Classification of protein crystallization images using Fourier descriptors

The two-dimensional Fourier transform (2D-FT) is well suited to the extraction of features to differentiate image texture, and the classification of images based on information acquired from the frequency domain provides a complementary method to approaches based within the spatial domain. The intensity, I, of the Fourier-transformed images can be modelled by an equation of power law form, I = Ar^α, where A and α are constants and r is the radial spatial frequency. The power law is fitted over annuli, centred at zero spatial frequency, and the parameters, A and α, determined for each spatial frequency range. The variation of the fitted parameters across wedges of fixed polar angle provides a measure of directionality and the deviation from the fitted model can be exploited for classification. The classification results are combined with an existing method to classify individual objects within the crystallization drop to obtain an improved overall classification rate.