MCRFS-Net: single image dehazing based on multi-scale contrastive regularization and frequency selection.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image dehazing aims to restore clarity in hazy images, but non-uniform haze presents a significant challenge for existing atmospheric scattering models.
• Current methods often address image-level or feature-level non-uniformity separately, with few effectively tackling both simultaneously.

Purpose of the Study:

• To develop a novel approach for image dehazing that effectively addresses non-uniform haze at both image and feature levels.
• To introduce an Adaptive Multi-Scale Frequency Selection (AMFS) module and a Multi-Scale Contrast Regularization (MSCR) loss function for improved dehazing performance.

Main Methods:

• Introduced the Adaptive Multi-Scale Frequency Selection (AMFS) module, comprising an Adaptive Multi-Scale Module (AMSM) for weighted feature fusion and a Frequency Selection Block (FSB) for frequency domain processing.
• The AMSM integrates multi-scale features to mitigate non-uniform dehazing issues, while the FSB uses an attention mechanism to highlight important frequency components and suppress noise.
• Proposed a Multi-Scale Contrast Regularization (MSCR) loss function utilizing cross-scale contrastive learning to enhance feature consistency.

Main Results:

• The proposed algorithm demonstrated superior performance compared to existing methods on four benchmark datasets.
• Achieved enhanced detail preservation and improved robustness in dehazing under non-uniform haze conditions.
• The AMFS module effectively handled non-uniform haze by integrating multi-scale features and processing them in the frequency domain.

Conclusions:

• The novel AMFS module and MSCR loss function provide an effective solution for non-uniform image dehazing.
• The proposed method offers significant improvements in detail restoration and robustness, outperforming current state-of-the-art techniques.
• This work advances the field of image dehazing by addressing simultaneous image and feature level non-uniformity.