Maritime small ship detection is a challenge problem in computer vision. At present, YOLOv3 network is widely used for object detection, but it gets low recall rate and detection accuracy for small objects in the complex ocean environment. Addressing this problem, we improve the backbone and predicted network of YOLOv3 network for detecting maritime small ship. Firstly, we build a maritime small ship dataset including four kinds of scenes: small traffic flow and heavy traffic flow in sunny and foggy weather. Secondly, we use K‐means to re‐cluster the anchor box for matching the shape of maritime ship. Thirdly, we introduce spatial pyramid pooling (SPP) module and frequency channel attention (FCA) module, and redesign the structure of YOLOv3 network, called it as SPP‐FCA‐YOLOv3. Here SPP module is used to fuse local features with global features and enriches the expression capability of the feature maps. FCA module emphasizes important object feature and suppresses unnecessary noise. Experimental results show that proposed SPP‐FCA‐YOLOv3 has higher detection accuracy for maritime small ship detection, getting a 2.2% improvement in average precision compared with YOLOv3, and a 1.2% improvement in average precision as well as higher speed compared with YOLOv5.