previous
|
next
Figure 2.
Improved segmentation results with multispectral reflected light imagery. (A) We take the same image of an archaeocyathid boundstone sample in a traditional red-green-blue (RGB) colorspace, as well as a false color ultraviolet (UV)-yellow-red space and sample the same pixels for four feature classes in each (colored boxes). (B) In the RGB image, all classes show covariance between color channels, and most pixels fall around the brightness line. (C) In the UV-yellow-red (UV-Y-R) image, covariance between channels is removed for all classes, as evidenced by the fourfold increase in average distance between each pixel and the brightness line (reported as total least squares, TLS). The movement of all classes away from the brightness line into distinct regions of the color space eases segmentation. (D–F) Using a support vector machine (SVM), an automated classification of the RGB image is 65% accurate and does not give high-resolution borders between classes and regions (D, E). In contrast, an SVM segmentation of the UV-yellow-red image is 91% accurate and gives sharp region and class boundaries more suitable for measurements (D, F).
previous
|
next