Abstract
This study describes an innovative methodology designed to filter out redundant visual information during the construction of underwater photo-mosaics of the seabed. Collecting visual data of the seabed using underwater robots typically requires taking hundreds or even thousands of images when missions are performed in large areas and/or have long durations. All these images typically contain multiple overlapping parts that need to be filtered out appropriately in order to eliminate useless repetitive information, questionable pixel aggregations, unrealistic color blends, and the collapse of computational resources dedicated to the mosaic-building process. The global process presented here has two phases: a first phase to extract and preprocess images, which involves decoding from Bayer to RGB format, resolution reduction, rectification, and contrast enhancement. The second phase includes image filtering, discarding those recorded outside the effective mission time and those with hashes that are sufficiently similar to be considered overlapping. Experiments have been conducted with datasets collected from an Autonomous Underwater Vehicle while observing marine habitats of special ecological interest. Experiments on real-world benthic datasets have demonstrated an Image Reduction Ratio (IRR) of up to 49.91%, leading to a 30% reduction in total processing time. Crucially, this efficiency gain is achieved without significant loss of environmental data, as evidenced by a Feature Persistence Ratio (FPR) consistently above 0.97. This methodology provides a scalable solution for large-area seabed mapping, drastically reducing data management efforts while maintaining the high spatial resolution and informativeness required for ecological benthic assessment.