Object Detection In Remote Sensing Images Based On Deep Learning

With the rapid development of artificial intelligence and computer vision deep learning based object recognition is a research hotspot in the field of remote sensing image processing.With the development of remote sensing satellite technology,the resolution of remote sensing satellite images is getting higher and higher.At the same time,the demand for accurate detection and identification of remote sensing objects is also increasing.Traditional remote sensing object recognition methods are based on the characteristics of artificial design,and the accuracy of recognition is not high,so it is difficult to meet the requirements of all kinds of object recognition in high-resolution remote sensing images.Deep learning has achieved great success in the field of object detection and has great potential in the field of remote sensing object detection.In this thesis,remote sensing object detection algorithm based on deep learning is lucubrated.First,in view of the remote sensing object direction,more difficult to e×tract the characteristics of high quality problems,in this thesis,based on the deformable convolution network,using deformable convolution operation and deformable pooling operations,the posture,remote sensing object feature e×traction of scale sizes,can get higher quality characteristics,and then the multi-layer characteristic figure forecast candidate area,in order to improve the detection performance of the network for small object,finally realizes the detection accuracy has increased,but not satisfactory in terms of processing speed.Therefore,we adopted the network structure based on YOLOv3 and Darknet-53 for remote sensing object detection,and the processing speed was effectively improved when the accuracy rate was almost not decreased and some categories were even improved.Furthermore,in this thesis,the more lightweight Tiny Darknet is used to replace Darknet-53,and the detection speed is greatly improved at the expense of some accuracy.Parts of the remote sensing image object often presents the intensive distribution,the characteristics of object detection is very commonly used in the field of the great inhibitory action will give too much overlap testing box,but it is easy to cause the residual in the remote sensing object detection,this article puts forward the direction based on adaptive YOLOv3 network,introduce the detection method of rotation in the YOLOv3,through the rotation of the joint goals more rectangular box,reducing the object detection under intensive distribution box overlap rate,thus greatly reducing the miss rate.The research in this thesis is oriented to the actual needs of remote sensing object detection field.Considering the particularity of remote sensing object detection,it improves the mainstream method and improves the detection rate and speed.