Object detection from satellite imagery

Abstract

This report is about explaining how to apply the Faster R-CNN network structure on Object detection from satellite imagery. It explains different parts, including preparation, implementation, experiment results, and conclusion, and the purpose is trying to find out the best model for object detection. Comparing to the last “generation” CNN network, Fast R-CNN, RPN is the radically different part that implied in Faster R-CNN. It gives up the traditional selective search method but uses generated small “window”(anchor) to find the proposal region. There are lots of features that may affect the network's training and performance, like chosen convolutional neural network, learning rate, size of the dataset, and the testing dataset. The experiment and discussion part examines and discusses all the mentioned factors above in the report, and the discussion depends on the experiment results.