EYNet: Extended YOLO for Airport Detection in Remote Sensing Images
Nowadays, airport detection in remote sensing images has attracted considerable attention due to its strategic role in civilian and military scopes. In particular, uncrewed and operated aerial vehicles must immediately detect safe areas to land in emergencies. The previous schemes suffered from various aspects, including complicated backgrounds, scales, and shapes of the airport. Meanwhile, the rapid action and accuracy of the method are confronted with significant concerns. Hence, this study proposes an effective scheme by extending YOLOV3 and ShearLet transform. In this way, MobileNet and ResNet18, with fewer layers and parameters retrained on a similar dataset, are parallelly trained as base networks. According to airport geometrical characteristics, the ShearLet filters with different scales and directions are considered in the first convolution layers of ResNet18 as a visual attention mechanism. Besides, the major extended in YOLOV3 concerns the detection Sub-Networks with novel structures which boost object expression ability and training efficiency. In addition, novel augmentation and negative mining strategies are presented to significantly increase the localization phase's performance. The experimental results on the DIOR dataset reveal that the framework reliably detects different types of airports in a varied area and acquires robust results in complex scenes compared to traditional YOLOV3 and state-of-the-art schemes.