Remote Sensing Image Classification Based On Faster RCNN

In the continuous development of satellite remote sensing and hardware and software technology,more and more computer vision methods and deep learning methods are applied in remote sensing image detection.Remote sensing images are much slower to develop than natural images because remote sensing images are large in size,complex in information,low in contrast,and confusing in foreground and background.In recent years,artificial intelligence technology has begun to develop in a blowout manner,which brings hope to the development of remote sensing images.The Hinton team participated in the Image Net image recognition competition for the first time in 2012.The convolutional neural network created by Alex Net was the first in this competition,and the classification accuracy was far better than the second place in the competition.(SVM method),they successfully demonstrated the potential of deep learning in image recognition.Because of this image recognition contest,the convolutional neural network officially entered the line of sight of many researchers,thus starting the research trend of convolutional neural networks.Later,in 2014,Girshick et al.proposed a model of image detection,R CNN(Regions with Convolutional Neural Network Features),which became the basis of deep learning in the field of image detection.In the following year,the team also proposed two models,Fast RCNN and Faster RCNN,which laid the foundation for two-stage image detection,especially Faster RCNN,which has become a field of image detection.Benchmark.The improvement of remote sensing image technology and convolutional neural networks,especially the success of the Faster CNN target detection field,has greatly promoted the application of Faster RCNN in the field of remote sensing image detection.This paper mainly studies the application of Faster RCNN in remote sensing image object detection,and has completed the following research work:1.A new convolutional neural network layer Echo Layer is proposed.In recent years,feature extraction in deep learning is generally performed by sequential concatenation between convolutional layers,but subsequent feature information cannot be interactively and shared with the feature information extracted in the previous stage.The method uses the multiplexing of the convolutional layer to refer the feature information extracted by the previous channel to the subsequent channel for reference,which can make the feature connection more compact.The experimental results show that the method can improve the effect of the model and can be used in different models;2.Improved the model structure of Faster R CNN.In Fast RCNN,there are fewer network sharing layers.Compared with Faster R CNN,the training rate is slower and the model occupies more memory.Therefore,the Faster RCNN combines the original starting convolution layer with the last fully connected layer based on the Fast RCNN to achieve a two-stage training shared layer structure.This paper shows through experiments that the two different tasks of regression and classification,using the same shared layer structure,will lead to weakening of the model.This paper proposes to split some shared layers,which can further improve the performance of the model;3.A variety of Faster RCNN-based models were used to detect multiple types of targets for remote sensing images.Trying to combine the models of Mobile Net,Res Net,Dense Net and their ideas with the Faster RCNN model,through the migration learning,fine-tuning the Image Net pre-training model,and then migrating into the various improved models of Faster R CNN.