<![CDATA[Accurate segmentation of shipwreck targets in sonar imagery is important for underwater archaeology, marine monitoring, and search operations, but the task remains difficult because labeled sonar masks are scarce and full adaptation of transformer models can be computationally expensive. This study evaluates whether parameter-efficient fine-tuning can provide a practical alternative for binary sonar shipwreck segmentation. Using SegFormer-B0 initialized from a pretrained checkpoint, three adaptation strategies were compared under a consistent protocol: full fine-tuning of all model parameters (FullFT), training only the segmentation head (Head-only), and LoRA-based adaptation of selected linear layers together with head training (LoRA-A+Head). Models were selected by the best validation epoch and evaluated on a held-out test set. Across three random seeds, FullFT achieved the best performance, with a Dice score of 0.614 ± 0.008 and IoU of 0.487 ± 0.007. LoRA-A+Head achieved a Dice score of 0.546 ± 0.010 and IoU of 0.401 ± 0.008 while updating only 1.57% of the parameters, whereas Head-only reached 0.494 ± 0.010 Dice and 0.354 ± 0.008 IoU. These results show a clear accuracy efficiency trade off, full fine-tuning gives the highest accuracy, whereas LoRA-A+Head offers a practical option when reducing the number of updated parameters is important. The findings support the use of parameter-efficient adaptation for sonar segmentation in compute-limited settings.]]>
Copyrights © 2026
