Answers are grounded in this paper and cite their sources.
The Adaptive Multi-Frequency Sharpening Model (AMFSM) is a lightweight, training-free approach designed to enhance underwater images by mitigating the effects of light absorption and scattering. Based on multi-frequency Gaussian decomposition with an adaptive weighting factor (α) that incorporates a Laplacian-based noise-related indicator, the proposed approach is capable of adapting the degree of sharpening based on the local properties of the image. Consequently, the model selectively enhances edges and fine details while suppressing noise and artefacts in homogeneous regions. Following the sharpening process, a colour restoration phase is implemented. This involves red channel compensation and luminance-based CLAHE to eliminate dominant blue–green colour casts while simultaneously improving overall contrast. On a 100-image subset of the LSUI benchmark, on which its sharpening gain and fusion weights were also tuned, AMFSM obtained a PSNR of 23.45 dB, an SSIM of 0.89, and a UCIQE of 0.68, the highest mean values among the evaluated traditional, fusion-based and deep learning baselines on all eight reported metrics. Paired Wilcoxon signed-rank and t-tests against the strongest baseline (INGC-GAN), with Holm correction, indicate statistically significant differences for seven of the eight metrics, whereas the UCIQE difference is not statistically significant. AMFSM requires 0.12 s per image on a laptop CPU.
No takes yet. Share an insight, caveat, or question.
Dhoom et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: