| SINGLE IMAGE DEHAZING ON AERIAL IMAGERY USING CONVOLUTIONAL NEURAL NETWORKS |
| Paper ID : 1406-SMPR |
| Authors: |
|
Mojgan Madadikhaljan *1, Seyed Majid Azimi2, Peter Reinartz2, Uwe Soergel3 1Deutschen Zentrums für Luft- und Raumfahrt (DLR) 2Remote Sensing Technology Institute (IMF), German Aerospace Center (DLR) 3University of Stuttgart |
| Abstract: |
| As a physical phenomenon, the particles floating in the air called as haze can be seen in many areas and cause significant visibility attenuation in aerial imagery. Obstruction of the objects due to the haze corrupts the performance of object detection and image segmentation algorithms. Current single-image dehazing algorithms using convolutional neural networks use ground-imagery to obtain a haze free image form a single hazy image. Therefore, all of the previous single-image dehazing neural networks use datasets such as NYU containing only indoor and outdoor images. Unlike previous works, in this paper, we focus on single image dehazing task using aerial imagery. We implement the-state-of-the-art CNNs-based single-image dehazing methods using airborne images and compare their performance. In order to prepare training dataset, we generate synthetically-hazed aerial images to train convolutional neural networks. We compare the outputs of the networks trained with the synthetically-generated hazed aerial images with the ones using in-situ images. Since the training data have been changed to be aerial instead of indoor/outdoor images, we expect significant improvement on the dehazing task. |
| Keywords: |
| Single-image Dehazing, Convolutional Neural Neworks, Aerial Imagery, Haze, Hazy Image Generation |
| Status : Conditional Accept (Oral Presentation) |
