基于Faster R-CNN的高分二号遥感影像特定目标识别

Faster R-CNN based specific target recognition of GF2 remote sensing images

针对当前遥感影像特定目标识别效果较差的现状,基于深度学习网络对高分二号遥感影像特定目标进行准确识别。首先通过引入部分噪声增强样本数据建立样本库并在TensorFlow框架下配置Faster R-CNN网络学习目标特征建立可用于高分二号遥感影像特定目标识别的卷积网络。而后为判别深度学习网络的识别效果,选取遥感影像目标识别效果较好的约束能量最小化(constrained energy minimization,CEM)算法与之比较。最后在待识别遥感影像内生成房屋的包围框并标注识别房屋的置信度,得到总体房屋识别的置信度为95.61%以上。实验中CEM法房屋目标识别率为76.4%,而深度学习法可达到90.9%,深度学习法目标识别率比CEM法高14.5%。实验结果表明Faster R-CNN适用于高分二号遥感影像的特定目标识别,相较于CEM法识别率有明显提升。

In view of the current situation that the recognition effect of specific targets in remote sensing images is poor,based on deep learning network to accurately recognize specific targets in GF2 remote sensing images.Firstly,by introducing some noise-enhanced sample data to establish the sample database,and configuring Faster R-CNN network to learn target features under TensorFlow framework,a convolution network for GF2 specific target recognition is established.Then,in order to distinguish the recognition effect of deep learning network,the constrained energy minimization(CEM)algorithm with better recognition effect of remote sensing images is selected to compare with it.Finally,in the remote sensing image to be identified,the surround frame of the house is generated and the confidence of the identified house is marked.The confidence of the overall house recognition is more than 95.61%.In the experiment,the recognition rate of CEM method is 76.4%,while that of deep learning method is 90.9%.The recognition rate of deep learning method is 14.5%higher than that of CEM method.Conclusion Faster R-CNN is suitable for target recognition of GF2 remote sensing image.Compared with CEM,the recognition rate of Faster R-CNN is obviously improved.