IMPROVING UNDERWATER SEMANTIC SEGMENTATION VIA ADAPTIVE FREQUENCY-AWARE ENHANCEMENT AND THE SEGMENT ANYTHING MODEL

Abstract

The segmentation of underwater images is a challenging task in computer vision and marine research due to factors such as low contrast, non-uniform illumination, and color distortion. Despite advances in deep learning, segmentation performance often remains suboptimal. Effective preprocessing of underwater images is therefore essential for improved accuracy. This study proposes a hybrid deep learning framework that combines image enhancement with robust segmentation. Low- and high-frequency components of the image are filtered using Bilateral and Difference-of-Gaussian filters, while an adaptive variational strategy prevents oversaturation and over-enhancement of color and contrast. The enhancement process reduces noise, optimizes contrast, and recovers critical features, facilitating more accurate segmentation. The deep learning component leverages a robust architecture and extensive datasets to further improve segmentation performance. Evaluation using quantitative metrics and visual results demonstrates that the proposed approach significantly enhances underwater image segmentation, benefiting applications in marine exploration, monitoring, and research.