HsgNet: A Road Extraction Network Based on Global Perception of High-Order Spatial Information

Abstract

Road extraction is a unique and difficult problem in the field of semantic segmentation because roads have attributes such as slenderness, long span, complexity, and topological connectivity, etc. Therefore, we propose a novel road extraction network, abbreviated HsgNet, based on high-order spatial information global perception network using bilinear pooling. HsgNet, taking the efficient LinkNet as its basic architecture, embeds a Middle Block between the Encoder and Decoder. The Middle Block learns to preserve global-context semantic information, long-distance spatial information and relationships, and different feature channels’ information and dependencies. It is different from other road segmentation methods which lose spatial information, such as those using dilated convolution and multiscale feature fusion to record local-context semantic information. The Middle Block consists of three important steps: (1) forming a feature resource pool to gather high-order global spatial information; (2) selecting a feature weight distribution, enabling each pixel position to obtain complementary features according to its own needs; and (3) inversely mapping the intermediate output feature encoding to the size of the input image by expanding the number of channels of the intermediate output feature. We compared multiple road extraction methods on two open datasets, SpaceNet and DeepGlobe. The results show that compared to the efficient road extraction model D-LinkNet, our model has fewer parameters and better performance: we achieved higher mean intersection over union (71.1%), and the model parameters were reduced in number by about 1/4.

Road Extraction from Very-High-Resolution Remote Sensing Images via a Nested SE-Deeplab Model

“…Alshehhi et al [42] designed an improved single patch-based CNN structure to achieve the extraction of buildings and roads; in the post-processing stage, the extracted features by CNN can be combined with the low-level features of roads and buildings such as asymmetry and compactness of adjacent regions, which improved the accuracy and integrity of extraction tasks. Xie et al [43] combined the efficient LinkNet with a Middle Block to develop a HsgNet model, which made use of global semantic information, long-distance spatial information and relationships, and information of different channels to improve the performances in roads extraction with fewer parameters compared with D-LinkNet. An important aspect of research on road detection is center line extraction, which is not limited to semantic segmentation.…”

FuNet: A Novel Road Extraction Network with Fusion of Location Data and Remote Sensing Imagery

“…Of course, there is a shortage of local information loss due to the use of dilated convolution. At present, the emerging attention mechanism [12][13][14] for global information learning has also achieved success in the field of semantic segmentation, such as Non-local [15], PSANet [16], A2Net [17], EMANet [18], and HsgNet [19]. Graph convolution networks (GCN) [20] are also brought into focus because of its strong reasoning learning ability.…”

Road Extraction from High Resolution Remote Sensing Images Based on Vector Field Learning

“…An important modification of FCN by U-net is that there are a large number of feature channels in the up-sampling part, which allows the network to propagate context information to higher resolution layers. The research ideas of road extraction using U-net include multivariate loss function [ 7 , 8 ], modification, of network architecture such as new network unit or adding jump connection [ 9 , 10 , 11 , 12 , 13 ], pre-training [ 14 ], multitask learning strategy [ 15 , 16 ], etc. In addition to U-net, some new encoder-decoder networks are also proposed for road extraction [ 17 , 18 , 19 , 20 ].…”