Speaker: Minsung Jang
Abstract: Roads are one of the data types for sample points in NRI and their detection plays important roles in remote sensing imagery. Our main goal is to improve the performance of road segmentation from NRI satellite image data using Deep Learning techniques. We provide two different approaches to achieve the goal: 1) to sample inputs from the original images in pre-classified subdivisions with higher resolution and 2) to modify architecture of Convolutional Neural Networks (CNNs) based on U-Net and DenseNet. Various experiments are presented and evaluated in accordance with dice scores. We preliminarily find that imagery with higher resolution and larger input size of the networks can enhance classification performances. Besides, the overall performance turns out to be more sensitive to data characteristics than a choice of network models. Finally, we discuss still remaining issues when comparing our model with a benchmark and future work to consider both downsized and cropped images simultaneously.