NEW NETWORK BASED ON D-LINKNET AND RESNEXT FOR HIGH RESOLUTION SATELLITE IMAGERY ROAD EXTRACTION

DlinkNet[1] (LinkNet With Pretrained Encoder and Dilated Convolution) has been proved to be an effective method for road extraction in remote sensing fields as it won the champion in the DeepGlobe's Road Extraction Challenge. However, as the number of hyperparameters increases (such as the number of channels, filter size, etc.), the difficulty and computational overhead of network design will increase. Focused on this problem. this paper put forward effective ideas to improve D-LinkNet: (1) Applying ResNeXt as its backbone instead of ResNet to rebuild D-LinkNet: (2) Replacing initial block with stem block in the beginning of the network. The results of mad extraction which was trained with our own dataset was evaluated with IoU scores. The evaluation results shows that the improved network has higher IoU scores than D-LinkNet when maintaining the model complexity and number of parameters.