Multi-scale ship detection algorithm based on global channel attention

The current detection methods concerning ship detection suffer from inadequate extraction of image target feature information, false detection and omission under complex background disturbance, and ship scale diversity. The primary factor lies in how the feature extraction process of a Convolutional Neural Network (CNN) focuses on local spatial operations, overlooking the correlation among pixel points within the image and across the entire region, along with the interplay between different channels, and the existing ship detection models are not sufficient for the semantic information of the high-level features and the fine-grained feature fusion of the underlying features. Based on the YOLOv7 detection model, we propose a CAPN module that combines the traditional asymmetric pyramidal nonlocal network with channel attention to capture more contextual information and enhance the correlation between channels. Additionally, the concept of Adaptive Spatial Feature Fusion (ASFF) is brought in to blend the semantic details of the high-level feature output from the network with the intricate fine-grained features, making the feature layer fusion more adequate. In this paper, the improved detection algorithm is named YOLOv7-CAPA.Experimental results on a dataset based on a coastal complex background (SeaShips) show that the algorithm improves the detection of ship targets in complex backgrounds and also copes with various complex scenarios that may be encountered in handling ship detection tasks, providing a better basis for subsequent complex tasks such as ship tracking and ship intent recognition.