随着人工智能技术的进一步发展，深度学习方法在船舶检测领域发挥着重要作用。然而，深度学习算法出现的虚警和漏检，对船舶检测领域技术的应用存在一定的阻碍。虽然经典的深度学习方法能够有效处理单一背景的海面，但是当面对复杂背景之下的数据时，经典模型很容易得出岸上的虚警。并且在常规训练中，模型常常对一些显著特征过于关注，出现特征过拟合现象，当这些显著特征发生改变时极易出现漏检。在模型对输入进行前向传播的过程中，模型中不同网络层会对输入生成对应的映射，也就是特征图。充分利用特征图的语义信息和空间信息是一种有效减少虚警和漏检的方法。与传统模型相比，我们提出的特征图强化网络可以充分利用特征图生成自适应特征图掩码与水陆分割掩码，避免模型的特征过拟合与削弱复杂背景造成的影响，最终达到减少虚警与漏检的目的。在现有公共数据集上与算法模型的对比实验结果表明，本文所提出的方法的性能更为出色，超过了其他SOTA算法。

Monitoring ships in target waters is a vital guarantee of territorial sea security. The classical deep learning method has notable results in a single background sea surface. If there are complex backgrounds such as ports and coastlines, classical models are easily prone to poor performance. This letter proposed a novel method to deal with this problem by utilizing feature-maps. Compared with traditional models, the FMRNet we proposed can significantly improve detection accuracy and effectively overcome problems caused by complex backgrounds. Experimental results on available public datasets show that the proposed method obtains excellent performance and a low false alarm rate.