Color in Computer Vision: Fundamentals and Applications

Chapter 14: Color Feature Description

With contributions by Gertjan J. Burghouts

In the previous chapters, we have outlined the theory of invariant feature extraction from color images.¹ The advantage of the full invariants described in Chapter 6 is that they capture intrinsic scenes or object properties, robust to various arbitrary imaging conditions such as local illumination, shadows, and color of the light source. Hence, these invariant features are well suited to characterize the image content in the so-called image descriptors. In this chapter, we demonstrate the appropriateness of such invariant color descriptors. Much of the methodology described here is adopted from Burghouts and Geusebroek [258] and from van de Sande et al. [259].

Many computer vision tasks depend heavily on local feature extraction and matching. Object recognition is a typical case where local information is gathered to obtain evidence for recognition of previously learned objects. Recently, much emphasis has been placed on the detection and recognition of locally (weakly) affine invariant regions [55, 57, 260–262]. The rationale here is that planar regions transform according to well-known laws. Successful methods rely on fixing a local coordinate system to a salient image region, resulting in an ellipse describing local orientation and scale. After transforming the local region to its canonical form, image descriptors should be well able to capture the invariant region appearance. As pointed out by ...