基于迁移学习的大规模遥感图像语义分割与提取

Semantic Segmentation and Extraction of Large Scale Remote Sensing Images Based on Transfer Learning

摘要: 图像语义分割是计算机视觉领域的热点研究课题，随着全卷积神经网络的迅速兴起，图像语义分割和全卷积神经网络的融合发展取得了快速发展。本文建立了基于特征融合的大规模卷积神经网络语义分割训练模型。通过迁移学习的监督式训练方式，对目前图像分割领域主流的模型进行了训练与比较，建立了评估指标PR参数以及预测图像的噪声分析模型。在训练模型时将其分成两个分支，利用主流的语义分割模型特性，分别做降噪分支和提取空间语义分支等，引入采集源高度作为权重参数，对不同分支进行特征融合，以提高其鲁棒性。最后进行特征融合，借助机器学习完成遥感图像分割任务并对本文模型的有效性进行验证。

Abstract: Image semantic segmentation is a hot research topic in the field of computer vision. With the rapid rise of fully convolutional neural networks, the fusion of image semantic segmentation and fully convolutional neural networks has achieved rapid development. In this paper, a large-scale convo-lutional neural network semantic segmentation training model based on feature fusion is estab-lished. Through the supervised training method of transfer learning, the current mainstream mod-els in the field of image segmentation are trained and compared, and the evaluation index PR pa-rameters and the noise analysis model of the predicted image are established. When training the model, it is divided into two branches, using the characteristics of mainstream semantic segmenta-tion models to do noise reduction branch and extraction spatial semantic branch, etc., and intro-ducing the height of the acquisition source as a weight parameter, and perform feature fusion on different branches to improve its robustness. Finally, feature fusion is performed, the remote sens-ing image segmentation task is completed with the help of machine learning and the effectiveness of the model in this paper is verified.