On the detection of low-resolution skin regions in surveillance images

This paper presents a study into the detection of skin regions in images where faces may be low-resolution. We focus on surveillance footage and no assumptions are made about fine facial features being visible. This type of data presents the further challenge of changes in appearance of skin regions due to changes in both lighting and resolution. We investigate the use of common colour spaces (YIQ, YCbCr, HSV and RGB) and the effects of histogram size (number of bins) and dimensions (number of channels) by comparing the results to ground-truthed data. Error is measured per-pixel using the raw likelihood weights. We first use a non-parametric classification scheme based on a histogram similarity measure the Battacharyya coefficient. Comprehensive results indicate that the YIQ colour space with 16 histogram bins gives the most accurate performance over a wide range of imaging conditions for nonparametric skin classification. We then compare this nonparametric method to established classification techniques. These are: Gaussian, Bayesian and Thresholding methods. We demonstrate better performance of the non-parametric approach vs. the other methods. We also show results of face-detection via a simple aspect-ratio.