Abstract
Traditional motion deblurring often overlooks the fact that blurring occurs in the irradiance domain rather than the final intensity image. This discrepancy, combined with the lack of automatic detection for saturated pixel blocks, often leads to artifacts like ringing and poor noise suppression. This paper proposes a deblurring framework based on irradiance. By combining an energy-accumulation model of the motion process with Camera Response Function (CRF) estimation and introducing a block-based saturation detection algorithm, the method restores clear images with higher fidelity. Qualitative and quantitative results demonstrate superior performance, including higher Signal-to-Noise Ratios (SNR) and reduced ringing compared to traditional direct deblurring methods.
Introduction: The Physics of Motion Blur
Digital images are the result of a complex pipeline. Light hitting a sensor (irradiance) is integrated over the exposure time. If the camera or object moves, this integration results in motion blur. However, before a user sees the image, the camera applies a non-linear Camera Response Function (CRF) and clips values that exceed the sensor's capacity (saturation).
Most existing deblurring algorithms operate directly on the final pixel intensities. This is mathematically problematic because the convolution model of blur—where the blurred image is the product of a sharp image and a kernel—is only physically accurate in the linear irradiance domain. Applying these models to non-linear intensity values introduces errors that manifest as "ringing" artifacts and amplified noise.
Methodology
The proposed method shifts the deblurring process back to the irradiance domain through three key innovations:
1. CRF Estimation via Energy Accumulation
To move from intensity back to irradiance, one must know the inverse CRF. While traditional methods require multiple exposures or specific geometric constraints, this paper introduces a model that solves for the CRF by analyzing energy accumulation along the motion path. This allows for accurate CRF estimation from a single blurred image, offering greater flexibility for real-world applications.
2. Block-Based Saturation Detection
Saturated regions, such as bright lights in night scenes, contain no useful blur information and violate the assumptions of deconvolution. Rather than using simple intensity thresholds, which are sensitive to noise, the authors propose a block-based detection algorithm. By using a sliding window and dual thresholds, the system identifies contiguous regions of saturation, allowing the deblurring process to mask these areas and prevent them from corrupting the restoration of the rest of the image.
3. Irradiance-Domain Restoration
Once the CRF is estimated and saturated blocks are identified, the intensity image is mapped to the irradiance domain. Deconvolution is performed in this linear space using the Richardson-Lucy algorithm. Finally, the restored irradiance is mapped back to the intensity domain for the final output.
Experimental Results
The framework was tested on a variety of real-world and simulated datasets, including challenging night scenes and indoor environments.
- Qualitative Performance: The irradiance-based approach significantly reduced ringing artifacts around high-contrast edges compared to direct deblurring methods. It also demonstrated superior noise suppression in darker regions of the image.
- Quantitative Performance: In comparative tests using Peak Signal-to-Noise Ratio (PSNR), the proposed method achieved a score of 26.15 dB on a "Building" test case, substantially outperforming direct deblurring (18.20 dB) and other state-of-the-art methods like Kim et al. (23.38 dB).
Conclusion
By accounting for the physical reality of how cameras capture light, this research demonstrates that deblurring in the irradiance domain is fundamentally more robust than intensity-based approaches. The combination of flexible CRF estimation and robust saturation detection provides a practical path forward for high-quality image restoration in consumer photography.
Summary updated