Remote Sensing Image Object Detection Based On Faster R-CNN

Remote sensing image object detection is to determine the category and location of the object to be detected in a given remote sensing image.Traditional object detection algorithms based on machine learning usually require manual feature design,but there is no universal feature design method,so this kind of method is not universal in the application of remote sensing image object detection.With the continuous development of deep learning,the feature expression ability and generalization ability learned by deep convolutional neural network are stronger.Faster R-CNN algorithm,as the most widely used and widely applied deep learning model at present,performs object detection on images obtained from natural scenes with generally high accuracy.Therefore,Faster R-CNN algorithm was selected for remote sensing image object detection.However,due to the complexity of remote sensing image background and object to be detected is comparatively small,Faster R-CNN algorithm can not be used for remote sensing image object detection with high accuracy.Aiming at this problem,an improved Faster R-CNN algorithm is proposed,which mainly includes the following three aspects.(1)The size and diversity of the training set directly affect the training effect of the network model.The more training samples and the better diversity,the better performance of the trained network model.Data expansion processing was carried out through defogging operation,horizontal inversion operation and counterclockwise rotation operations of 90,180 and 270 degrees,and the expanded training set was used for model training.The experiment verifies the expansion of the training data set and the diversification of samples,so that the network can fully learn the characteristics of remote sensing images,so as to train a better tobject detection model of Faster r-cnn.(2)Due to the complexity of remote sensing image background and object to be detected is comparatively small,the feature extraction network in Faster R-CNN algorithm cannot fully extract the features of small and medium objects in the image.At the same time,Region Proposal Network in Faster R-CNN algorithm extracts a large number of background as candidate region,so that more background samples and fewer foreground samples can be obtained,and Faster R-CNN algorithm cannot fully learn the foreground information,thus reducing the detection performance.Therefore,aiming at the former,eature pyramid balance strategy is adopted to make full use of features at different levels and achieve the balance between features at different levels,so as to fully extract small object features.Aiming at the latter,class balance strategy is applied to reduce the influence of a large number of background on the network model and to mine difficult samples,so as to balance the foreground and background categories and guide the model to converge better.These two strategies were used to improve Faster R-CNN algorithm.On homemade remote sensing image data set has carried on the experiment,the contrast experimental results show that the improved Faster R-CNN algorithm has good performance,and detection accuracy is improved by approximately 7% in comparision with original Faster R-CNN.Inaddtion,in NWPU VHR-10 and RSOD-Dataset on two remote sensing image data sets were tested,the test results show that the improved model has stronger generalization ability.(3)A remote sensing image object detection system based on Faster R-CNN was designed and implemented.The system is used for the selection,detection and display of the results and performance evaluation of the images to be tested.And the system can automatically detect objecs in remote sensing images,and can quickly and accurately detect interested objects and display the results.