Semantic-Aware Dehazing Network With Adaptive Feature Fusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Single image dehazing is challenging due to semantic confusion in hazy scenes.
• Existing convolutional neural networks (CNNs) struggle with natural and realistic reconstructions.
• Semantic information is crucial for accurate object representation in haze-free images.

Purpose of the Study:

• To develop a novel dehazing network that incorporates semantic priors for improved image quality.
• To address the limitations of current methods in recovering faithful textures and natural appearances.
• To enhance the semantic understanding within the dehazing process.

Main Methods:

• Proposed a semantic-aware dehazing network (SDNet) utilizing semantic priors as a color constraint.
• Designed a densely connected block for capturing global and local image information.
• Implemented adaptive feature fusion from shallow and deep layers to refine object appearance.

Main Results:

• The proposed SDNet effectively recovers textures and improves the naturalness of dehazed images.
• Experimental results show favorable performance compared to state-of-the-art single image dehazing methods.
• Semantic prior incorporation leads to more reasonable scene configurations and realistic reconstructions.

Conclusions:

• Incorporating semantic priors into dehazing networks is a viable approach for achieving high-quality, realistic results.
• The developed SDNet demonstrates the effectiveness of semantic-aware processing in single image dehazing.
• Future work can explore more sophisticated semantic prior integration and feature fusion techniques.