Attention-Based Complementary Learning for Active Target Classification With Limited Sonar Data

We propose an active sonar target classifier using attention-based complementary learning (ABCL) to mitigate poor generalization with scarce active sonar data. ABCL utilizes the information of different features (spectrograms of short-time Fourier transform and constant-Q transform in the present study), which provide distinct perspectives for the same raw data. The complementary learning module of ABCL first summarizes the input features using a transformer encoder layer, including self-attention. Subsequently, the attention mechanism is applied to the refined input features from the transformer encoder layer to generate a comprehensive feature with better generalization using limited active sonar data. The comprehensive feature from two different features enables a shallow network to discriminate between active targets and nontargets and allows ABCL to be light. Two active sonar datasets from two oceanic environments were used for training and testing. The generalization of ABCL was examined using a receiver operating characteristic curve. ABCL showed a comparable performance to state-of-the-art deep learning (DL) models in vision in a test scenario and achieved the best performance with low variance in the other test scenario, where other DL models showed a poor probability of detection at a low probability of false alarms.