Rapid Aircraft Classification in Satellite Imagery using Fully Convolutional Residual Network
Shah Nawaz Khan, Syed Irteza Ali Khan, Zain UI Abideen, Muhammad Salman Khan, Shahzad Anwar · 2020
Advancement in high-performance computing technology has paved way for development of Deep Learning algorithms for computer vision to provide unprecedented performance both in terms of accuracy and speed. Image recognition, a subfield of computer vision, is one of the key application areas in which deep learning based Convolutional Neural Networks (CNN have achieved ground-breaking performance). Majority of the algorithms of object classification in CNN are focused on street view imagery that is of high resolution and have small size. The problem with satellite imagery is that it has objects in small size and images are especially in large size. There are two main objectives of this research: time reduction of processing large dimensions satellite images, and achieving acceptable accuracy of classifying small aircrafts. For this purpose, ResNet-50 has been modified such that it can process high-resolution satellite imagery of large dimension in one go instead of processing it in small patches sequentially, without affecting the accuracy of object classification. ResNet-50 with sliding-window scanning technique and the proposed model trained on satellite imagery are compared. The proposed method reduces the processing time by 99.9% by keeping the accuracy at the same level.