Road Extraction in Remote Sensing Imagery Based on PVT-Unet.

What is it about?

Road extraction from remote sensing images has gradually become a prominent research hotspot in the field of autonomous driving and smart city construction. In recent years, with the developments of computing power, deep learning has been widely used in this field and convolution neural networks are usually used to extract roads. However, since the roads in the remote sensing images are easy to be occluded by trees and buildings, the roads extracted by these methods are usually fragmented. In this paper, a U-shaped Neural Network based on Pyramid Vision Transformer (PVT-Unet) is designed. This network combines Transformer's long term learning capability with U-shaped network multi-scale feature extraction capability to predict the roads well. Experimental results show that PVT-Unet outperforms the state-of-the-art methods in all evaluation metrics on the Istanbul City Road Dataset.ork

Featured Image

Photo by NASA on Unsplash

Photo by NASA on Unsplash

Why is it important?

• We propose PVT-Unet, a novel approach which extract roads from remote sensing images more accurately by capturing multi-scale feature information. • In the shallow layers, to improve the efficiency of the model, we use fewer Transformer blocks to learn feature information. In the semantic layers, as the number of tokens continues to decrease, we use more Transformer blocks to learn semantic information. And the calculation amount of the model will not increase significantly due to the increase of Transformer blocks. • We evaluate on Istanbul City Road Dataset that is publicly available and the performance outperforms the state-of the art networks.

Perspectives