Binarizing Training Samples with Multi-threshold for Viola-Jones Face Detector

We propose an alternative method to Viola-Jones face- detector: in learning stage, we replace each of the gray-scale training images with multiple binary images using multi-threshold; and in detection stage, we use a binarized input-image instead of a gray-scale one. We call this method "TMBMT" (Training by Multiple Binarized samples using Multi-Threshold). Using face images of 1040 individuals from the CAS-PEAL face database, we show that proposed face-detector improves the conventional Viola-Jones facedetectors in terms of numbers of both missed-faces and false-alarms. We also discuss (hypothetical) reasons for the improved performance with (partially) supporting evidence: (1) binarization leads to sharpening of feature distribution; (2) use of multi-threshold leads to better selection of feature location. We speculate these two reasons explain the improved performance.