Region-based convolutional neural networks for object detection in very high resolution remote sensing images

Abstract

Recently, the automatic object detection in high-resolution remote sensing images has become the key point in the application of remote sensing technology. The traditional methods, such as bag-of-visual-words (BOVW), could perform well in simple scenes, but when it used in complex scenes, the performance drops quickly. This paper we first try to use the current hot deep learning technology: Region-based convolutional neural networks (R-CNN), to detect aircrafts under the complex environments in high-resolution remote sensing images. This method has been proved to be very efficiency when using in object detection in natural images. Here, we tried to introduce this method into the field of the remote sensing. During our experiments, we also compared the impact of different proposal generate methods on the final detection results. And we also proposed some practical tips to accelerate the detection speed. After detection, we proposed to use a novel algorithm which we called box-fusion, to eliminate the redundant and repetitive boxes that covering the same object. As experiments and results shows, the R-CNN method is much more effective and robust than the traditional BOVW method when dealing with aircrafts detection under complex scenes in high-resolution remote sensing images.