Deep Neural Networks for Reliable Underwater Vision

Reliable computer vision in underwater environments is essential for camera-based systems utilized for environmental monitoring, inspection, and navigation. However, underwater image data is characterized by unique properties such as turbidity, low contrast, and limited availability of annotated datasets. Deep neural networks
are data-driven methods, and thus particularly sensitive to these problems. This thesis investigates how deep neural networks (DNNs) can be evaluated and tested in underwater settings through the lenses of environmental monitoring and navigation.
The work is carried out in four studies under increasing operational complexity, ranging from static long-term cameras setups to towed camera devices and cameras mounted onto autonomous underwater vehicles. First, visual anomaly detection methods are evaluated in static camera setups for biodiversity monitoring where predicted anomaly scores from DNNs are compared with a multi-annotated marine life dataset. The evaluation shows that reverse distillation aligns most closely with human perception under varying levels of visibility. Second, DNN-based image classification is applied to seagrass presence detection in video footage from a towed camera system. Vision transformers perform with high accuracy across different deployment sites, and their predictions enable automated estimation of seagrass coverage for environmental monitoring. Third, neural radiance fields are examined as alternatives to simulators by generating realistic and diverse test data for visionbased navigation. By rendering new viewpoints from shifting camera trajectories, the generated test data reveals limitations in the robustness of feature extraction methods. Fourth, preliminary work shows how 3D Gaussian splatting can be used to generate targeted test videos for evaluating loop closure detection mechanisms, a critical component of underwater navigation pipelines. Initial results showed a sharp decline in performance with increasing loop length of generated videos. By integrating novel computer vision techniques in visual anomaly detection, image classification and 3D scene reconstruction, this thesis advances the application of state-of-the-art methods in underwater image analysis and provides insights into the reliability and limitations of DNNs in complex underwater environments.