In microscopy image analysis, images of pollen are particularly difficult to analyze automatically. The major reason of this difficulty emerges from the complex nature of the pollen, which presents great variability and similarities between classes and even on the same classThis paperpropose a feature selection methodology with dimensionality reduction for pollen image classification. First, we extracted five descriptors to represent images: The histogram of image (HISTO), the histogram of the oriented gradients (HOG) and three alternatives of the local binary pattern: the histogram of the LBP (LBP), the Fast Fourier Transform Applied to the histograms of LBP (LBPFFT) and the Discrete Cosine Transform applied to the histograms of LBP (LBPDCT). Secondly, we concatenated the different descriptors before selecting the most discriminative set of features for pollen images classification by the analysis of neighborhood components of the descriptors. We experimented the proposed methodology on a dataset of pollen images of the Adamawa region of Cameroon, named LESIA_Pollen. We performed the post-hoc Bonferroni test and the analysis of variance for the statistical analyze of the results. Our result demonstrate that we get more precision with overall accuracy of 98.83%.
Keywords: Histogram; LBP; Fourier transform of LBP; Cosine transform of LBP; HOG; Neighborhood component analysis; pollen image