This article presents both a geometric methodology and an empirical experiment with real plant images to evaluate the performance of different color indices for plant segmentation from two-dimensional digital images. Six well-known color indices given in the literature were selected for the evaluation: the normalized green, the difference between normalized green and normalized red, the normalized difference index, the excessive green index, the modified excessive green index, and the hue. Segmentation performance was assessed on the normalized green-red plane (the rg plane) by using Type I and Type II errors, commonly used in statistics to estimate the errors associated with null hypothesis testing. The two errors were applied in this study to quantifying the misclassified pixel percentages for plant and background, respectively. The rg plane is shown to facilitate visualizing geometrically the behavior of the color indices. An experiment with 240 digital images of four plant types under different lighting and background conditions was carried out to empirically investigate their segmentation performance. Segmentation results showed that the hue achieved the least amount of Type II error with a small loss of plant pixels, which confirms expectations from the geometric analysis on the rg plane.

Keywords