The Research On Object Detection In Remote Sensing Images Based On Convolutional Neural Network

Remote sensing technology is a technique that uses sensors to detect objects over long distances.With the rapid development of remote sensing technology,the data we can obtain is also becoming more and more abundant.Remote sensing will have broad application prospects in many fields such as agriculture,forestry,ecology and military.How to deal with remote sensing data efficiently will become an important problem to be solved.Remote sensing images can be divided into aerial photography,microwave image,etc.,and aerial images are closer to ordinary images.The research of this paper is mainly in aerial image.The main study of this paper is about object detection in aerial images.We introduce a convolutional neural network algorithm in deep learning to achieve object detection in aerial images.Aerial images usually have some special features: high resolution,scale diversity,dense arrangement of small objects,arbitrary orientation,and so on.In order to adapt to the particularity of aerial images,we have designed new data augmentation methods to facilitate the learning of neural networks.We designed two regression methods for remote sensing images,based on vertex regression and angle regression,and optimized both methods to further enhance the learning ability of the network.At the same time,above the detection task,we only use the rectangular box information and add a branch to do the segmentation task,and use segmentation result to optimize detection result.We experimented based on the generic Faster R-CNN algorithm framework using new data augmentation methods.We improved both regression methods and used the segmentation task for optimization.We obtained an optimal mAP of 73.69% on the DOTA dataset,which is nearly 10 percent higher than the benchmark based on Faster R-CNN algorithm.