University of Twente Student Theses


Semantic segmentation of urban airborne oblique images

Liu, Li (2019) Semantic segmentation of urban airborne oblique images.

[img] PDF
Abstract:Computer recognition and classification of remote sensing digital images is an important research direction in remote sensing image processing.A lot of classification methods has been widely used for urban scene understanding,the large availability of airborne images necessitates an image classification approach that automatically learns relevant features from the rawimage and performs semantic segmentation in an end-to-end framework. For this, deep learning methodologies werepreferred, specifically in the form of convolutional neural network (CNN)which is widely used in computer vision. In this research, airborne oblique imagesare used as data source for semantic segmentation in urban areas, we investigate the ability of Deep Convolutional Neural Network(DCNN)structures that is designed based on U-net architectureand adaptitto our dataset.Different from its original architecture based on VGG11,U-net based on VGG16 architecture with deeper layers isappliedin this study.However, the training time ofdeep neural network is too long because of enormous parametersandpresent some challenges for model training. Here, depthwise separable convolution is coupled with convolution block in our architecture to reduce the model parametersand improve model efficiency.Outputs of deep neural network are sometimesnoisy and the boundaryof objects are not very smoothbecause continuous pooling layers reduce the ability of localization.As a result, the mean field approximation to the fully connected conditional random field (CRF) inference which models label and spectral compatibility inthe pixel-neighbourhood is applied to refine the output of other classifiers.Finally, we achieved anend-to-end modelwhich connect the scores of the U-net classification with a fully connected Conditional Random Field (CRF) architectureand make the inference as Recurrent Neural Network (RNN), to consider the local class label dependencies in an end-to-end structure and make full use ofDCNNs and probability random field model. We compare the fully connected CRF optimized segmentation results with those obtained by applying the trained U-net classifier.Segmentation results indicate that our classification U-net modelwith deeper layershave performed favourably, the modified network not only ensures the accuracy but also greatly shortens the time of model training in the large-scale image classification problem,furthermore, this end-to-end model which combines fully connected CRF has effectively improved the segments of U-net.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page