Anomaly Detection via Reverse Distillation For Ocean Environments Preservation

Monitoring underwater environments is essential for ocean preservation, particularly for detecting marine debris and other anomalies. This study extends the application of the Unsupervised Anomaly Detection (UAD) technique using Reverse Distillation (RD) for semantic segmentation in deep-sea trash detection. The proposed method was evaluated against the DeepLab V3+ supervised model under two experimental protocols: one utilizing the full training dataset and another with only 10% of the original training data to assess performance under limited supervision. Our findings reveal that although the supervised deep learning model delivers higher accuracy and more precise object delineation when abundant labeled data is available, the UAD technique shows promising results as an automated alternative by outperforming DeepLab V3+ under limited labeled data availability.