Object Detection of Optical Remote Sensing Image Based on Improved Faster RCNN

Abstract

Object detection of optical remote sensing image is an important and challenging problem. And it is widely used in the field of aerial and satellite image analysis. With the rapid increase of optical remote sensing image data and popularity of convolutional neural network, the problem has attracted lots of attention recently. However, the detection result of images with complex background is unsatisfactory, so as images with dense and small objects. Aiming at these problems, we propose a method that combined Feature Pyramid Network(FPN) and Deformable Convolution Network(DCN) to improve the Faster RCNN framework, which helps to improve the detection result. The improved network combines the low-level structural information and the high-level semantic information together to enhance the feature representation. The shared convolutional layer makes end-to-end training come true. Additionally, deformable convolution network makes feature extraction better. We adopt the proposed framework to implement experiments on DOTA dataset, attaining mean average precision(mAP)value of 0.834 on the testing dataset, which is an increase of 23% than the classic Faster RCNN.