ACENTRIST: An Augmented Census Transform Histogram for Aerial Image Classification

Abstract

Aerial image classification has become one of the most important topics to the computer vision researchers because of its numerous real world application. A great number of census transform based descriptors have been introduced in recent years to classify the aerial images. But the major drawback of these census transform based techniques is, most of these techniques works only with the center pixel information of an image with respect to their neighboring pixels. Hence, no information about the relationship among the neighboring pixels is obtained. To mitigate this problem, we introduce an Augmented Census Transform Histogram (ACENTRIST) for aerial image classification which encodes both the center pixel information and neighboring pixel information. The proposed technique augments two local binary pattern based descriptor which encodes the center pixel information with respect to the neighboring pixels and information of the angular difference of the neighboring pixels. We have conducted thorough experiments in two of the well-known aerial image dataset, UC Merced Land Use (Land Use 21) and In-House (Banja Luka), and the experimental result shows that the proposed methodology gains considerable higher accuracy over the state of the art methods.