Some new color features and their application to cervical cell classification

摘要

New color features derived from multi-dimensional histograms of multi-color image data are described. For two colors, these features are derived from the summation of the normalized two-dimensional histogram and its transpose in the same manner as Haralick's textural features are derived from the gray value co-occurrence matrix. When applied to cervical cell classification, these features resulted in the error rate being reduced from 5.3% to 2.5% for distinguishing cancerous and precancerous cells from normal cells in cervical smears.In a second aspect of this paper, Fourier descriptors of cell boundaries and the advantages of 0.7 micron over 1.0 micron sampling of the images are also investigated in the context of cervical cell classification. Neither the inclusion of the boundary features nor the higher scanning resolution resulted in as significant a reduction in error rate as the new color features did.