AFF-LightNet: A Lightweight Ship Detection Architecture Based on Attentional Feature Fusion

Efficient mobile detection equipment plays a vital role in ensuring maritime safety, and accurate ship identification is crucial for maritime traffic. Recently, the most advanced learning-based methods have markedly improved the accuracy of ship detection, but these models often face huge challenges on resource limited mobile devices due to their large size and high computational requirements. Thus, we propose a lightweight ship detection network based on attentional feature fusion, called AFF-LightNet. To enhance the fusion of multi-scale features and semantically inconsistent features, we introduce iterative attentional feature fusion (IAFF) into the proposed neck network, improving the efficiency of feature fusion by introducing a multi-scale channel attention module. Also, deep and cross network version 2 (DCNv2) is replaced by Convolution (Conv) in the backbone network to further improve the detection accuracy of the proposed network. It enhances the spatial sampling position in convolution and Region of Interest (Rol) pooling by introducing offsets. Moreover, a lightweight convolution GhostConv was introduced into the head network to reduce the number of parameters and computation cost. Last, Scalable Intersection over Union (SIOU) loss was leveraged to improve the convergence speed of the model. We conduct extensive experiments on the publicly available dataset SeaShips and compare it with existing methods. The experimental results show that compared with the standard YOLOv8n, the improved network has an average accuracy of 98.8%, an increase of 0.4%, a reduction of 1.9 G in computational complexity, and a reduction of 0.19 M in parameter count.