采用改进DeepLab网络的遥感图像分割

Remote Sensing Image Segmentation Based on Improved DeepLab Network

中文关键词:  遥感图像处理  深度学习  卷积神经网络  语义分割

英文关键词:remote sensing image processing  deep learning  convolutional neural network  segmantic segmentation

中文摘要:

      基于深度学习的语义分割方法具有较好的应用前景，但针对海量的遥感数据，深度学习算法在训练速度上仍有较大提升空间。在前人研究基础上，提出一种改进的全卷积网络，改进的网络以DeepLab为网络前端，结合Inception结构，在不降低特征提取能力的前提下，通过减少网络参数数量，降低网络运算复杂度，有效提升了网络的训练速度。基于INRIA Aerial Image Dataset高分辨率航空遥感数据集开展语义分割对比实验，实验结果显示改进的网络在训练速度和精度上均较DeepLab网络有所提升。所提出的改进DeepLab网络对为处理数据量庞大的遥感影像提供了可参考的解决方案。

英文摘要:

      The semantic segmentation based on deep learning has a desirable application prospect.However,there is great space for it to improve the training speed for the massive remote sensing data.Based on the previous studies,an improved fully convolutional network was proposed in this contribution.Combined with the Inception structure,the new network uses DeepLab as the front end,the training speed for massive remote sensing data is improved by decreasing the parameters amount and computational complexity of the network without reducing the feature extraction capability.A semantic segmentation contrast experiments were performed by using the INRIA high resolution aerial remote sensing datasets.The results show that the improved DeepLab network has better training speed and accuracy than the original DeepLab.The new network proposed provides a reference solutions for the semantic segmentation application of remote sensing images with massive data.