Rotated Ship Detection with Improved YOLOv5X

Abstract

Ship detection in optical remote sensing images is a vital yet challenging task. Now, more attention has been focused on increasing detection accuracy, while the detection speed is ignored. However, detection speed is as important as detection precision for ship detection. In this paper, we propose a new model, named ImYOLOv5X, which is based on YOLOv5X combined with a Squeeze-and-Excitation Module for fast and accurate rotated ship detection. Firstly, we incorporate a Squeeze-and-Excitation (SE) module into backbone of YOLOv5X, which enables the model to focus on detection objects, thus improving detection accuracy. Then we design an easy-to-insert module, containing a Convolution Set and Squeeze-and-Excitation Module (CS-SE), which can extract features and weigh the channels of features for prediction. Finally, we introduce the Gaussian Wasserstein Distance (GWD) loss as the regression loss of the model. The GWD loss resolves the boundary discontinuity and inconsistency in training and final detection metric. Extensive experiments on the HRSC2016 dataset show that our model can achieve highest detection accuracy and still maintain fastest detection speed compared with some other models, which proves the effectiveness of our model.