Side-scan sonar (SSS) is a reliable tool for underwater detection. However, the existing SSS image recognition methods often perform poorly in practical deployment due to their idealistic training. It is important to develop more practical and effective methods for SSS image recognition. Because the real underwater data often encounters open-set and long-tailed distribution issues. To address the problem of open-set long-tailed recognition performance in SSS images, a feature space optimization network (FSONet) is proposed in this paper. The pre-training model based on transfer learning is introduced to address the issue of small datasets, and the feature space of some network branches is optimized to enhance the model’s generalization ability. Secondly, to address the noise problem in sonar images, we propose a high-resolution denoising multi-scale feature block (HDMB) to effectively reduce image noise and extract features of different scales. Finally, the evaluation metrics are given to assess the open-set and long-tailed recognition in sonar images (Sonar-OLTR). The experiment results show that the proposed method performs well and is comparable to the state-of-the-art method (SOTA).