FM-YOLO: A Ship Detection Algorithm Based on Feature Fusion and Multi-Scale Attention

Ship detection technology is one of the key components of ship intelligence and an important aspect of maritime transportation safety. However, due to the diversity of ship types and the complexity of coastal environments, traditional detection algorithms are not ideal. The reason lies in the fact that Convolutional Neural Networks (CNN) overly focus on local spatial operations during feature extraction. Therefore, based on the YOLOv7 model, we propose a detection algorithm that incorporates feature fusion and multi-scale attention mechanisms (FM- YOLO). First, we introduce the lightweight GhostConv into the extended efficient aggregation network layer of YOLOv7, resulting in a new model (FELAN). This model can more comprehensively capture contextual information, preventing overemphasis on local features and thereby enhancing target features. Second, we enhance the feature fusion capability by combining a lightweight spatial pyramid pooling (SPPCSPC) with global channel pooling, further boosting detection performance. Additionally, we conceptualize a multi-scale channel spatial attention mechanism, CASANet (Coordinate and Spatial Attention Network), to enhance attention on target regions, improve small-scale target detection, and adapt to multi-scale targets. The model is tested on the SeaShips dataset with complex coastal backgrounds. Compared to the traditional YOLOv7 model, the results show that this algorithm improves the ability to detect ships in complex backgrounds.