Illegal Constructions Detection in Remote Sensing Images based on Multi-scale Semantic Segmentation

Urban planning is an important application field of remote sensing images. Using semantic segmentation to deal with this matter shows great potential. However, there is still a long way to go to achieve complex semantic segmentation. To improve the learning ability of complex rules in a semantic segmentation network, and can explicitly indicate the context relationship between categories. This paper proposes a new convolution structure based on the current semantic segmentation network with the encoding-decoding structure. The traditional multi-layer convolution structure is replaced by a new multi-scale convolution parallel structure. In addition, a full connection conditional random field under certain rules are added to constrain the segmentation results. For the segmentation accuracy, we first compare it with the current segmentation network on a open datasets. And it has shown good practicality in detecting illegal constructions in Jiangxi province, China.