Recent years have observed an immense refinement in the domain of object localization, especially over remote sensing images. Object localization aims to predict the objects within an image as well as its boundaries. Achieving accurate object localization over remote sensing images is always challenging due to the complex context information. The paper mainly concentrates on solving the problem of accurate object localization using a three-stage localization pipeline. The introductory stage in the pipeline focus on generating regions of interest (RoI) called candidate regions using selective search algorithm. Then, each of these candidate features is extracted by passing it through a hybrid convolutional neural network (CNN) model. Finally, to improve the localization accuracy we propose an optimal object localization technique called Unsupervised Score Based Bounding Box Regression using Feature Descriptors (USBBBR-FD) algorithm integrated with non-maximum suppression (NMS) algorithm to optimize the bounding boxes of regions which are detected as objects. Analysis shows that the detection precision of the hybrid CNN model is higher than that of any other single CNN model and the dimension-reduction CNN model performs far better than retrained CNN models. Experiments show that the proposed USBBBR-FD algorithm can more accurately locate objects within remote sensing images and also shows robustness in complex background as compared with traditional features extraction methods.

This content is only available via PDF.
You do not currently have access to this content.