CoRRN: Cooperative Reflection Removal Network.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Removing reflections from images captured through glass is crucial for many computer vision applications.
• Existing non-learning methods struggle with real-world reflection properties due to limited descriptive capabilities.

Purpose of the Study:

• To develop a unified framework for reflection removal that integrates image context and multi-scale gradient information.
• To address the challenge of strong local reflections using a novel statistic loss function.

Main Methods:

• A novel network architecture employing a feature-sharing strategy for cooperative reflection removal.
• Integration of image context and multi-scale gradient information within a unified framework.
• Introduction of a statistic loss function that considers gradient level statistics between background and reflections.

Main Results:

• The proposed network effectively removes reflections by leveraging integrated contextual and gradient features.
• The statistic loss function aids in removing strong local reflections.
• Experiments demonstrate superior performance compared to state-of-the-art methods on a public benchmark dataset.

Conclusions:

• The proposed feature-sharing network offers an effective solution for reflection removal in computer vision.
• The integration of diverse image information and a novel loss function advances the field of reflection removal.
• The method shows significant potential for real-world applications requiring clear imagery through glass.