Two-stage Airplane Detection with NMS Filtering in Remote Sensing Images

Abstract

In the past few years, object detection based on deep learning have attracted attention from more and more organizations and researchers. Compared to one-stage object detection methods, two-stage methods would display a better performance of accuracy and precision. As airplane detection is a basic task in remote sensing images, we propose an airplane-detection method based Faster R-CNN and Feature Pyramid Networks (FPN), with Non-maximum Suppression (NMS) postprocessing. The Faster R-CNN is the most widely used two-stage detection framework, and is still the mainstream box-detection method in many famous detection research platforms like Detectron2 and MMDetection. For the various sizes of airplanes objects, the FPN is used as an excellent technique in recognition systems for detecting objects at different scales. Due to the prominent similarity between different classes of airplanes, the naive two-stage method would yield many duplicate boxes of multiple airplane classes for one object. To improve the recall and precision of the detection model, an NMS Filtering is proposed to prevent the phenomenon of multiple duplicate boxes for one object. The experiment showed that our method is able to accomplish the task in remote sensing for the detection and recognition of airplanes in 24 different classes including helicopter and wing aircrafts, and the NMS postprocessing would have a positive influence on improving the recall and mean average precision (mAP) metrics. The future work would be expanded on improving the precision of classification task.