Comparison of supervised and unsupervised methods to classify boar acrosomes using texture descriptors

This work compares supervised and unsupervised techniques to classify images of boar sperm heads according to their membrane integrity. We have used 5 different descriptors to characterize the texture of the acrosomes: Laws method, Legendre moments, Zernike moments and 4 and 13 of the features proposed by Haralick extracted from the co-occurrence matrix. We have carried out the classification using Fisher Linear Discriminant Analysis (LDA). Quadratic Discriminant Analysis (QDA), k-Nearest Neighbours and Backpropagation Neural Networks to classify them. Results show that unsupervised classification methods have better performance than supervised ones: The former yield a best error rate 6.11%, while the latter achieved a best error rate of about 9%.