Hyperspectral images involve abundant spectral and spatial information, playing an irreplaceable role in land-cover classification . The architecture of most deep learning models is manually designed, relies on specialized knowledge, and is relatively tedious . In HSI classifications,datasets captured by different sensors have different physical properties . The overlap regions ofpatches of adjacent pixels are calculated repeatedly, which increasescomputational cost and time cost . We propose a new fast classification framework, i,e., pixel-to-pixel classificationframework, which has no repetitive operations and reduces the overall cost . The networks designed by our 3D-ANAS achieve competitiveperformance compared to several state-of-the-art methods, while having a muchfaster inference speed. Furthermore, we propose a 3D asymmetricdecomposition convolutions. The networks were designed by the 3D ANAS achieve faster inference speed than those used by other methods, including those designed by their 3DANAS in HSI networks designed in this way . The network was designed to achieve competitive performance compared to the networks designed to

Author(s) : Haokui Zhang, Chengrong Gong, Yunpeng Bai, Zongwen Bai, Ying Li

Links : PDF - Abstract

Code :
Coursera

Keywords : designed - d - networks - classification - achieve -

Leave a Reply

Your email address will not be published. Required fields are marked *