Convolutional Neural Network with Dilated Anchors for Object Detection in Very High Resolution Satellite Images

Abstract

Nowadays, object detection has acquired a great concentration either in ordinary images or satellite images. For satellite images, object detection is a challenging problem because objects have different scales and sparsity with very complicated background. Recent deep learning approaches have achieved breaking results for object detection than traditional ones. The ability of bounding boxes to catch existing objects with a complete and precise manner is still a challenging problem. We propose a dilated anchor method based on You Only Look Once version 3(YOLOv3) algorithm to make object detection more flexible and precise. The proposed method uses greater size anchor bounding boxes with about 30 % to 40 % larger than the traditional ones. This increase in anchor size increases the ability to catch more class objects with less influence on location detection. The experimental results using public NWPU VHR-10 dataset demonstrate the effectiveness of the proposed method in object detection of most classes and increase the overall accuracy with minimal effect on the precise location.