Abstract: This paper proposes a multi-scale adaptive detection algorithm for Circular Synthetic Aperture Sonar (CSAS) images based on an improved YOLOv8 framework. First, a multi-scale feature enhancement module is developed, which leverages multi-branch convolutional operations and cross-scale feature fusion mechanisms to strengthen the model’s capability in characterizing small and indistinct targets. Second, a hybrid attention mechanism is designed, integrating channel compression and self-attention query strategies to dynamically enhance the feature response of target regions while suppressing background noise. Furthermore, to mitigate the scarcity of real-world sonar data, a neural network-based data augmentation method is introduced, generating synthetic training images that adhere to sonar-specific imaging characteristics, thereby effectively augmenting the dataset and improving data diversity and physical consistency. Experimental results demonstrate that the proposed algorithm achieves a mean Average Precision (mAP) of 89.7%, surpassing the baseline model by 6.3%, and validates its effectiveness in detecting CSAS images.