An Optimized Residual Network with Block-soft Clustering for Road Extraction from Remote Sensing Imagery

the task of road extraction from remote sensing imagery faces many challenges, traditional methods require complex extracting processes with relatively low precision. Deep learning methods such as convolutional neural network, VggNet, AlexNet, GoogleNet can obtain higher accuracy of road extraction, but requires lots of computing resources, training time and unsatisfactory real-time performance. Based on the reasons mentioned above, this paper proposes an optimized residual network with block-soft clustering (ORNBSC) for road extraction from remote sensing imagery. The block-soft clustering module aims at extracting essential features from satellite images and reducing the dimensionality, therefore accelerating the extraction speed. Meanwhile, the residual neural network module to improve the accuracy of extraction. Groups of experiments using Massachusetts roads dataset demonstrate that the ORNBSC model achieves better performance than traditional methods on precision of road extraction from remote sensing imagery.