Compact color features with bitwise quantization and reduced resolution for mobile processing

Color features for recognition are often extracted from 8-bit images. However, some studies recommended the use of an arbitrary number of colors, often less than 256 colors. Because the use of less colors can led to a lower computational cost and less power consumption, this paper investigates the extraction of features using images with different pixel depth, i.e., number of colors, and using two resolution settings. We show that it is possible to obtain compact and effective descriptors, by extracting features with images of lower quantization and resolution parameters. Also, we propose a bitwise quantization algorithm that codifies the most significant color features. While it reduces the number of colors, the distances between the feature vectors are kept similar, benefiting mobile image applications.