Airplane Recognition from Remote Sensing Images with Deep Convolutional Neural Network

An automatic airplane recognition algorithm is proposed in this paper, which sequentially uses an object detection convolutional neural network (CNN) and a semantic segmentation CNN to accomplish the airplane recognition task. Experimental results on a collected airplane dataset demonstrate the effectiveness of the proposed method. Our results demonstrate that about 4 points increase of mIOU can be achieved by even reducing the shrinkage rate (SR) with a factor of 2, and better performance can be achieved by further reducing the SR in CNN. We also observed that post-processing techniques such as CRF could be unnecessary when the SR is relatively low. We also adopted the focal loss function, which was originally proposed for object detection, into the semantic segmentation task, and observed that better segmentation results can be achieved by assigning a larger weight to hard samples than to easy samples in the training procedure.