Core Concepts
Proposing a semi-supervised model for real-world nighttime dehazing with spatial-frequency awareness and realistic brightness constraint.
Abstract
Existing research focuses on daytime image dehazing, neglecting nighttime hazy scenes.
Proposed model addresses issues of multiple light sources, glow, noise, and unrealistic brightness in nighttime dehazing.
Spatial and frequency domain interaction module handles localized, coupled, and frequency inconsistent characteristics.
Retraining strategy with pseudo labels and local window-based brightness loss for realistic brightness.
Outperforms state-of-the-art methods on public benchmarks.
Stats
"Experiments on public benchmarks validate the effectiveness of the proposed method and its superiority over state-of-the-art methods."
"The brightness intensity corresponding to x0i ∈ RX and y0i ∈ RY are µ(x0i) and µ(y0i), respectively."
Quotes
"We propose a spatial and frequency domain aware semi-supervised nighttime dehazing network (SFSNiD)."
"The experimental results on synthetic and real-world datasets show that the proposed method can achieve impressive performance."