<p>Underwater sonar images, characterized by low resolution, weak textures, and high-noise levels, pose significant challenges for object detection tasks. Given the demands for real-time perception and deployment efficiency during autonomous underwater vehicle operations, designing lightweight and robust detection models becomes crucial. To address this, an improved detection model named lightweight sonar YOLO (LS-YOLO) is proposed based on the YOLOX-Nano framework. Compared with existing models, LS-YOLO integrates three key components: (i) a frequency-aware cross-layer fusion module based on discrete cosine transform introduced in the multi-scale feature fusion stage to enhance information interaction across layers; (ii) a residual shrinkage and enhancement module designed for the feature extraction stage to suppress noise interference and improve response to blurred boundaries and small objects; (iii) an uncertainty-guided multi-task loss based on uncertainty modeling with Laplace distribution, enhancing model robustness to ambiguous targets and outlier samples. The experimental results show that LS-YOLO achieves notable improvements over the original YOLOX-Nano in key metrics, such as mean average precision and inference speed, while significantly reducing the number of model parameters. Compared with other state-of-the-art lightweight detectors, LS-YOLO maintains excellent detection performance with fewer parameters and lower computational complexity. In addition, Grad-CAM visualization confirms the effectiveness of each module in terms of feature perception and target localization.</p>