Enhanced LPQ Based Two Novel Blur Invariant Face Descriptors in Light Variations

Among numerous local descriptors persisting in literature, LPQ is one of most influential blur invariant descriptor. But when high content blurring is mixed with light variations (moderate and high) then LPQ and many others fails to achieve required robustness to declare the efficient face descriptor. This gap is filled by introducing advance versions of LPQ so-called Enhanced LPQ1 (ELPQ1) and ELPQ2. Precisely, blurred affected image (with light variations), by using Gaussian low pass filtering, is passed to 2 novel pre-processing approaches before LPQ feature extraction. In former approach edge oriented (enhanced) images are produced by 3 methods i.e. Sobel (magnitude gradient), Image sharpening and Kirsch (magnitude gradient). In latter approach, Sobel (magnitude gradient) is replaced by more robust magnitude gradient, called Sobel + Prewitt (magnitude gradient), produced by using Sobel horizontal gradient and Prewitt vertical gradient. Rest 2 methods remains similar in both the pre-processing approaches. Further LPQ is imposed for feature extraction. As each pre-processing approach contains 3 methods therefore 3 transformed images are produced after LPQ. Histograms under respective category are integrated to build feature size of 2 robust blur invariant representations ELPQ1 and ELPQ2. Compressed feature size is attained from FLDA and classification is assisted from SVMs. Experiments on GT and EYB proves efficacy of ELPQ1 and ELPQ2 against benchmark methods.