Abstract
Under-display camera (UDC) systems enable full-screen displays by mounting the lens beneath the screen. However, the OLED pixel array causes significant light diffraction and attenuation, leading to complex degradations like flare and haze. We present D$^2$RNet, a novel video restoration network that decouples these degradations. By employing Decoupling Attention Modules (DAM), the network separates flare and haze components to apply targeted temporal processing. Our approach achieves state-of-the-art performance on a new large-scale UDC video benchmark, VidUDC33K.
The UDC Challenge: Flare and Haze
The primary hurdle for UDC systems is the OLED display itself. The pixel grid acts as a diffraction grating, scattering light before it reaches the sensor. This results in two distinct types of degradation:
- Flare: Intense light sources produce saturated streaks and "rainbow" artifacts.
- Haze: Darker areas suffer from a loss of contrast and blurry textures.
In video, these effects are dynamic. As the camera or objects move, the Point Spread Function (PSF) shifts, making temporal consistency difficult to maintain with standard image restoration techniques. To address this, the authors modeled the dynamic PSF to create realistic training data.
Methodology: D$^2$RNet and the DAM
The core of D$^2$RNet is the Decoupling Attention Module (DAM). Unlike previous methods that treat all degradations uniformly, DAM uses a soft mask generation function to partition frames into flare and haze regions based on light intensity.
Targeted Temporal Processing
- Long-Term Flare Removal: Because flare often saturates pixels (destroying local content), the network uses long-term features to retrieve information from distant frames where the light source might not have been present or was in a different position.
- Short-Term Haze Removal: Haze is primarily a texture blur issue. The network leverages short-term features from adjacent frames to sharpen details through precise temporal alignment.
The overall architecture is a multi-scale bi-directional recurrent network, allowing it to handle objects changing size and distance while maintaining high efficiency.
VidUDC33K: A New Benchmark
To train and evaluate D$^2$RNet, the authors introduced VidUDC33K, the first large-scale UDC video dataset. It consists of 677 high-quality HDR videos degraded using a realistic PSF measured from a commercial ZTE Axon 20 smartphone. This ensures the model learns to handle the specific optical characteristics of real-world UDC hardware.
Results and Performance
D$^2$RNet sets a new standard for UDC video restoration:
- Quantitative Superiority: It achieves 31.91 dB PSNR on VidUDC33K, a 1.02 dB improvement over the previous state-of-the-art (RVRT).
- Efficiency: With only 5.76M parameters and a runtime of 0.44s per frame, it is significantly more efficient than heavy Transformer-based models like VRT, making it viable for mobile deployment.
- Qualitative Quality: Visual results demonstrate superior removal of diffraction artifacts and better restoration of shadow textures compared to general-purpose video restoration models.
Conclusion
By "decoupling" the complex optical degradations of UDC systems into manageable components, D$^2$RNet provides a targeted and efficient solution for full-screen camera restoration. This divide-and-conquer strategy not only improves visual quality but also ensures the temporal consistency required for high-quality video.
Summary updated