Research On Remote Sensing Image Scene Classification And Detection Based On CNN

Owing to high resolution of remote sensing images,they are generally used to city planning,military affairs and so on,so remote sensing images have significant research value on research and practical applications.The following is the main content of this thesis:Firstly,some introductions of application about remote sensing images by interpreting the characteristics are made,which appeals to so many researcher in this field devoting themselves into the research about pattern recognition of remote sensing images.Next,list some results in this field all over the world.In addition,some conceptions and theory of training of convolutional neural networks are elaborately described here.Secondly,in this thesis,a method fine-tuning a pre-trained network model to train convolutional neural networks is employed to accomplish remote sensing scenes classification.By means of fine-tuning,network training is less time consuming and can be avoided over-fitting.In the meantime,those more distinguishing features can be extracted from the remote sensing images.Thus,the proposed method provided an excellent performance than the state-of-the-art methods.Thirdly,a framework name Faster R-CNN constructed by two convolutional neural networks is employed in this thesis.Both convolutional neural networks share the hidden convolutional layers and make alternating optimization at the phase of training.When testing on an image,the region proposal network will generate a large number of proposal regions,then recognition can be done using the second network.In fact,the resolution and scales of images dataset in the experiment is very large.The priority is clipping the images and making annotation,then do the training using those small scale images while using large scale image at the testing phase,in order to obtain good detection performance.By using this proposed method,not only in pure ocean scene but also in those complex scenes(such as around islands or harbors),can be obtain significantly excellent performance.Besides,this method is also an end-to-end one and cost extremely short time in detecting,satisfies the requirement of real-time,which serves better for many tasks.