Uncertainty Driven Active Learning for Image Segmentation in Underwater Inspection

TL;DR

This work tackles the labeling bottleneck in semantic segmentation for underwater infrastructure inspection by applying epistemic-uncertainty–driven active learning via MC-dropout and mutual information. It evaluates two segmentation models, DenseNet-56 and HyperSeg-S, on CamVid and a large underwater pipeline image set, showing that uncertainty-based sampling significantly reduces labeled data requirements while approaching or matching full-dataset performance. A key contribution is a threshold-based active learning framework with selective labeling and a reproducible implementation, achieving a substantial mean IoU gain on underwater data (e.g., $67.5\%$ meanIoU with only $12.5\%$ of data) and demonstrating robustness through repeatability analyses. The findings imply practical benefits for underwater inspection tasks by lowering annotation costs and improving handling of class-imbalanced, low-quality imagery in real-world deployments.

Abstract

Active learning aims to select the minimum amount of data to train a model that performs similarly to a model trained with the entire dataset. We study the potential of active learning for image segmentation in underwater infrastructure inspection tasks, where large amounts of data are typically collected. The pipeline inspection images are usually semantically repetitive but with great variations in quality. We use mutual information as the acquisition function, calculated using Monte Carlo dropout. To assess the effectiveness of the framework, DenseNet and HyperSeg are trained with the CamVid dataset using active learning. In addition, HyperSeg is trained with a pipeline inspection dataset of over 50,000 images. For the pipeline dataset, HyperSeg with active learning achieved 67.5% meanIoU using 12.5% of the data, and 61.4% with the same amount of randomly selected images. This shows that using active learning for segmentation models in underwater inspection tasks can lower the cost significantly.

Figures (8)

• Figure 1: Our active learning process: After pre-training the model with a small set of randomly chosen images, we start selecting images with uncertainty above a threshold $\textrm{TR}$ (top-left) for training and validation of Model A. Due to space constrains only one $TR$ is represented. However, in each iteration two new thresholds are defined: $TR_t$ for training, and $TR_v$ for validation images. We discard images with an uncertainty below the mean minus 1.5 standard deviations. For evaluation, the process is repeated on randomly selected images (bottom-left) training model B. We compare the meanIoU performance of the two models after each iteration (bottom-right). In this image, $D_U$ and $D_L$ are the pools of unlabeled and labeled images, respectively, both for uncertainty-based selection ($uncer$) and random selection ($rand$). The signs '$+$' and '$-$' indicate that the newly selected images are removed from $D_U$ and added to $D_L$. The arrows with dashed lines highlight that, for each iteration, the number of images to be selected randomly is defined by the number of images selected based on uncertainty.
• Figure 2: Example of the pipeline dataset: (a-b) an anode and pipeline, (c) a pipeline and a field joint, (d) a field joint, boulders and pipeline, (e) a pipeline alone, and (f) a pipeline, field joint, and part of a vehicle in the top right (best viewed in color).
• Figure 3: For each number of forward passes $T$, $\overline{EU}_{\textrm{img}}$ of the validation dataset of CamVid was calculated five times using DenseNet. The graph on the left shows the average of the five results obtained, and the graph on the right shows the standard deviation. Here $\overline{EU}_{\textrm{img}}$ was calculated using mutual information (MI).
• Figure 4: Results of the active learning experiments. The bottom graphs show the difference in mean IoU between the model trained with active learning vs. trained with random images, where positive values means active learning prevails.
• Figure 5: Test segmentation results from the CamVid model. GT is the ground truth. MCD is the average of the results obtained with MC-dropout. For the entropy and the mutual information (MI) plots, the warmer colors represent higher values (best viewed in color)