Online knowledge distillation network for single image dehazing.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Traditional dehazing methods face challenges like artifacts and color distortion.
• Learning-based methods offer better color fidelity but may yield under-dehazed results.

Purpose of the Study:

• To propose a novel online knowledge distillation network (OKDNet) for single image dehazing.
• To combine the strengths of model-based and model-free approaches for improved dehazing performance.

Main Methods:

• Utilized a multiscale network with attention-guided residual dense blocks for feature extraction.
• Employed dual branches for dehazing: one model-based (atmospheric scattering) and one model-free.
• Introduced an efficient feature aggregation block and an online knowledge distillation strategy for joint optimization.

Main Results:

• OKDNet achieved superior performance on both synthetic and real-world images compared to state-of-the-art methods.
• The proposed method demonstrated effectiveness in removing haze while maintaining image quality and color fidelity.
• Achieved superior performance with fewer model parameters.

Conclusions:

• OKDNet successfully integrates model-based and model-free techniques for effective single image dehazing.
• The proposed online knowledge distillation strategy enhances dehazing performance.
• OKDNet offers a promising solution for computer vision tasks requiring high-quality, haze-free images.