(b), (c) and (d) in different size are used in [8][10] to extract
features. And, Lienhart et al. [11] introduced an efficient
scheme for calculating 45˚ rotated features. And, Mita and
Kaneko introduced a new scheme which makes Haar-like
features be more discriminative [6]. Though Haar-like feature
provides good performance in extracting textures and cascading
architecture and integral image representation make it
computationally efficient, it is still not feasible on mobile
products.