School of Marine Science and Technology, Northwestern Polytechnical University, Xi'an 710072, China.
Naval Research Academy, Beijing 100161, China.
J Acoust Soc Am. 2020 Oct;148(4):2182. doi: 10.1121/10.0002172.
This paper investigates the problem of dim frequency line detection and recovery in the so-called lofargram. Theoretically, long enough time integration can always enhance the detection characteristic. But this does not hold for irregularly fluctuating lines. Deep learning has been shown to perform very well for sophisticated visual inference tasks. With the composition of multiple processing layers, very complex high level representations that amplify the important aspects of input while suppressing irrelevant variations can be learned. Hence, DeepLofargram is proposed, composed of a deep convolutional neural network and its visualization counterpart. Plugging into specifically designed multi-task loss, an end-to-end training jointly learns to detect and recover the spatial location of potential lines. Leveraging on this deep architecture, performance limits of low SNR can be achieved as low as -24 dB on average and -26 dB for some. This is far beyond the perception of human vision and significantly improves the state-of-the-art.
本文研究了所谓的 lofargram 中暗频线检测和恢复的问题。从理论上讲，足够长的时间积分总能增强检测特性。但这不适用于不规则波动的线。深度学习在复杂的视觉推理任务中表现非常出色。通过多个处理层的组合，可以学习到非常复杂的高级表示，这些表示放大了输入的重要方面，同时抑制了不相关的变化。因此，提出了 DeepLofargram，它由一个深度卷积神经网络及其可视化对应物组成。通过专门设计的多任务损失，端到端训练可以共同学习检测和恢复潜在线的空间位置。利用这种深度架构，可以实现低至平均-24 dB、某些情况下低至-26 dB 的低 SNR 性能极限。这远远超出了人类视觉的感知范围，并显著提高了现有技术水平。
