A Feature Enhanced Scale-adaptive Convolutional Network for Ship Detection in Maritime Surveillance*
Tingting Yao, Bo Zhang, Yuan Gao, Yuxin Ren, Zhiyong Wang · 2023
Accurate ship detection is of great importance in intelligent maritime surveillance and numerous methods have been proposed over the years. However, due to the interference caused by water dynamics and the different sizes of various ships, existing methods are often prone to produce a large number of false alarms or missing detection. Therefore, in this paper, we propose a feature enhanced scale-adaptive convolutional network for effective ship detection in complex maritime scenarios. The proposed model is built on a one-stage deep convolutional network based detector and consists of feature enhancement modules, scale-adaptive modules, and an improved detection head. First, a feature enhancement module is proposed and incorporated into multiple convolutional layers. The statistical information calculated from both channel and spatial domains is fused to alleviate irregular noise and increase the weight of ship feature representation. In addition, the receptive field of deep layer is enlarged via a scale-adaptive module devised, so that the robustness of the proposed model over varying ship scales could be enhanced. Finally, an improved detection head is further applied on different feature maps. A centerness branch is introduced following the combination of classification and regression branches to suppress low-quality bounding boxes for more accurate ship detection. Experimental results on two real-world maritime surveillance datasets demonstrate that our proposed model is able to improve ship detection accuracy with less computational cost against a number of state-of-the-art ones.