Multi-class object recognition using boosted linear discriminant analysis combined with masking covariance matrix method

TL;DR: A new algorithm, boosted linear discriminant analysis (bLDA), for classification of a non-linear pattern distribution, and masking covariance matrix method (MCM) to find optimal local features, instead of traditional exhaustive search in huge number of candidates feature, for robust and realtime object recognition.

Abstract: We propose a new algorithm, boosted linear discriminant analysis (bLDA), for classification of a non-linear pattern distribution, and masking covariance matrix method (MCM) for robust and fast computation of object recognition. bLDA integrates classifiers on multiple linear discriminant spaces. Each linear discriminant space is spanned by eigenvectors so as to maximize ratio of within-class variance and between-class variance of training data. The weights of samples are updated for each boosting step; increasing weights for misclassification while decreasing weights for correct classification. bLDA performs well to classify a complicated data distribution, such as face images. In addition, we propose MCM to find optimal local features, instead of traditional exhaustive search in huge number of candidates feature, for robust and realtime object recognition. In MCM, the covariance vectors of training data set are restricted to be locally correlated, by multiplication of covariance mask. bLDA combined with MCM automatically and effectively extract spatially-local features. Especially, the covariance mask on Haar space induces anisotropic collinearity of object’s contours. bLDAMCM algorithm performed 98.70% correct for face/nonface classification task, after 100 rounds of boosting.