Multi-Stream Fusion Network With Generalized Smooth L₁ Loss for Single Image Dehazing

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Single image dehazing is a critical challenge in computer vision.
• Existing methods often struggle with complex haze patterns and real-world conditions.

Purpose of the Study:

• To introduce an end-to-end convolutional neural network, the Multi-Stream Fusion Network (MSFNet), for single image dehazing.
• To enhance feature extraction and fusion for improved dehazing performance.

Main Methods:

• Developed an encoder-decoder architecture with a three-stream encoder using Residual Dense Blocks (RDBs).
• Implemented a feature fusion block with stream-wise and channel-wise attention mechanisms.
• Utilized a generalized smooth L1 loss function and a multi-scale supervision learning strategy for training.

Main Results:

• MSFNet demonstrated superior performance in dehazing synthetic and real-world images.
• The proposed attention mechanisms and multi-scale supervision effectively guided feature capture.
• The generalized smooth L1 loss provided robust training by managing outliers.

Conclusions:

• MSFNet offers a state-of-the-art solution for single image dehazing.
• The network's architecture and training strategies contribute to its high performance.
• This work advances the field of image restoration and computer vision applications.