Style Transfer-Based Augmentation for Side-Scan Sonar Images

Side-scan sonar(SSS) has been extensively adopted in ocean exploration because of its stability and efficiency when deployed on autonomous undersea vehicles(AUVs). Nevertheless, the difficulty in acquiring SSS images and the limited availability of training samples severely constrain the performance of the deep neural network(DNN)-based SSS image classification. To mitigate this limitation, this paper proposed a multi-scale attention network(MSANet) that utilized optical-acoustic image pairs for data augmentation to enhance the generalization capacity of SSS image classification models. First, shallow and deep features were extracted from multiple encoder layers to comprehensively capture both content and style information. Next, a multi-scale attention module(MSAM) was introduced to extract both local and global contextual information of style features along the channel dimension. These style features were then effectively fused with optical features to achieve precise spatial alignment of optical and acoustic features. Finally, the fused multi-scale features were aligned and input to a decoder to generate high-fidelity SSS images that were subsequently used to train the classification network. Extensive experiments on real-world SSS datasets demonstrate that the proposed style transfer-based augmentation strategy can effectively generate high-quality simulated SSS image samples, thereby improving the performance of SSS image classification based on DNN.