maanyabhatia/ship-hull-biofouling-segmentation-dinov2

DINOv2 semi-supervised segmentation for ship hull biofouling detection. CAIR Lab, IIT Mandi. Paper submitted to IEEE OCEANS 2026. # Ship Hull Biofouling Segmentation Using DINOv2 Semi-supervised semantic segmentation for automated ship hull biofouling detection, built from just 24 manually annotated images. ## Overview This project, completed during a research internship at the Centre for Artificial Intelligence Research (CAIR), IIT Mandi (supervisor: Prof. Jagadeesh Kadiyam), develops a DINOv2-based segmentation pipeline for detecting biofouling (barnacles, mussels, biofilm) on ship hulls. Starting from only 24 manually annotated images, an iterative pseudo-label expansion pipeline grows the training set to 407 images, cutting annotation cost by roughly 94 percent while substantially improving segmentation accuracy over a purely supervised baseline. A paper describing this work has been submitted to IEEE OCEANS 2026. ## Key results | Model | Training images | mIoU | Dice | Pixel accuracy | | --- | --- | --- | --- | --- | | V1 (supervised baseline) | 24 | 0.0897 | 0.1102 | 0.2828 | | Round 1 (pseudo-labeled) | 407 | 0.4383 | 0.4952 | 0.6629 | | Round 2 (degraded) | 407 | 0.0823 | 0.1348 | 0.1457 | | Improved architecture (best) | 24 | 0.4624 | 0.5063 | 0.6857 | | U-Net baseline (supervised) | 24 | 0.3251 | 0.4264 | 0.5707 | | DeepLabV3+ baseline (supervised) | 24 | 0.2939 | 0.3855 | 0.5556 | The improved DINOv2 architecture outperforms a U-Net baseline by roughly 42 percent mIoU and DeepLabV3+ by roughly 57 percent mIoU, under the same 24-image annotation budget. ## Method The pipeline starts by training a DINOv2 ViT-S/14 backbone (self-supervised, pretrained on 142M images) on the 24 hand-annotated images, with the last three transformer blocks unfrozen and a 4-layer convolutional decoder head, using a combined CrossEntropy and Dice loss. This V1 model is then used to generate pseudo-labels for a larger pool of unlabeled ship-hull images, expanding the training set from 24 to 407 images for Round 1. A second pseudo-labeling round was also tested and found to degrade performance, which …