PingCNN: A Low-Latent and Low-Cost Network for Mine Detection in Side Scan Sonar Signal

Underwater mine detection using side scan sonar is of paramount importance to the safety of ships and submarines. Traditional 2-D mine detection algorithms are computationally intensive and induce substantial latency. Real-time operation of mine detection algorithms is essential, especially for underwater vehicles with limited computational capacity. To address this, we propose ping based convolutional neural network (PingCNN), a novel approach leveraging raw 1-D sonar pings, yet not explored in literature. We hypothesize that the characteristics of reflected pings are distinctive to the objects being scanned, thus employing a time series deep learning approach to advance underwater mine detection. PingCNN articulates mine detection as a detection-by-classification problem in a 1-D signal domain. In addition, we propose a novel likelihood-based sensitivity enhancer module, built with statistical and signal processing techniques, to mitigate false negatives. We have thoroughly validated the proposed approach using a custom data set recorded under different scenarios. Validation on a custom data set shows PingCNN performs comparably with 2-D methods, with minimal latency and computational requirements. In summary, PingCNN offers a low-latency, computationally efficient solution suitable for resource-constrained underwater vehicles.