An evaluation of 2D-wavelet filters for estimation of differences in textures of pharmaceutical tablets
Journal article, 2006
In chemical imaging spectra are acquired over a surface with one spectrum for each pixel of the image. The obtained spectra usually carry a mixture of chemical and physical information. One may view the properties that vary over the image, the mean spectral magnitude from separate wavelength intervals, or better, PCA scores may be shown as images.
In this way a multitude of images are compressed to a few images that in the PCA case are representative for the main variation in the sample images. These images may be viewed manually and deductions as to e.g. differences in homogeneity can be made. At an increased rate of samples, the observer will have difficulties coping with the repetitive work and different observers will most likely have slightly different interpretations. In order to automate the process of estimation of e.g. homogeneity and particle density, image filters can be used to calculate a small set of texture descriptors for each image. Calculations based on the 2D versions of the discrete wavelet transform (DWT) using Daubechies 14 and the dual tree complex wavelet transform (DT-CWT) using near-symmetric 13, 19 tap filters in combination with q-shift 14, 14 tap filters were evaluated for this purpose.
The aim with this work is to evaluate texture descriptors based on a combination of 2D-wavelet filters and energy, i.e. l(1)-norm, calculations for each wavelet scale. These descriptors are then used as observations for overview in e.g. PCA. In this way the texture differences can be ranked by ordinary use of PCA or PLS.
This method is tested on multivariate near infrared images of pharmaceutical tablets. Score images are selected to represent variations of the aggregate density and sizes in the compressed tablets. Images are shifted and rotated to compare shift and rotational independence of the texture descriptors.
principal components analysis