Real-time gender recognition with unaligned face images

In image-based face profiling classification, face alignment, which transforms the face images in order that the 2D coordinates of the facial features are consistent with each other, is considered as an important preprocessing step before the actual classification to maximize classification accuracy. Unfortunately, accurate face alignment is a time-consuming process and a challenging problem especially for low-resolution face images, where eye localization is a tough task. In this paper, we aim to develop a real-time gender recognition system where face alignment is omitted and the detected face bounding box is used directly. Three measures are employed to compensate for the errors caused by the misalignment: firstly, unaligned faces are included in the training set; secondly Local Binary Patterns (LBPs) and Gabor features are extracted to represent the face images and components, respectively and thirdly a probability model is proposed to fuse holistic- and component-based features with a boosting scheme. The experiments using unaligned and manually aligned faces show that a comparable accuracy can be obtained even without alignment.