMGN advances the field by providing a single framework that surpasses specialized methods for individual restoration tasks.