Problem
Research questions and friction points this paper is trying to address.
nighttime image dehazing
domain drift
data scarcity
distribution mismatch
low illumination
CLIP-Guided Data Augmentation for Night-Time Image Dehazing
🤖 AI Summary
Nighttime image dehazing is challenged by complex degradations including haze scattering, low illumination, non-uniform lighting, and strong light interference, which often lead to domain shift and training instability under limited supervision. This work proposes a practical dehazing framework that eliminates the need for reconstruction networks by leveraging a CLIP vision encoder to select external data for effective domain alignment. The approach employs NAFNet in a two-stage training scheme and enhances robustness during inference through an ensemble of test-time augmentation (TLC), x8 self-ensemble, and weighted snapshot fusion. The proposed method effectively mitigates domain shift and training instability, demonstrating superior performance and output consistency in the NTIRE 2026 Nighttime Dehazing Challenge.
📝 Abstract
Nighttime image dehazing faces a more complex degradation pattern than its daytime counterpart, as haze scattering couples with low illumination, non-uniform lighting, and strong light interference. Under limited supervision, this complexity aggravates domain drift and training instability, since target-domain samples are scarce while naively introducing external data may weaken adaptation due to distribution mismatch. This paper presents our solution to the NTIRE 2026 Night Time Image Dehazing Challenge, built as a unified framework that integrates domain-aligned data construction, stage-wise training, and inference-time enhancement. Specifically, a pre-trained CLIP visual encoder screens candidate external samples by similarity to construct training data closer to the target domain. NAFNet is then trained in two stages, first adapting to the target domain and then expanding to broader degradation patterns. At inference time, TLC, x8 self-ensemble, and weighted snapshot fusion are combined to improve output stability. Rather than relying on complex network redesign, the proposed framework offers a practical and effective pipeline for nighttime image dehazing.
Innovation
Methods, ideas, or system contributions that make the work stand out.
CLIP-guided data augmentation
domain-aligned data construction
stage-wise training
nighttime image dehazing
inference-time enhancement
🔎 Similar Papers
No similar papers found.
