Underwater Acoustic Object Discrimination for Few-Shot Learning

Abstract

Deep neural networks are the go to algorithm when it comes to image classification [1-5]. This is partly because they can have arbitrarily large number of trainable parameters. However, this comes at a cost of requiring a large amount of data, which is sometimes not available. In the field of underwater acoustic object discrimination, data are scarce, no public dataset can be used for training, and data can only be collected by sea trials. In order to mitigate such an issue, we choose a few-shot learning with Siamese networks method. Siamese networks are a special type of neural network architecture. Instead of a model learning to classify its inputs, the neural networks learn to differentiate between two inputs. We trained our model to use only eighty single beam trajectories, divided into eight types of objects, and calculated the differences between each type of object, and the result show that our Few-shot Learning with Siamese networks can achieve good performance, we can achieve the target stable tracking on the multi-beam graph by the result of the difference between each type. This provides an important reference method for the application of deep learning in the field of underwater acoustic object recognition.