Road Segmentation And Extraction Of High Resolution Aerial Imagery Based On Recurrent Neural Networks

Object detection and recognition is an important research field of computer vision.In the past decade,the coverage and the resolution of Aerial imagery have been improved gradually,which provide a rich data source for target detection and recognition.However,relying solely on traditional manual interpretation is clearly far from meeting the needs of users.Therefore,in decades,introducing the artificial intelligence technology,especially the technology of machine learning and deep learning into automatic target classification and recognition algorithm is currently a hot topic in the field of computer science and information engineering.Automatic extraction of road network information from Aerial images is one of the most important research topics.With the resolution of the image gradual improvement,the original linear road extracting methods suitable for the low resolution images are not suitable for the automatic extraction of the road surface-area in high resolution images.In this thesis,taking road surface area automatic extraction as the research object,and the road segmentation and extraction problem of Aerial images is studied.The research contents can be conclude as follows:Firstly,introducing the theory of machine learning,the support vector machine(SVM)and hypergraph method for road segmentation of Aerial images are studied.For the support vector machine method,the super pixel segmentation technology is used to accelerate the processing speed.At the same time,the Hough transform is used to extract the road boundary,which improves the result of road segmentation.However,the hypergraph method takes the neighborhood information into account,so the result is better used in the road segmentation of the aerial images.At the same time,super-pixel segmentation is used to improve the efficiency,and morphological algorithm is used to optimize the results.Secondly,introducing the theory of deep learning,the recurrent neural network is studied,and the convolution neural network and the conditional random field are combined to form an end-to-end network.The method of multi-scale sample training and morphological post-processing is used to improve the road segmentation effect of Aerial image effectively.Experiments show that for 120 images with 500*500 resolution,scheme using support vector machine and Hough transform takes 35s,which reach a recall rate of 93.28%,i.e.,3.75%higher than the second best method,with a quality score of 78.77%.In contrast,the recurrent neural network costs 0.5s per frame,with a recall rate of 89%and a quality score of 82.84%,which is close to the best performed method.However,the overall performance of RNN based method is outstanding and more robust.Quantitative and qualitative analysises show that the two methods have respective advantages and disadvantages:The former is applicable to the scenarios with limited training data,while the latter is suitable for the scenarios with complex road conditions and large training sets.