Airplane Object Detection in Satellite Images Based on Attention Mechanism and Multi-scale Feature Fusion

The order of magnitude of object detection instances in satellite images is larger than that in conventional images, and many small object instances are clustered together in satellite images. Most objects in conventional object detection datasets are perpendicular to the ground, while objects in satellite images are parallel to the ground and their orientation varies greatly. Conventional detection with horizontal bounding box, a horizontal bounding box may contain multiple dense instances and may contain a lot of background information, resulting in a disproportionate proportion of the background in the detection box. The oriented bounding box helps to locate the object more accurately and obtain the azimuth information of the object more easily. Therefore, in this work, we used oriented bounding boxes to detect airplane objects in satellite images. In order to target the detection of airplane objects, we produced a high-resolution dataset containing only airplane objects and more small instances. We use ResNet as backbone and FPN as Neck. By adding an SE Block module after each stage of ResNet to pay attention to important features, and connecting a pyramid convolution (PConv) module after FPN to enhance feature fusion. Cropping the input image to 1024*1024 using a sliding window allowed detection accuracy improved on DOTA v1.0(90.53 mAP) and our own dataset Satellite-Airplane2309(83.48 mAP) on the GTX TITAN X.