The utilization of forward-looking sonar (FLS) has become increasingly significant for detecting underwater objects. However, despite its widespread use, target recognition tasks for underwater unmanned vehicles (UUVs) are still predominantly carried out through postprocessing techniques. This can be attributed to the limitations of computing devices deployed on UUVs, which have restricted computational capabilities due to power constraints. Furthermore, conventional object detection approaches designed for natural images are inadequate when applied to sonar images with high levels of noise. To address these challenges and enable UUVs to autonomously detect underwater objects, we propose an innovative solution that involves the integration of an improved version of YOLOv8 with attention mechanism, allowing the model to obtain more feature information while reducing the number of parameters required. Additionally, we developed a comprehensive preprocessing module consisting of removal, contrast enhancement, and noise filtering techniques. Our experimental results demonstrate that the proposed method achieves high accuracy, robustness, and can be efficiently implemented on hardware processors. Our code is available on https:/github.com/wangzitao7777/underwater-detection.