Ocean Remote Sensing Image Quality Assessment via Multidirectional Perception Fusion and Deviation-Aware Processing

Despite the visibility improvements achieved by existing methods, ocean remote sensing images often suffer from over-enhancement and texture distortion due to light attenuation and scattering, particularly when analyzing maritime objects like coral reefs, marine life, or submerged structures. Traditional assessment methods struggle to handle both over- and under-enhancement. To address this issue, we propose a novel ocean remote sensing quality assessment method that accurately captures diverse image quality deviations and aligns with human visual perception. To deeply analyze the relationship between the overall structure and details in ocean remote sensing enhancement, we introduce multidirectional perception fusion to enhance the perception of image details. To address the over-enhancement or under-enhancement regions, a diff capture block is designed to accurately detect and handle these deviations. In addition, a parallel processing architecture with a symmetric transform block performs a multidimensional analysis of score and weight features, balancing deviation regions against true reference regions. This process yields a comprehensive quality score derived from weighted calculations of activation scores and weights for each image block. Extensive experiments on ocean datasets featuring ocean objects and environments demonstrate that our method outperforms current quality assessment methods. In addition, we have conducted cross-dataset testing on remote sensing datasets that include complex terrains and natural landscapes, providing a more reliable assessment for geoscience applications such as ocean habitat mapping, underwater archaeology, and oceanographic research.