Fast aircraft detection based on region locating network in large-scale remote sensing images

Abstract

Nowadays, we get more and more remote sensing (RS) images which cannot be well processed or used by manual analysis or existing automatic methods. In the past few years, the object detection technology has greatly developed, especially after the usage of CNN in object detectors. However, Object detection in large scale RS images is still a challenging tasks which needs further study. Compared to natural images, RS images include much more objects in different sizes with a larger scope. Therefore, it is extraordinarily time-consuming to detect small objects in a large-scale RS image, since this work needs more scale and location traverses. Algorithms for common images cannot tackle the problem of some special object detection, like aircraft detection, in RS images. In this paper, we introduce an extra Region Proposal strategy named Region Locating Network (RLN) to improve the Faster RCNN framework. The proposed RLN locates spectacular areas where aircrafts are usually found, like parts of the runway and the parking apron. Based on the locating result, we can use Faster RCNN to detect airplanes in several smaller image regions. Extensive experiments show that the proposed method has oblivious improvement in recall rate, accuracy and computing efficiency.