A Novel Method for the Separation of Overlapping Pollen Species for Automated Detection and Classification

Table 7

Dataset comparison of the proposed method and other methods, as appeared in the literature.

Reference	Detection	Description	Classification results	Results
Proposed	A database of 12 pollen species is generated and a MS-Otsu filter applied to separate and regroup overlapping pollen grains using morphological operations and GVFS	Shape, first- and second-order texture	Multilayer perceptron (MLP)	PR	RE	FM
			10-fold cross validation (0.9–0.1)	0.974	0.975	0.974
			5-fold cross validation (0.8–0.2)	0.961	0.962	0.961
			2-fold cross validation (0.5–0.5)	0.961	0.962	0.961
Kaya et al. [6]	Classifies 20 different pollen types obtained from the genus Onopordum L. (Asteraceae) by a rough set-based expert system. For each pollen grain, 30 different images were photographed (600 total)	Microscopic features: polar axis (P), equatorial axis (E), P/E, exine, intine, tectine, nexine, columella, colpus L, and colpus W	The 440 samples were used for training and the remaining 160 samples were used for testing (600 total)	The overall success of the RS method in recognition of the pollen grains; PR = 90.625%; (145/160 pollen samples)
Dell’Anna et al. [7]	Discriminates and automatically classifies pollen grains from 11 different allergy-relevant species belonging to 7 different families	Fourier transform infrared (FT-IR) patterns	Applied statistical analysis unsupervised (hierarchical cluster analysis, HCA) and supervised (k-NN neighbors classifier, k-NN) learning method in the process of pollen discrimination	Obtained accuracy of 80% for the 11 species classified and 84% for 9 species
Mitsumoto et al. [5]	Used in autofluorescence images to simplify the problem by splitting pollen into RGB channels. Assuming circularity on the particles	Particles size	Presents the relationship between the grain diameter and ratio of the pollen grains	The results show that values for the pollen grains of a given species tend to cluster within a limited area of the graph (lack of quantitative results)
Ranzato et al. [4]	Blurring the image into two bandwidths and , to calculate Gaussian diffrences	Local jets (shape descriptor and texture information)	Bayesian classifier. Others are tried, but there is very little improvement; train-test (90%–10%) random selection… 10 times (100%–6.8%) 100 times (100%–23.2%) after based on the use of false classifications in the training data	Train-test (90%–10%) random selection… 10 times (100%–6.8%) 100 times (100%–23.2%)
Rodriguez-Damian et al. [3]	Looks for circular grains, as most pollen grains present have this shape. Tests some edge detection techniques to find a good shape of each pollen grain	Shape: common geometrical features (CGF); statistical moments; statistical moments; Fourier descriptors; Texture: Haralicks’s coefficients; gray level run length statistics; gray level run length statistics	Minimum distance classifier using preselected attributes; SVM	Texture (88%); boundary features (80%); they try fusing classifiers and improving the result (89%)