Style Transfer-Based Data Augmentation for Side-Scan Sonar Image Classification

Side-scan sonar (SSS) has been widely employed in ocean exploration due to its stability and efficiency on autonomous underwater vehicles (AUVs). However, the difficulty in acquiring SSS images and the limited availability of training samples significantly limit the performance of the deep neural network (DNN)-based SSS image classification. To address this challenge, this paper proposes a multi-scale attention network (MSANet) that utilizes optical-acoustic image pairs for data augmentation to enhance the generalization ability of SSS image classification models. Specifically, shallow and deep features are extracted from different layers of the encoder to comprehensively capture content and style information. Subsequently, a multi-scale attention module (MSAM) is introduced to extract both local and global contextual information of style features along the channel dimension. These style features are then effectively fused with optical features to achieve precise spatial alignment of optical and acoustic features. Finally, the fused features from different layers are rescaled and fed into a decoder to generate high-fidelity synthetic SSS image samples, which are used to train the SSS image classification network. Experimental results on real-world SSS datasets demonstrate that the proposed style transfer approach effectively generates high-quality synthetic SSS image samples, thereby improving the DNN-based SSS image classification performance.