Hyperspectral preprocessing is typically used to reduce disorder of the measurement setup (varying lighting, etc.) and, in general, aims at the transformation of spectroscopic data into a form, most suitable for further processing like information extraction.
Therefore, information caused by the chemistry of objects is aimed to be more dominant compared to influencing disorders.
This page summarizes strategies on how to find a proper preprocessing for an application.
Note: preprocessing can generate noise which can get dominant too, and can potentially influence the preprocessed signal, and therefore can worsen the results. In general it is good practice to decide for a preprocessing mode that doesn't generate too much noise.
The Preview feature tends to show points of similar spectral information in similar color, while points of differing spectroscopic information are shown in distinct colors.
Keep in mind that
In the example below three plastic plates of different chemistry (PA, PMMA, PP) are shown at different preprocessing modes. A color image illustrates the Preview for each preprocessing mode. Next to it, selected spectra from the plates and background are shown.
We assume the three plastic plates to be distinct in chemistry (PA, PMMA, PP) - therefore we look for the preprocessing mode which best describes the plates by distinct color information.
In the following table the influence of the preprocessing method "derivative" is summarized:
The left part of the table above shows reflectance data and data obtained by applying derivation to the data. Please note: By applying the derivation to spectra, constant factors like offset is rejected. Higher order derivation tends to have noisy spectra.
The 2nd derivation differentiates the chemistry best by different colors, but also shows some noise in the images.
In the following table the influence of the preprocessing method "normalization" and "derivative" is summarized:
The table above summarizes the same data as before but additionally normalization was applied.
Please note: By applying normalization to the data the absolute values get lost - each spectra is in the same value range. Background pixels in the scene tend to cause noisy output information because normalization is done without respect of object or background.
The color impression gained is more or less comparable to those of the data without normalization.
When blurring effects aren't critical for the application of interest, the preprocessing "1st derivation normalized" might be suitable since the Preview corresponds to our expectations (different colors). Please note: by applying normalization the spectra are much more decoupled to external influences like light variations - so the spectra are in general much more "stable" compared to those without normalization. In case relevant information gets lost by normalization, which is often the case, this preprocessing will not be the best choice.
|Polymer identification (e.g. sorting)||1st derivative|
|Food identification||Intensity, 1st derivative|
|Impurity detection (in e.g. food)||1st derviative|
|Medical tissue investigation||2nd derivative|
|Minerals identification||Intensity, 1st derivative, 2nd derivative|
|Wood inspection||Intensity, 1st derivative|
|Pharmaceutical investigation||Intensity, 1st derivative|
*) The combination with normalization can be beneficial.
© 2019 by Perception Park GmbH
The content of this page and any attached files are confidential and intended solely for the addressee(s). Any publication, transmission or other use of the information by a person or entity other than the intended addressee is prohibited. If you receive this in error please contact Perception Park and delete copied material. Perception Park GmbH, Wartingergasse 42, A-8010 Graz; Austria; FN 400381x