Ship detection has important practical significance in military and livelihood fields, and the rapid identification and accurate positioning of ships help monitor maritime traffic, thus ensuring maritime navigation safety and safeguarding national maritime rights and interests. However, in practical scenarios, the arrangement of ships mostly has the characteristics of dense and different directions, which makes the horizontal frame detection no longer applicable to ship detection. Therefore, in this paper, we use an improved rotating frame detection network based on RetinaNet, named IS-net. In order to fully extract ship features to improve detection accuracy, the ConvNext network is used as the backbone network. In addition, we use Kullback-Leibler Divergence (KLD) as the regression loss. KLD (and its derivative) can dynamically adjust the parameter gradient according to the characteristics of the object, which makes the regression effect more accurate. In this paper, we use the well-known HRSC2016 dataset, and a large number of ablation experiments show that IS-net has excellent detection accuracy.
IS-Net: Improved Ship Detector Based on RetinaNet
2022-11-01
590500 byte
Conference paper
Electronic Resource
English
Ship Detection from Satellite Imagery Using RetinaNet with Instance Segmentation
Springer Verlag | 2023
|