Ultra-Lightweight Network for Ship-Radiated Sound Classification on Embedded Deployment
Sangwon Park, Dongjun Kim, Sung-Hoon Byun, Sangwook Park
TL;DR
The paper addresses ship-radiated sound classification on resource-constrained embedded platforms, a setting where real-time inference and small model size are critical. It proposes ShuffleFAC, a lightweight architecture that combines Frequency Adaptive Separable Convolution (FASC) with a ShuffleNet-style backbone to achieve frequency-sensitive feature extraction at low cost. On the DeepShip dataset, ShuffleFAC with channel scaling γ=16 achieves a macro F1 of $71.45\pm1.18\%$ using $39\,K$ parameters and $3.06\,M$ MACs, with a Raspberry Pi 5 inference time of $6.05\pm0.95$ ms, improving macro F1 by $1.82\%$ while reducing model size by $9.7\times$ and latency by $2.5\times$ compared to MicroNet0. These results demonstrate the practicality of real-time, energy-efficient UATR on embedded hardware and highlight the value of frequency-aware design for spectrogram-based acoustics.
Abstract
This letter presents ShuffleFAC, a lightweight acoustic model for ship-radiated sound classification in resource-constrained maritime monitoring systems. ShuffleFAC integrates Frequency-Aware convolution into an efficiency-oriented backbone using separable convolution, point-wise group convolution, and channel shuffle, enabling frequency-sensitive feature extraction with low computational cost. Experiments on the DeepShip dataset show that ShuffleFAC achieves competitive performance with substantially reduced complexity. In particular, ShuffleFAC ($γ=16$) attains a macro F1-score of 71.45 $\pm$ 1.18% using 39K parameters and 3.06M MACs, and achieves an inference latency of 6.05 $\pm$ 0.95ms on a Raspberry Pi. Compared with MicroNet0, it improves macro F1-score by 1.82 % while reducing model size by 9.7x and latency by 2.5x. These results indicate that ShuffleFAC is suitable for real-time embedded UATR.