Object Detection on Aerial Image by Using High-Resolutuion Network

Abstract

To detect trespassing in images captured by drones for water conservancy facilities inspection, this paper proposes a method that adapts Hight-Resolution Net to reserve high resolution features for improving detecting results. To detect trespassing target with small scale, this method parallels low-resolution and high-resolution conventical feature maps to reserve high-resolution features, besides that multi-scale fusions are conducted to enhance feature maps with different resolutions. Compare to Faster R-CNN, proposed method achieves 1.7% higher mAP on small targets.