SEMANTIC SEGMENTATION OF SHIP HULLS FOR UNDERWATER CLEANING ROBOTS

Underwater barnacle cleaning robots require a semantic segmentation algorithm deployable onboard to effectively segment barnacles, obstacles, and background on ship hulls. This study independently constructed a dedicated ship hull semantic segmentation dataset and implemented an improved BiSeNet network. The approach integrates features from both a spatial path and a context path. Building upon the cross-entropy loss function, class weights were strategically assigned during training according to the distribution of different target categories within the dataset. This method achieved significantly high Intersection over Union (IOU) and F1 scores.