Discovering unknowns: Context-enhanced anomaly detection for curiosity-driven autonomous underwater exploration

Discovering unknown objects from visual information as curiosity is highly demanded for autonomous exploration in underwater environment. In this research, we propose an end-to-end deep neural network for anomaly detection in the highly dynamic unstructured underwater background faced by a moving robot. A novel patch-level autoencoder combined with a context-enhanced autoregressive network is introduced to differentiate abnormal patterns (unknowns) from normal ones (knowns) in fine-scale regions. The autoencoder and autoregressive network share the same encoder to extract latent features. The autoregressive branch learns semantic dependence based on conditional probability to identify anomaly in a latent feature space. The overall anomaly score is weighted by both image reconstruction loss and feature similarity loss. The model outperforms state-of-the-art anomaly detection, demonstrated on the benchmark dataset CIFAR-10. Average discrimination performance AUROC improved 2.18%, and inception distance between normal and anomalous classes improved 9.33% in Z-score. The network has been tested using three underwater datasets from underwater simulation, a real-world undersea video and public SUIM data. The AUROC accuracy improved 6.36%, 32.45% and 40.17% respectively by using the proposed patch learning paradigm. It is the first report on unknown detection as navigation clues for curiosity-driven autonomous underwater exploration.