GPU-accelerated registration of hyperspectral images using KAZE features

  • PDF / 1,006,121 Bytes
  • 15 Pages / 439.37 x 666.142 pts Page_size
  • 44 Downloads / 230 Views

DOWNLOAD

REPORT


GPU‑accelerated registration of hyperspectral images using KAZE features Álvaro Ordóñez1   · Francisco Argüello2   · Dora B. Heras1   · Begüm Demir3 

© Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract Image registration is a common task in remote sensing, consisting in aligning different images of the same scene. In the particular case of hyperspectral images, the exploitation not only of the spatial information contained in the image but also of the spectral information helps to improve the registration. An example of registration method exploiting all the information contained in the images is HSI–KAZE, which is based on feature detection and detects keypoints using nonlinear diffusion filtering. The algorithm is oriented toward extreme situations in which the images are very different in terms of scale, rotation and displacement. In this paper, an efficient implementation of the HSI–KAZE algorithm on GPU using CUDA is proposed. A detailed analysis of the implementation as well as a performance comparison to an OpenMP multicore implementation is also presented. The resulting algorithm is suitable for on-board processing of high-resolution images. Keywords  Image registration · Hyperspectral data · KAZE features · Remote sensing · CUDA · GPU

* Álvaro Ordóñez [email protected] Francisco Argüello [email protected] Dora B. Heras [email protected] Begüm Demir demir@tu‑berlin.de 1

Centro Singular de Investigación en Tecnoloxías Intelixentes (CiTIUS), Universidade de Santiago de Compostela, Santiago de Compostela, Spain

2

Departamento de Electrónica e Computación, Universidade de Santiago de Compostela, Santiago de Compostela, Spain

3

Faculty of Electrical Engineering and Computer Science, TU Berlin, Berlin, Germany



13

Vol.:(0123456789)



Á. Ordóñez et al.

1 Introduction Nowadays, the capture of hyperspectral images is easier thanks to the new advances in image sensor technologies. Hyperspectral images consist of hundreds of continuous spectral bands. This large amount of information allows extending the use of these images to a plethora of applications such as land use classification [4], quality control [12], agriculture [9] and medical science [22], among others. Image registration is a fundamental preprocessing task in many of these applications. It consists in estimating a geometrical transformation that maps one image to another. Registration methods can be classified according to their nature into two categories [25]: area-based methods and feature-based methods. The first group, area-based methods, work directly with image intensity values, while the second group, feature-based methods, seek to detect relevant features such as regions, lines or points. This representation at a higher level makes the featurebased methods more suitable for multisensor registration or illumination changes. The scale-invariant feature transform (SIFT) [10] is the most popular featurebased method. It tries to detect points with distinctive features called keypoints. Keypoints ar