An Attention Augmented Convolution based Improved Residual UNet for Road
Extraction
Abstract
Recently remote sensing images have become more popular due to improved
image quality and resolution. These images have been shown to be a
valuable data source for road extraction applications like intelligent
transportation systems, road maintenance, and road map making. In recent
decades, the use of highly significant deep learning in automatic road
extraction from these images has been a hot research area. However,
fully automated and highly accurate road extractions from remote sensing
images remain a challenge due to topology differences, complicated image
backgrounds, and complex contexts. This paper proposes novel attention
augmented convolution-based residual UNet architecture (AA-ResUNet) for
road extraction, which adopts powerful features of self-attention
mechanism and advantageous properties of residual UNet structure. The
self-attention mechanism uses attention augmented convolutional
operation to capture long-range global information; however, traditional
convolution has a fundamental disadvantage: it only performs on local
information. Therefore, we use the attention augmented convolutional
layer as an alternative to standard convolution layers to obtain more
discriminant feature representations. It allows to develop a network
with fewer parameters. We also adopt improved residual units in standard
ResUNet to the speedup training process and enhance the segmentation
accuracy of the network. Experimental results on Massachusetts, DeepGlob
Challenge, and UAV Road Dataset show that the AA-ResUNet performs well
in road extraction, with Intersection over Union (IoU) (94.27%), lower
trainable parameters (1.20 M), and inference time (1.14 sec).
Comparative results on the proposed method have proven the supremacy or
compatibility in road extraction with ten recently established deep
learning approaches.