Hyperspectral image super-resolution using recursive densely convolutional neural network with spatial constraint strate
- PDF / 1,155,700 Bytes
- 11 Pages / 595.276 x 790.866 pts Page_size
- 83 Downloads / 220 Views
(0123456789().,-volV)(0123456789(). ,- volV)
EXTREME LEARNING MACHINE AND DEEP LEARNING NETWORKS
Hyperspectral image super-resolution using recursive densely convolutional neural network with spatial constraint strategy Jianwei Zhao1,2
•
Taoye Huang1 • Zhenghua Zhou1
Received: 29 January 2019 / Accepted: 29 August 2019 Ó Springer-Verlag London Ltd., part of Springer Nature 2019
Abstract Hyperspectral images (HSIs) have been widely applied in real life, such as remote sensing, geological exploration, and so on. Many deep networks have been proposed to raise the resolution of HSIs for their better applications. But training their huge number of model parameters (weights and biases) needs more memory for storage and computation, which may bring some difficulties when they are applied in mobile terminal devices. In order to condense the deep networks and still keep the reconstruction effect, this paper proposes a compact deep network for HSI super-resolution (SR) by fusing the idea of recursion, dense connection, and spatial constraint (SCT) strategy. We name this method as recursive densely convolutional neural network with a spatial constraint strategy (SCT-RDCNN). The proposed method uses a novel designed recursive densely convolutional neural network (RDCNN) to learn the mapping relation between the low-resolution (LR) HSI and the high-resolution (HR) HSI and then adopts the SCT to improve the determined HR HSI. Compared with some existing deep-network-based HSI SR methods, the proposed method can use much less parameters (weight and bias) to attain or exceed the performance of methods with similar convolution layers because of the recursive structure and dense connection. It is significant and meaningful for the practical applications of the network in HSI SR due to the limitations of hardware devices. Some experiments on three HSI databases illustrate that our proposed SCT-RDCNN method outperforms several state-of-the-art HSI SR methods. Keywords Hyperspectral image Super-resolution Deep convolutional network Recursion Dense connection
1 Introduction A hyperspectral image (HSI) is a series of images that record the spectrum of scene radiance by a distribution of intensity in a contiguous band over a certain electromagnetic spectrum. As HSI contains the spectral signatures of different objects, it has played a vital role in numerous applications, such as geological exploration [1], face recognition [2], object segmentation [3], and so on.
& Jianwei Zhao [email protected] 1
College of Sciences, China Jiliang University, Hangzhou 310018, Zhejiang Province, People’s Republic of China
2
State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210093, People’s Republic of China
Although HSI can achieve high spectral resolution, it has severe limitations in spatial resolution when compared against regular RGB (a.k.a. multispectral) cameras in visible spectrum. The reason is that hyperspectral imaging systems need a large number of exposures to acquire many bands simultaneously
Data Loading...
