recolorize: improved color segmentation of digital images (for people with other things to do)

Abstract

Color is an important source of biological information in fields ranging from disease ecology to sexual selection. Despite its importance, most metrics for color are restricted to point measurements. Methods for moving beyond point measurements rely on color maps, where every pixel in an image is assigned to one of a set of discrete color classes (color segmentation). Manual methods for color segmentation are slow and subjective, while existing automated methods often fail due to biological variation in pattern, technical variation in images, and poor scalability for batch clustering. As a result, color segmentation is the common bottleneck step for a majority of existing downstream analyses. Here we present recolorize, an R package for color segmentation that succeeds in many cases where existing methods fail. Recolorize has three major components: (1) an effective two-part clustering algorithm where color distributions are binned and combined according to perceived similarity in a frequency-independent manner; (2) a toolkit for minor manual adjustments to automatic output where needed; and (3) flexible export options. This paper illustrates how to use recolorize and compares it to existing methods, including examples where we segment formerly intractable images, and demonstrates the downstream use of methods that rely on color maps.