Semantic Segmentation In Remote Sensing Through A Heavy Feature Reuse Based Network

Semantic segmentation is the task of clustering pixels into an ob-ject class.In the field of remote sensing semantic segmentation has wide applications ranging from scene cover classification to change detection for scene understanding.With the success of deep learning algorithms for classification tasks,there has been much work to apply convolutional neu-ral networks in remote sensing with much success.However,feature ex-traction of high resolution remote sensing imagery poses a challenge when applying such networks.In particular,there is a need to extract high level features while maintaining an objects resolution in the networks feature space.This masters thesis proposes an efficient deep fully convolution architecture that obtains high level features without loss of spatial resolu-tion by replacing the standard convolutional layers in U-Net with dense residual blocks.By stacking identity blocks,we allow the input to flow through the network at every proceeding layer.Our network is termed DRU-Net,and is shown to outperform standard U-Net.We further add to this by using a pyramid pooling layer as a global context prior before the decoding layers of our FCN.This additional step further improves this re-sults.We then perform several analysis of different data set,and show an application of fully convolutional networks in the context of radar images.