Research On Object Detection In Multi-source Remote Sensing Images Based On Deep Learning Algorithm

Remote sensing is the observation of far-away objects without making physical contact.Modern RS enjoys a boom in images of SAR,multi/hyper spectrum,infrared and laser,which are specialized in a certain aspect.When one technic can not complete certain task alone,the fusion of several of these technics will be adopted alternatively.Fusion technology is to extract and mix information from different sources of images and to interpret images from one another.As the development of RS,the interpretation of massive high-resolution RS images automatically arouses widespread attention recently.Because of the big data tendency,deep learning plays an indispensable role in object detection in RS images.The deep networks extract features that are most useful automatically from images rather than select features manually in traditional ways.These features are always such information mined by deep networks,which are not proposed by researchers.The main work of this paper includes the following aspects.Firstly,Faster RCNN algorithms objects detection are discussed.Taking ZFNet as an example,the specific structure of convolutional neural networks is shown and the BP algorithm is deduced.Furthermore,the idea to transplant convolutional neural networks from image classification to object detection is analyzed,during which,the whole development of Faster RCNN algorithm is reviewed.A tendency that algorithms are to become faster and more unitive is concluded.Secondly,transfer learning is used to finetune Faster RCNN networks on strengthened DOTA database to detect plane objects in remote sensing images.The performance of networks is tested in relation to the multi-directions and multi-scales in remote sensing images.The experimental results show that the deep networks are robust to objects with multi-orientation and multi-scale.More specifically,when the scale of objects become smaller,the counterpart scores are less.Thirdly,DS evidence theory is applied in a feature level to improve the performance of ship detection in GF-2 and Sentinal-1 images.The whole algorithm scheme include pre-process of multi-source images,image registration,water-land classification,proposal region detection and features extraction and fusion.The experimental results show that fusion result is better than that with one single source of images.