Abstract
Poisson image deconvolution is a critical task in low-light and medical imaging, typically solved using Kullback-Leibler (KL) divergence. However, KL-based optimization is often slow and unstable due to a lack of Lipschitz continuity. This paper proposes learning a Weighted Least Squares (WLS) data term using a lightweight neural network integrated into an unrolled optimization algorithm. This hybrid approach offers a 0.8 dB PSNR improvement over standard methods while ensuring numerical stability and model interpretability through a parameter-efficient architecture.
The Challenge of Poisson Deconvolution
In image processing, deconvolution aims to recover a sharp image from one corrupted by blur and noise. For Poisson noise, common in low-light scenarios, the Kullback-Leibler (KL) divergence is the statistically optimal data fidelity term. However, KL divergence presents significant optimization challenges, leading to slow and unstable convergence.
A common alternative is Weighted Least Squares (WLS). While the theoretical weights for Poisson noise are $w = 1/y$ (where $y$ is the observed data), small pixel values cause these weights to explode, forcing optimization algorithms to take infinitesimally small steps. This research seeks to bridge the gap between statistical optimality and numerical efficiency.
Methodology: Hybrid Unrolled Optimization
The authors introduce a hybrid framework that combines classical optimization with deep learning:
- Weight Estimation Module ($g_\theta$): A tiny Convolutional Neural Network (3,211 parameters) predicts an optimal weight map $w$ from the noisy image. A sigmoid activation ensures weights remain in a stable range $[0, 2]$.
- Unrolled Optimization: The Forward-Backward Splitting (FBS) algorithm is unrolled for 10 iterations, treating optimization steps as network layers.
- End-to-End Learning: The system simultaneously learns weights, regularization strength ($\lambda$), and step size ($\gamma$).
Experimental Results
The model was evaluated on the TAMPERE17 dataset, consisting of high-quality images corrupted with Gaussian blur and Poisson noise. The proposed learned WLS method achieved an average PSNR of 27.71 dB, outperforming standard Least Squares by 0.8 dB.
Qualitative analysis confirms that the network naturally learns a distribution similar to the theoretical $1/y$ model but "clips" the values to maintain stability. This results in superior edge preservation and noise reduction in dark regions, as seen in the clear reconstruction of text on magazine covers.
Conclusion and Future Work
By using a parameter-efficient model, the authors demonstrate that high-quality deconvolution does not require massive "black-box" architectures. The model remains interpretable and robust even under strong blur. Future work will explore more sophisticated learned priors and the application of this framework to mixed noise profiles or non-diagonal weighting operations.
Summary updated