Satellite Image Detection Based On Deep Learning

In recent years,all countries in the world have vigorously developed the aerospace industry,the application of satellite imagery in various industries has developed rapidly,especially in military reconnaissance,marine ships and fishery management.Due to the scarcity of valuable information in satellite images and the huge scale of satellite image data,intelligent auxiliary tools are urgently needed to efficiently obtain accurate and intuitive information from satellite images.This paper mainly uses deep learning technology to solve the problem of satellite image in the field of target detection..This paper first studies the Faster R-CNN-RBB satellite image target detection model based on non-rotating space.The overall structure is improved based on Faster R-CNN.First,in order to solve the degradation problem caused by the depth of the network and deal with small target objects,a combination of residual network and feature pyramid network is used to form a basic backbone network;Secondly,the RoI Align layer is used instead of the RoI Pooling layer to solve the problem of mismatching of the characteristic regions caused by the loss of the quantized information in the Rol Pooling layer twice..Faster R-CNN-RBB cannot accurately locate the target on the satellite image,and its detection speed is slow.In response to the appeal problem,this paper studies the Darknet53-RBB satellite image target detection model based on rotating space.First of all,in order to deal with the problem that the angle of the target instance will change as the image zooms,this paper implements an angle conversion algorithm.Secondly,two rotation frame coordinates are designed to improve regression training.Then add angle anchors in the prior frame to improve target detection Effect.Finally,an algorithm for calculating the cross-sectional area of the rotating rectangle using the edge-by-side cropping algorithm and Green-Riemann’s theorem and a non-maximum suppression algorithm based on the rotating rectangular space are proposed.In order to verify the effectiveness of the proposed algorithm,the proposed algorithm is verified using three data sets:HRSC2016,Dota and UCAS_AOD.On the HRSC2016 dataset L1,L2 and L3 recognition tasks,Darknet53-RBB’s mAPis 85.8%,72.3%,and 56.9%,exceeding the benchmarks given by the HRSC2016 dataset of 10%,9%,and 6%,Darknet53-RBB’s mAP on aircraft and car identification tasks in the UCAS_AOD dataset is 92.1%.From the perspective of visualization,the detection effect of Darknet53-RBB is better than Faster R-CNN-HBB.The prediction of Darknet53-RBB is closer to the target instance and the detection effect is clearer when the objects are densely arranged.In comparison of computing speed,The detection speed of Darknet53-RBB can reach 59 FPS.This paper designs experiments to verify the effectiveness of the non-maximum suppression method based on the rotated rectangular space and the two rotation box coordinates for regression training.It proves that the algorithm proposed in this paper has certain practical value and research prospects in the field of satellite image detection.The whole work in this paper has played a positive role in promoting the development of satellite image detection technology.