基于Mask-RCNN的建筑物目标检测算法

Building target detection algorithm based on Mask-RCNN

针对在航空影像中,城区80%的人工目标物为建筑物和道路,建筑物是遥感影像中主要地物的类别,所以建筑物的检测会直接影响到地物提取的自动化水平这一问题。该文提出了一种基于Mask-RCNN的建筑物目标检测方法,是基于卷积神经网络思想,在深度学习框架下通过多线程迭代训练,将无人机影像作为训练样本,在卷积神经网络中得到目标特征再通过区域建议网络(RPN)与ROIAlign操作将特征输入不同的全连接分支。最后得到具优化的权重参数的目标检测模型。在不同场景图像中,该模型可以检测出建筑物目标。实验结果达到了预期要求,提高了航空影像中建筑物检测的准确性。

Aiming at the problem that in aerial imagery,80%of the artificial targets in urban areas are buildings and roads,and buildings are the main types of features in remote sensing images,so the detection of buildings will directly affect the automation level of feature extraction.This paper proposed a building object detection method based on Mask-RCNN?It was based on the idea of convolutional neural network.Through the multi-thread iterative training in the deep learning framework,the unmanned aerial vehicle(UAV)image was used as the training sample.The target features were obtained in the neural network and then the features were input into different fully connected branches through the region proposal network(RPN)network and the ROIAlign operation.Finally,the target detection model with optimized weight parameters was obtained.In different scene images,the model could detect building targets.The experimental results met the expected requirements and improved the accuracy of building detection in aerial imagery.