Zero-shot lightweight submarine cable detection in side-scan sonar images

• A synthetic dataset that closely mimics the appearance of real side-scan sonar images is created, enabling effective training of our neural network. • An adjustable C2f module is proposed for customization of channels and sizes of feature map according to the physical information of interested objects. • A convolutional block attention module is integrated into the model for improved focus on target regions with limited computational overhead. • Validations on simulated and real-world side-scan sonar images demonstrates effectiveness of our method. Side-scan sonar imaging is a critical tool for submarine cable maintenance. However, the lack of available datasets for training deep-learning models poses a significant challenge for cable detection in side-scan sonar images. In this paper, we propose a zero-shot lightweight detector to address this issue. First, a lightweight deep neural network is built, which consists of a backbone with an advanced cross-stage partial fusion module and a rotating box detection head, making it highly suitable for detecting slender targets such as cables. Moreover, an improved cross-stage partial fusion module is designed with adjustable convolution kernels and channel parameters, enabling flexible feature map customization. Furthermore, the convolutional block attention module is introduced to enhance the ability to focus on useful features. A simulated dataset is created by fusing a real side-scan sonar image with a large number of seabed topography images. Experimental results demonstrate that our model outperforms the latest and well-established models, achieving a 37.40 % reduction in training time, a 13.04 % increase in detection speed, and a 17.57 % improvement in mean average precision.