Autonomous underwater garbage grasping and collection pose a great challenge to underwater robots. To assist underwater robots in locating and recognizing underwater garbage objects efficiently, a modified U-Net-based architecture consisting of a deeper contracting path and an expansive path is proposed to accomplish end-to-end image semantic segmentation. In addition, a dataset for underwater garbage semantic segmentation is established. The proposed architecture is further verified in the underwater garbage dataset and the effects of different hyperparameters, loss functions, and optimizers on the performance of refining the predicted segmented mask are examined. It is confirmed that the focal loss function will lead to a boost in solving the target-background unbalance problem. Eventually, the obtained results offer a solid foundation for fast and precise underwater target recognition and operations.
Keywords: U-Net; deep learning; semantic segmentation; underwater garbage collection.