Two-stage framework for correcting underwater distorted image sequences

In the process of imaging across the water-air interface, captured images of underwater targets are often subject to geometric distortion and detail blur due to water surface fluctuations. Although various correction methods have been proposed, their results still exhibit significant residual distortion, primarily because most existing methods rely on low-quality reference templates and lack effective mechanisms for screening anomalous samples. To address these limitations, this paper proposes what we believe to be a novel two-stage framework that separates preliminary correction from refined reconstruction. In the first stage, a sharpness evaluation mechanism utilizing Rényi entropy is introduced to select a high-quality initial reference frame, and Cox-Munk theory is employed to compensate for centroid displacement. This approach effectively resolves the localization bias inherent in traditional methods that use blurry average frames as reference templates. In the second stage, a median selection-based inversely proportional weighting reconstruction (MSIPWR) strategy is proposed. By utilizing principles of robust statistics, this module adaptively identifies and excludes various types of outliers, such as specular reflection artifacts and transient occlusions, while optimizing weight distribution based on the degree of deviation. This achieves precise suppression of complex environmental noise while simultaneously enhancing fine texture details. Experimental results demonstrate that the proposed method outperforms existing methods in terms of edge sharpness and structural integrity, showing exceptional robustness in suppressing specular highlights. This framework provides a stable and effective solution for high-quality imaging in environments with complex water surface fluctuations.