Object detection in VHR optical remote sensing images via learning rotation-invariant HOG feature

Object detection in very high resolution (VHR) optical remote sensing images is one of the most fundamental but challenging problems in the field of remote sensing image analysis. As object detection is usually carried out in feature space, effective feature representation is very important to construct a high-performance object detection system. During the last decades, a great deal of effort has been made to develop various feature representations for the detection of different types of objects. Among various features developed for visual object detection, the histogram of oriented gradients (HOG) feature is maybe one of the most popular features that has been successfully applied to computer vision community. However, although the HOG feature has achieved great success in nature scene images, it is problematic to directly use it for object detection in optical remote sensing images because it is difficult to effectively handle the problem of object rotation variations. To explore a possible solution to the problem, this paper proposes a novel method to learn rotation-invariant HOG (RIHOG) features for object detection in optical remote sensing images. This is achieved by learning a rotation-invariant transformation model via optimizing a new objective function, which constrains the training samples before and after rotation to share the similar features to achieve rotation-invariance. In the experiments, we evaluate the proposed method on a publicly available 10-class VHR geospatial object detection dataset and comprehensive comparisons with state-of-the-arts demonstrate the effectiveness the proposed method.