Vehicle-focused super-resolution of remote sensing imagery

Abstract

Vehicle detection in remote sensing imagery has various applications in traffic analysis,planning, and rescue operations after natural disasters. Using super-resolution as apre-processing step to increase the spatial resolution of remote sensing imagery benefitsvehicle detection performance. This thesis proposes a novel procedure to train the superresolutionstep of this pipeline in a vehicle-focused manner, by cropping the trainingset to images centered around vehicles. The Residual Dense Network is selected assuper-resolution architecture and Faster R-CNN is utilized for vehicle detection. Six existing annotated datasets are combined and unified to create the vehicle-focused crops, a conventional dataset for super-resolution training, and a dataset for vehicledetection training. Additionally, testing on a seventh, completely unseen dataset allows a generalization error to be estimated. The effect of this super-resolution training methodon subsequent vehicle detection is quantified by training an identical super-resolution model on unfocused data for comparison. Extensive evaluation shows on par performance of the vehicle-focused approach, while allowing faster training.