Identifying Tonal Frequencies in a Lofargram with Convolutional Neural Networks

SONAR signal detection is widely used to detect and recognize objects in ocean environment, such as fishes, sea mines, ships or submarines. Because the analysis process of sonargram is time-consuming and difficult even to the expert sonar technician, many previous approaches attempted to automate the process. Recently, convolutional neural networks (CNN) are used in many computer vision problems as feature extractor and predictor, and the performance overwhelms existing approaches. In this paper, we use convolutional neural network models to identify tonal frequencies in a lofargram. We divide a lofargram into several small patches, and a CNN model predicts the probability that the patch is from a tonal frequency. Our model shows 92.5% of precision and 99.8% of recall, and 0.150 seconds of processing time for an inference batch at a specific time frame.