Deep Video Dehazing with Semantic Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Single image dehazing using convolutional neural networks (CNNs) shows promise.
• Video dehazing can benefit from temporal information across frames.

Purpose of the Study:

• To develop a deep learning solution for video dehazing by exploiting temporal correlations.
• To incorporate semantic priors for improved transmission map estimation.

Main Methods:

• A CNN is trained end-to-end to accumulate information across video frames for transmission estimation.
• Global semantic priors are integrated to regularize transmission maps, ensuring smoothness within object regions.
• A synthetic video dataset based on the NYU depth dataset was generated for training.

Main Results:

• The proposed method effectively removes haze from outdoor videos.
• Learned features generalize to various outdoor scenes.
• The algorithm demonstrates superior performance compared to state-of-the-art methods on both synthetic and real-world videos.

Conclusions:

• Deep learning models can effectively perform video dehazing by utilizing temporal information.
• Incorporating semantic priors enhances the quality of estimated transmission maps.
• The developed approach offers a robust solution for real-world video haze removal.