Half-Quadratic Splitting Method in Image Precompensation: Accelerating Optimization and Parameter Selection

Software-based image precompensation enables users with visual impairments (e.g., myopia, hypermetropia, or astigmatism) to view images on digital displays clearly without corrective lenses or hardware modifications. This is achieved through image processing using information about user's refraction anomalies. The practical viability of software-based methods hinges on balancing retinal image quality with computational efficiency. Currently, a persistent trade-off exists: optimization-based approaches, such as Montalto et al.'s method, yield high-quality restorations but suffer from low computational performance, limiting their real-time applicability.

To overcome this bottleneck, we propose an accelerated optimization of the functional underlying the Montalto method by applying a half-quadratic splitting technique. Furthermore, we introduce a novel training scheme for parametric precompensation. By tuning the optimization parameters across different numbers of iterations, we generated a family of algorithmic variants, allowing for dynamic, tunable trade-offs between restoration quality and processing speed.

Evaluations on the SCA-2023 dataset demonstrate that our proposed algorithm provides a massive performance gain with negligible quality loss. Across multiple metrics, restoration quality differs from the baseline Montalto method by no more than 1-2%. Crucially, the proposed modifications execute hundreds of times faster on a GPU and tens of times faster on a CPU. Specifically, the five-iteration variant achieves 35 frames per second for 512p images on a standard CPU.

The application of half-quadratic splitting successfully resolves the quality–speed bottleneck in image precompensation. By dramatically accelerating optimization without compromising image quality, our algorithm achieves the frame rates required for CPU-based real-time systems.