基于改进的Faster Rcnn遥感图像目标检测

OBJECT DETECTION BASED ON IMPROVED FASTER RCNN FOR REMOTE SENSING IMAGE

遥感图像因为其自身小目标多、密集的特点而对于目标检测任务是一个挑战。设计一种多层特征融合的Faster Rcnn,丰富各特征层的信息、平衡位置信息和分类信息。算法采用ResNet作为骨干网络提取特征,通过自上而下的特征融合,得到多尺度特征图,从而增强位置信息和分类信息以得到更加精准的检测结果。与Faster Rcnn算法相比,该算法对位置信息更加敏感,准确率提高了2.7百分点。相对于经典的目标检测框架SSD,Yolo v3等的检测效果,结合了特征融合的Faster Rcnn效果得到了明显提升。

Remote sensing image is a challenge for object detection task because of its many small objects and dense objects.A multi-level feature fusion based on the Faster Rcnn is designed,which enriches the information of each feature layer,the balance of position information and classification information.The algorithm used ResNet as the backbone network to extract features.Multi-scale feature maps were obtained with the top-down feature fusion to enhance location information and classification information so that the algorithm obtained more accurate detection results.Compared with the Faster Rcnn,it is more sensitive to location information.And the accuracy rate is improved by 2.7 percent point.Compared with the state-of-the-art object detection framework(such as SSD,Yolo v3,etc.),the Faster Rcnn combined with feature fusion has improved efficiency significantly.