Significant progress has been made in class-balanced benchmarks for ship detection. Nonetheless, their performance in real-world scenarios is subpar, even falling short of expectations. This is primarily attributed to the long-tailed distribution of categories. To address the aforementioned challenges, this paper proposes a ship long-tailed network, named SLT-Net, for fine-grained detection of ships in optical remote sensing images. In particular, SLT-Net employs ResLT as the backbone network, combined with Mutual Information (MI) Loss to enhance the detection performance. Firstly, ResLT takes into account the parameter space and aims to reserve specific capacity for tail classes, thereby rebalancing the representation of head and tail classes. Then, MI Loss, instead of Cross-Entropy, is employed to minimize the discrepancy between the real probability distribution and the predicted probability distribution by fitting mutual information. Experimental results on the ShipRSImageNet dataset demonstrate that SLT-Net outperforms existing SOTA networks, showcasing the best performance in terms of ship detection.