基于Faster R-CNN的卫星图像污水处理厂识别

Wastewater treatment plant recognition in satellite images using Faster R-CNN

针对卫星图像中污水处理厂目标识别性能低的问题,提出了更快速区域卷积神经网络(Faster R-CNN)和工艺环节相结合的方法,检测污水处理厂、生化池和污泥泵房目标。在污水处理厂识别过程中,使用数据扩充技术、引入负样本等方法来扩充训练集样本;选用ZFNet、VGG和ResNet三种神经网络进行特征提取,采用Faster R-CNN方法训练目标检测模型,同时根据图像检出的工艺环节与污水处理厂之间的从属关系,过滤掉孤立的污水处理厂目标和工艺环节目标,提升污水处理厂的目标识别性能。实验结果表明,结合ResNet、Faster R-CNN和工艺环节方法的识别效果最好,相较于ResNet结合Faster R-CNN方法:准确率可以达到79.68%,提升了5.92%;召回率可以达到93.45%,提升了3.32%;F-measure可以达到86.2%,提升了4.84%。实验结果表明,该方法对不同结构、不同工艺环节的污水处理厂都有不错的识别效果,能夠兼顾识别精确率和召回率。

In order to deal with low target recognition performance of wastewater treatment plant in satellite images,a Faster R-CNN(Faster Regions with Convolutional Neural Network features)method combining with technology processes was proposed to detect the target of wastewater treatment plant,biochemical pool and sludge pump house.In the recognition process of wastewater treatment plant,data expansion technique and increasing negative samples method were used to expand training set samples firstly.Secondly,ZFNet,VGG and ResNet were adopted to extract the features of input images.Lastly,Faster R-CNN method was used to train the target detection model,and at the same time,according to the subordinate relationship between the technology processes detected by the image and the sewage treatment plant,the targets of the isolated sewage treatment plants and the targets of the technology processes were filtered out to improve the target recognition accuracy of the wastewater plants.The experimental results show that the performance of combination of ResNet,Faster R-CNN and technology processes is the best,by comparing with the combination method of ResNet and Faster R-CNN,the preision,recall and F-measure can reach 79.68%,93.45%and 86.2%respectively,and improved by 5.92%,3.32%,4.84%respectively.It is verified that the proposed method can achieve a good recognition performance for wastewater treatment plants with different structures and processes,and take into account the precision and the recall.