Ship-Radiated Noise Recognition Based on Swin-Transformer

We propose a method for underwater object detection utilizing low frequency analysis and the recording spectrum (LOFAR) of ship-radiated noise, along with a hierarchical visual transformer improved by shifted windows. The LOFAR spectrum provides an intuitive representation of the time-frequency distribution of target signal energy, and the typical approach involves feature extraction using convolutional neural networks (CNNs). However, The limited receptive field of CNNs makes it difficult to capture global features such as color characteristics, texture features in the LOFAR spectra. Motivated by the inherent self-attention mechanism in Transformer models, we introduce the Swin Transformer model with the ability to capture global relationships. Building upon the traditional Transformer architecture, Swin Transformer incorporates the concept of feature extraction from convolutional neural networks and applies it to the task of obtaining multi-scale features of images in a hierarchical manner. By computing attention within shifted windows, it ensures that the computational complexity remains linear with respect to input image size. Moreover, employing the Shifted Window approach facilitates information propagation among different windows, significantly enhancing computational efficiency and the model's ability to capture global information. Therefore, we attempted to apply Swin Transformer to the recognition and classification of LOFAR spectra. Experimental results indicate that utilizing Swin Transformer for recognition comparable performance to using CNNs.