Bridge Detection Algorithm Based on Rotation and Scale Invariance

Abstract

With the development of remote sensing technology and deep neural network, high-resolution optical remote sensing image bridge target detection based on deep learning has become a research hotspot. Bridge target detection is a great challenge because of its arbitrary direction, diverse scale and complex background. In view of the characteristics of bridge targets in remote sensing image, we propose a bridge target detection algorithm based on rotation and scale invariance. Our method is improved based on the DetectoRS network. Aiming at the difficulties of bridge with different scales and multi-directions, we use Recursive Feature Pyramid (RFP) to extract the scale invariant feature and add orientation-invariant model (OIM) to extract rotation invariant feature. In addition, most of the bridge dataset are labeled with horizontal rectangle, it is difficult for network to extract the rotation invariant feature, and the scale feature of bridge will also be blurred. In this paper, a rotated box regression algorithm based on Boxinst, a weakly supervised learning method, is proposed to transform the annotation. A cloud and negative sample data enhancement strategy is proposed since the background of remote images is complicated and there are a lot of false alarms with similar shapes as bridges. The algorithm we proposed in this paper has greatly improved the accuracy of bridge target detection in remote images with complex scenes, and achieved the second place in the preliminary competition in the bridge detection track of the 2020 Gaofen Challenge on the Automated High-Resolution Earth Observation Image Interpretation, with the map of 84.48%.