DenseResSegnet: A Dense Residual Segnet for Road Detection Using Remote Sensing Images

Road map update in short interval is challenging due to the presence of various occlusion in remote sensing images. This article addresses this issue by using deep learning algorithm. Here in this paper road networks are detected using Dense Residual (DR) Segnet followed by post processing. The basic convolution block in Segnet architecture is replaced by dense residual block for aggregating the contextual information. The dense residual block has multiple level convolution layer with residual connection to aggregate contextual information. The post processing uses geometrical shape analysis to filter the non-road objects left by Segnet segmentation. The proposed method is trained on Massachusetts Road Dataset with fixed number of convolution layer filters in Segnet architecture. It is found experimentally that training with fixed number of filters compress the training time and memory consumptions. The proposed methods achieve promising quality metric of 74.41% as compare to the state of methods present in the literature.