A new face descriptor using local un-quantized patterns

摘要

We present a novel face representation based on local un-quantized patterns (LUP) descriptors. LUP descriptor is a simple yet powerful descriptor which measures the difference of intensities between surrounding pixel with the center in a local neighborhood, but preserves the finer local geometric structure unlike LBP, SIFT or HOG (which uses either the quantized version of local gray level patterns or quantized codes of image gradients). This descriptor also solves the problem of limited spatial support of LBP like operators, where increasing the size of local-neighborhood increases the histogram dimensions exponentially making it unsuitable for real-time needs. By applying principal component analysis (PCA) to LUP, we develop a new srepresentation, which gives better performance than LBP and comparable performance to LARK while only taking a fraction of the computation when compared to the latter.