A Remote Sensing Image Segmentation Model Based on CGAN Combining Multi-scale Contextual Information

摘要

In order to extract targets in different scale under complex scene, this paper proposes a remote sensing image segmentation model based on Conditional Generative Adversarial Network (CGAN) combing multi-scale contextual information. This end-to-end model consists of a generative network and a discriminant network. A SegNet model fusing multi-scale contextual information is proposed as the generative network. In order to extract multi-scale contextual information, the multi-scale features of the end pooling feature map in the encoder are extracted using different proportion of dilated convolution. The multi-scale features are further fused with the global feature. The discriminant network is a convolution neural network for two category classification, determines whether the input is a generated result or the ground truth. After alternate adversarial training, the experimental results on a remote sensing road dataset show that the road segmentation results of the proposed model are superior to those of the comparable models in terms of target integrity and details preserving.