Uncertainty Guided Multi-Scale Attention Network for Raindrop Removal From a Single Image.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Raindrop removal from images is challenging due to raindrop density and diversity.
• Existing methods using binary masks overlook the varying blur levels and patterns of raindrops.

Purpose of the Study:

• To develop an effective raindrop removal method that accounts for raindrop diversity and blur levels.
• To introduce a novel framework, the uncertainty guided multi-scale attention network (UMAN), for enhanced raindrop removal.

Main Methods:

• Proposed a soft mask approach with values in [-1,1] to represent raindrop blur levels.
• Explored multi-scale image fusion using deep features to capture complementary raindrop patterns.
• Developed an iterative mechanism within a multi-scale pyramid structure and incorporated an attention mechanism for effective fusion.

Main Results:

• The proposed soft mask effectively utilizes blur degree information for raindrop removal.
• Multi-scale fusion and attention mechanisms successfully highlighted raindrop features and reduced noise.
• UMAN achieved convincing results on benchmark datasets, outperforming existing methods.

Conclusions:

• The soft mask and multi-scale fusion strategy significantly improve raindrop removal performance.
• UMAN provides a robust framework for handling diverse raindrop characteristics in images.
• The method demonstrates the effectiveness of incorporating blur-level and multi-scale information for image restoration tasks.