Abstract:

High-resolution remote sensing image segmentation based on the traditional deep convolutional neural network needs hand-crafted architectures,which is excessively dependent on expert experience,time-consuming and laborious,and the network generalization ability is poor.A neural network architecture search method of resource balanced partial channel sampling is presented in this study.First,the resource-balanced strategy is added to the network architecture parameter to minimize the updating imbalances and discretization discrepancy during pruning,so the stability of the search algorithm is improved.Second,the partial channel is sampled for the mixed operation in search space,which can effectively reduce the computing cost,improve the search efficiency and alleviate the problem of network overfitting.Finally,according to the characteristic of complex features,the discrete distribution and the wide spatial range of high resolution remote sensing images,the Gumbel-Softmax trick is introduced to improve the sampling efficiency and make the sampling process backpropagate.The proposed method can achieve 90.93% and 69.53% MIoU on the WHUBuilding and GID dataset,respectively,which outperforms the prior work like SegNet,U-Net,Deeplab v3+ and NAS-HRIS.Experimental results show that this proposed method can help search the architecture for high resolution remote sensing image segmentation efficiently and automatically,and has the advantages of a high segmentation accuracy and low computing resources.

Key words: deep learning, high-resolution remote sensing, network architecture search, image segmentation, convolutional neural network