CRISP enhances low-light videos using progressive training and semantics
CRISP: Progressive Illumination Curriculum for Recurrent Low-Light Video Enhancement with Semantic Priors
Computer Vision and Pattern Recognition
Summary
Taking videos in dark places can make them look blurry, noisy, or unclear. The authors introduce a new method called CRISP that improves these dark videos by gradually teaching a computer to brighten and fix colors over time, while also using information about what's in the scene to make details clearer. This approach helps keep video frames looking smooth and connected, even when there's movement or noise. Tests show this method produces clearer and more natural-looking videos compared to others.
What this means in practice
- •For video editing software developers: Incorporate CRISP’s technique to improve visibility and stability of dark videos without slowing down processing times.$Commercial implications: CRISP enables more efficient and higher-quality low-light video enhancement features for consumer and professional video editing tools.
- •For security camera system engineers: Use CRISP’s method to better recover details and maintain frame consistency in nighttime or low-light surveillance footage.
Authors
Ruirui Lin, Xin Yan Lim, Yini Li, Zhiming Liu, Nantheera Anantrasirichai
Abstract
Low-light video enhancement requires the recovery of brightness, color, and fine spatial structures while preserving temporal coherence in the presence of noise and motion. Existing methods often struggle to balance these objectives, particularly in severely degraded regions. This paper presents CRISP, a recurrent framework trained using a progressive illumination curriculum. Its two-frame recurrent architecture propagates information across successive frames to exploit temporal context beyond the immediate input pair. During training, supervision progressively transitions from intermediate targets interpolated between the low-light input and the reference to the full reference, gradually increasing the required brightness and color correction. This curriculum stabilizes recurrent training without introducing additional inference cost and yields the largest performance improvement among the components evaluated in the ablation study. To further support structural reconstruction, the framework optionally incorporates frozen DINOv3 semantic features through adaptive layer normalization (AdaLN)-style modulation, without fine-tuning the foundation model. Experiments on DID and BVI-RLV demonstrate that CRISP achieves the highest PSNR on DID and the highest SSIM on BVI-RLV among the evaluated methods, while maintaining computational efficiency.