Adaptive haze pixel intensity perception transformer structure for image dehazing networks

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning networks for image dehazing often struggle with real-world haze, algorithmic inefficiencies, and color distortion.
• Existing methods face challenges in daytime environments and cross-dataset generalization.

Purpose of the Study:

• To propose SwinTieredHazymers (STH), an adaptive deep learning network for effective image dehazing.
• To address limitations in current dehazing algorithms concerning efficiency and color fidelity.

Main Methods:

• Developed STH, a novel dehazing network utilizing a three-branch hierarchical design.
• Integrated Transformer for global features and Convolutional Neural Networks (CNN) for local features to modulate haze residuals.
• Employed adaptive pixel intensity discernment for clarity restoration.

Main Results:

• STH demonstrated superior performance over advanced single-image dehazing methods.
• Achieved high quantitative metrics and visual fidelity in real-world hazy image dehazing.
• Showcased strong performance in cross-dataset dehazing scenarios, indicating robustness.

Conclusions:

• SwinTieredHazymers (STH) effectively restores clarity in hazy images by adaptively processing haze residuals.
• The hierarchical integration of global and local features enhances the algorithm's applicability and performance.
• STH represents a significant advancement in single-image dehazing for diverse real-world conditions.