Feature Enhanced Fully Convolutional Network for Ship Detection

Focusing on the problem of high false detection and missed detection in the existing ship detection methods,which is affected by the background noise on sea surface and the various scales of different types of ships,a novel feature enhanced fully convolutional network is proposed for ship detection.First,a multi-scale feature enhancement mechanism is introduced to improve the feature descriptive power of the model for ships.In each single-scale feature extraction stage,the statistical information is calculated and utilized to suppress the in-terference of background noise,such as wave and clutter,on the feature description of ships.Meanwhile,in the multi-scale feature fusion stage,two steps feature adaptive fusion strategy is utilized to improve the network awareness of multi-scale ships.Furthermore,an improved object detection head network is devised under the re-gression analysis framework,which optimizes the calculation of centerness through an attention enhancement mechanism in regression branch to better suppress the low-quality bounding boxes,and improve ship detection performance while ensuring detection efficiency.The mean average precision of the proposed method reaches 91.6%on the large-scale maritime surveillance dataset SeaShips,and the detection speed is 9 frames per second.Experimental results demonstrate the proposed method has good detection accuracy and robustness.