Open AccessSemantic Segmentation of Remote Sensing Images via Stepwise-Refined Large-Kernel Deconvolutional Networks
Author(s) -
Xinwei Zhao,
Haichang Li,
Rui Wang,
Zheng Chen,
Song Shi
Publication year - 2019
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1267/1/012050
Subject(s) - computer science , segmentation , pascal (unit) , artificial intelligence , kernel (algebra) , focus (optics) , convolutional neural network , remote sensing , computer vision , feature (linguistics) , image segmentation , pattern recognition (psychology) , geography , linguistics , philosophy , physics , mathematics , combinatorics , optics , programming language
Deep CNN based semantic segmentation has been developed for several years and many models are proposed. However, most of them are designed for natural scene images such as PASCAL VOC, and cannot perform very well on remote sensing images, in which objects are much smaller and more densely distributed than those in natural scene images. In this paper, we demonstrate the importance of high-resolution feature maps and the problem of large dilated convolutional kernels in semantic segmentation of remote sensing images. Furthermore, we propose a Stepwise-Refined Large-Kernel Deconvolutional Network with a focus on small and densely-distributed objects such as houses and buildings, or long and narrow ones such as roads and rivers. Experiments on a public available ISPRS Vaihingen Challenge Dataset and our self-compiled Fujian Dataset show that our model outperforms the state-of-the-art models in semantic segmentation of remote sensing images.