Abstract
Runway segmentation for polarized synthetic aperture radar has been a problem of great interest
in the field of synthetic aperture radar imagery. Its goal is to segment the radar image in order to determine
the exact location of the runway and its unique morphology, which is crucial for both military
and civil purposes. The extraction of runways for synthetic aperture radar pictures still relies heavily on
conventional methods because of the issue of a little amount of data. However, the segmentation model
of the convolutional neural network, which is popular in the segmentation task of optical pictures, has
the advantages of high accuracy and versatility. Therefore, this paper proposes an airport runway segmentation
algorithm based on Dilated U-net network for synthetic aperture radar, which can accurately
extract airport runways even with a small amount of data using deep learning methods. This algorithm
combines dilated convolution with a U-net network to extract the runway region. The addition of dilated
convolution gives the original U-net network a larger perceptual field in the process of feature extraction,
which is necessary for airport runway segmentation with a connected structure. After comparison
experiments, the algorithm in this paper uses a deep learning method under the circumstance of a small
amount of data to increase the accuracy of detection results and also reduce false alarms and missed
alarms.