Rapid Object Detection in VHR Optical Remote Sensing Images Based on Rotation-Invariant Discrete Hashing

Abstract

Object detection is one of the most fundamental but challenging problems faced for large-scale remote sensing image(RSI) analysis. Recently, learning based hashing techniques have attracted broad research interests because of their significant efficiency for high-dimensional data in both storage and speed. This paper proposes a novel object detection model which utilizes hashing methods to substantially improve the detection speed. In particular, firstly a selective search method is used to generate a number of high-quality object proposals that may contain objects. Then we propose rotation-invariant discrete hashing(RIDISH), which sloves the problem of object rotation variations in RSI, to quickly eliminate most non-object proposals in Hamming space. And finally the object detection task can be achieved by classifying the rest(very limited amount of) proposals with more discriminating classification model. Experimental evaluations on a publicly available very high resolution (VHR) remote sensing dataset point out that the presented object detection model is much faster, while keeping more superior performance than those typically used in VHR remote sensing images.