Backbones-Review: Satellite Object Detection Using Faster-RCNN

Abstract

Object detection in remote sensing (RS) images has recently played an important role in many applications, such as environmental monitoring, ship detection, fire detection, autonomous driving, robotic vision, and crowd counting. In addition, object detection in nature images faces challenges such as enhancements in speed and accuracy in training and testing. On the other hand, the challenges in aerial or satellite images include viewpoint variation, occlusion, background cloudiness, shadows, and noise reduction. Deep learning models have helped to solve many challenges in object detection problems, such as the enhancement of speed and accuracy, enhancement of the aerial or satellite image (noise reduction), generation of new examples for the database from previous examples to be a large-scale database, etc. The Faster-RCNN method is one of object detection's most important deep learning models. With the advent of convolutional neural networks (CNNs), feature extraction has become more automated and simpler. This paper investigates four backbones in the remote sensing domain: resnet-50, resnext50_324d, efficientnet_b0, and densenet121. Various experiments were conducted on the NWPU VHR-10 dataset to evaluate these backbones using precision, recall, f1-score, average precision (AP), and mean AP matrices. The obtained results indicate that resnext50_324d has suppressed other backbones by 0.847 in terms of mAP.