The utilization of sonar imaging for detecting underwater targets is crucial for both maritime trade and military protection at sea. Current target detection methods employing underwater sonar images primarily rely on traditional machine learning techniques. These approaches necessitate extensive a priori knowledge and face difficulties in adapting to the low resolution of sonar images, blurry target edges, and severe distortion. Consequently, target recognition accuracy is suboptimal, particularly for detecting smaller targets. To address these challenges, this study introduces an adaptive global feature enhancement network for small target detection in underwater sonar images, building upon the original YOLOv7 model. To achieve richer gradient information, the focus of the network is directed to the target rather than the background by parallelizing additional gradient flow branches, we integrate the C2f feature extraction module of YOLOv8 into Multi-Headed Self-Attention (MHSA) to form a new model C2fMHSA.Subsequently, SPD-Conv is incorporated to preserve fine-grained information and minimize the impact of less-effective features during the convolution process. This adaptation allows the model to accommodate the low-resolution nature of sonar images and the high prevalence of small targets. Lastly, by replacing the original Intersection over Union (IoU) metric with Normalized Wasserstein Distance (NWD), the model’s sensitivity to small targets is decreased, effectively enhancing the network’s detection results demonstrate that the proposed model significantly outperforms three baseline models in terms of recognition accuracy, highlighting the value of this innovative approach to underwater target detection.