Résultat de la recherche
Robust input layer for neural networks for hyperspectral classification of data with missing bands
Hyperspectral classification using artificial neural networks is commonly applied on camera dependent interpolated data, or on the results of a dimensionality reduction algorithm. While these methods usually produce satisfactory results, they have severe limitations when part of the spectrum is missing, for example when parts of the image are overexposed or affected by bad pixels. This article presents an input layer based on the Haar transform for artificial neural networks used for hyperspectral data classification. This input layer is designed to perform efficiently with incomplete data and is independent of the specific bands used by the camera. This could enable providing pre-trained neural networks, which can be used with a camera with different specifications than the one used for training. This paper shows that a classifier for mineral identification built using this approach performs better than standard normalization on incomplete spectra, and similarly on complete spectra. Additionally, it shows that such a classifier matches local spectral features, and therefore that the artificial neural network is matching the spectrum shape.
A 2D hyperspectral library of mineral reflectance, from 900 to 2500 nm
Mineral identification using machine learning requires a significant amount of training data. We built a library of 2D hyperspectral images of minerals. The library contains reflectance images of 130 samples, of 76 distinct minerals, with more than 3.9 million data points. In order to produce this dataset, various well-characterized mineral samples were scanned, using a SPECIM Short Wave Infrared (SWIR) camera, which captures wavelengths from 900 to 2500 nm. Minerals were selected to represent all the mineral classes and the most common mineral occurrences. For each sample, the following data are provided: (a) At least one hyperspectral image of the sample, consisting of 256 wavelengths between 900 and 2500 nm. The raw data, the high dynamic range (HDR) image, and the masked HDR image are provided for each scan (each of them in HDF5 format). (b) A text file describing the sample, providing supplementary information for the subsequent analysis (c) RGB images (JPEG files) and automated 3D reconstructions (Stanford Triangle PLY files). These data help the user to visualize and understand specific sample characteristics. 2D hyperspectral images were produced for each mineral, which consist of many different spectra with high diversity. The scans feature similar spectra than the ones in other available spectral libraries. An artificial neural network was trained to demonstrate the high quality of the dataset. This spectral library is mainly aimed at training machine learning algorithms, such as neural networks, but can be also used as validation data for other types of classification algorithms.
Short-wave infrared hyperspectral imaging of minerals
L'imagerie hyperspectrale est une technique prometteuse pour l'identification et la cartographie de minéraux. Elle est largement employée en télédétection, au moyen de capteurs aéroportés, mais est relativement peu utilisée en laboratoire. De fait, un ensemble de problèmes limite la qualité des données obtenues, et par conséquent celle des conclusions en découlant. L'identification de minéraux nécessite deux étapes: 1. acquisition des données et calibration, 2. classification, en utilisant une librairie spectrale en tant que données de références. Cette thèse présente une analyse de l'entier de ce processus et propose des améliorations.
La caméra a été soigneusement étudiée afin de comprendre son principe de fonctionnement. Les différents aspects de l'acquisition ont été soigneusement évalués, notamment le choix du type de source de lumière, la disposition des différents éléments lors de l'acquisition de données, ainsi que l'impact des éléments environnants. Des algorithmes ont été conçus afin d'améliorer l'efficacité de l'acquisition de données, notamment pour automatiser le focus ou pour éviter les difficultés à choisir l'exposition correcte. Des méthodes combinant plusieurs images afin d'améliorer la qualité ont aussi été développées: de l'imagerie à grande gamme dynamique (HDR) a été implémentée afin de combiner plusieurs images capturées avec des temps d'intégration différents, et une méthode a été développée pour séparer la transmittance et la réflectance d'objets translucides.
L'efficacité des réseaux de neurones artificiels a été évaluée pour la classification: ceux-ci produisent des résultats très précis, mais requièrent un spectre complet. Cette exigence nécessite l'interpolation de pixels défectueux et empêche la réutilisation d'un réseau déjà entraîné avec une autre caméra. Ainsi, une nouvelle couche d'entrée (input layer) a été conçue, laquelle gère de façon efficace des spectres dont certaines bandes manquent. Comme il n'existait pas de données d'entraînement à l'identification de minéraux d'un tel réseau de neurones, une librairie spectrale d'images de réflectance a été créée. Celle-ci contient 130 échantillons de 76 minéraux distincts. Une telle librairie est, à notre connaissance, unique.
Finalement, des outils logiciels et matériels ont été développés afin d'effectuer toutes ces opérations de façon efficace, et d'automatiser tous les équipements concernés. Ces outils ont été publiés sous licence Open Source.
Abstract
Hyperspectral imaging is a promising technology for mineral identification and mapping. This technology is commonly used for remote sensing, using aerial sensors. It is however less frequently used for laboratory measurement, as a range of issues currently undermine the quality of the data obtained, and consequently the resulting conclusions. Mineral identification consists in two steps: 1. data acquisition and calibration, 2. classification, using a spectral library as reference data. This thesis presents an analysis of this whole process and proposes improvements.
The camera was carefully studied in order to understand exactly its operating principles. Each aspect of the scene set-up, such as the type of light source, the positioning of the various elements of the set-up and the impact of surrounding objects was carefully evaluated. Algorithms were designed to improve the efficiency of the acquisition, such as automate focus or to avoid difficulties in choosing the correct exposition. Methods to combine multiple images in order to increase the data quality were also developed: high dynamic range was implemented to merge multiple images captured with different integration time, and a method was developed to separate the transmittance and the reflectance of translucent objects.
The efficiency of artificial neural networks was evaluated for classification. It was found that they were producing highly accurate results, but required complete spectra as input, which requires interpolation of bad pixels and prevents the re-use of an already trained network with a different camera. Therefore a novel input layer was designed which handles efficiently spectra with missing bands. As no training data was readily available for training an artificial neural network for mineral identification, a spectral library of 2D reflectance images of 130 samples of 76 distinct minerals was created. To the best of our knowledge, no similar dataset exists.
Finally, hardware and software tools were designed and implemented to carry out all these steps and to allow efficient automation of all the devices involved. These tools have been published under Open Source licenses.