Trustworthy Automatic Target Recognition via Variational Autoencoder-based Out-Of-Distribution Detection

Forward-Looking Sonar (FLS) images offer unique perspectives for underwater inspection, but their low resolution and speckles make Deep Learning detectors prone to high confidence errors whenever unfamiliar objects appear. To tackle this issue, we present a lightweight trustworthiness add-on for YOLObased Automatic Target Recognition (ATR). Each detected crop is analyzed by a class-specific Variational Autoencoder, and its Mean Squared reconstruction Error (MSE)-based loss is evaluated only on echo-bearing pixels. Consecutively, fitting a Gaussian Kernel-Density Estimate (KDE) to reference losses allows to produce a percentile-normalised Similarity Score that quantifies how typical a crop is for its predicted class. Experiments on the public Marine-Debris dataset show a clear separation between In-Distribution (ID) and Out-of-Distribution (OOD) loss densities thanks to our add-on, mirrored by the score reported in the accompanying tables. The method is training-data efficient (requires no OOD examples), computationally inexpensive, and potentially deployable on embedded hardware, making it a practical step toward reducing False Positive detections in FLS imagery, thus increasing trustworthiness in maritime ATR systems.