Estás aquí: start » hiperespectral » eadp_gpu

Extended Anisotropic Diffusion Profiles in GPU for Hyperspectral Imagery

Esta é unha revisión antiga do documento!


Extended Anisotropic Diffusion Profiles in GPU for Hyperspectral Imagery

Experimental results related to the paper Extended Anisotropic Diffusion Profiles in GPU for Hyperspectral Imagery by Álvaro Acción, Francisco Argüello, and Dora B. Heras.

Abstract

Morphological profiles are common approach for extracting spatial information from hyperspectral images by extracting structural features. Additional kinds of profiles can be built based on different approaches as, for example, differential morphological profiles, or attribute profiles. Another technique used for characterizing spatial information on the images at different scales is based on computing edge-preserving filters such as anisotropic diffusion filters. Their main advantage is to preserve the distinctive morphological features of the images at a the cost of an iterative calculation. In this paper, the high computational cost associated to the construction of Anisotropic Diffusion Profiles (ADPs) is drastically reduced with the usage of GPUs. In particular, we propose a low cost computational approach for computing ADPs in Nvidia GPUs as well as a detailed characterization of the method, comparing it in terms of accuracy and structural similarity to other existing alternatives.

Downloads

Scenes

The following link contains some of the public hyperspectral images used in the experiments: http://www.ehu.eus/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes

Indian Pines EADP experimental results

The following repository contains a subset of the data used for the experimental results detailed in the paper. The EADP generated with the proposed algorithm applied to the Indian Pines scene, as well as the samples and results for the classification experiments are available for downloading. EADP Dataset. The parameters used for the algorithm were N=7, C=8, sigma=1 and ts=65. The total ADP size is 9 and the total EADP size is 63.

The image below these lines corresponds to the EADP generated with the parameters detailed in the previous paragraph. Each row corresponds to one of the principal components extracted from the original hyperspectral image. The first column contains the principal component, and each subsequent column shows the filtered images produced by one iteration of the algorithm for a given process time. The process time is the same for all the images represented in the same column.

The average results for 100 classification experiments using the EADP are as follows:

Metric Value
OA 92.45%
AA 95.27%
Kappa 91.34%
Class Accuracy
C1 0.981
C2 0.928
C3 0.959
C4 0.987
C5 0.967
C6 0.989
C7 0.961
C8 1
C9 1
C10 0.753
C11 0.905
C12 0.830
C13 0.994
C14 0.998
C15 0.991
C16 0.999