An Effective Multi-class Object Detection Model for Remotely Sensed Image using Mask R- DCNN

Abstract

Object detection in remote sensing image has received increasing attention from the research community in recent days. Over the past few decades, variety of deep learning based detection model such as Region based Convolutional Neural Network (R-CNN), Fast R-CNN and Faster R-CNN has been applied for object detection.  However, most of the existing detection methods localize each object using the bounding box, but cannot segment the object from the background. So in order to tackle the issue, we introduce the Mask R- Dilated CNN model, which incorporates both object detection and segmentation. In Mask R-DCNN, ResNet-50 and ResNet-101 act as backbone for feature extraction, Region Proposal Network (RPN) is utilized to generate RoIs and RoIAlign is to carefully hold the exact spatial location to generate mask through Fully Convolution Network (FCN). The aim of Mask R- DCNN model is to incorporate more relevant information by increasing the receptive field of convolutional layer for improving the robustness. Experimental results on the NWPU VHR 10-class benchmark dataset demonstrated the effectiveness of the proposed model by providing 95.7% accuracy for Dilated ResNet-50 & 96.2% accuracy for Dilated ResNet-101, which is better than traditional Mask R-CNN model.