Analysis of LDA-based matching schemes for Face Verification

The performance of face verification system using linear discriminant analysis (LDA) depends on many factors such as the image geometric registration, the feature space dimensions, photometric normalisations, matching schemes adapted in the LDA subspace, etc. In this paper, the relationships among these factors were explored by illustrating experimental results obtained on a publicly available face database using corresponding experimental protocol. A new innovative matching (NIM) scheme was proposed and investigated by comparing with two most often used the Euclidean distance (EUD) measure and the normalised correlation (NOC) measure in the LDA subspace obtained on the same training images. Experimental results shown that different matching schemes exhibited various efficients on photometric normalisations but these efficients kept consistent to the geometric normalisations. Among these three matching schemes, the new innovative matching scheme (NIM) achieved the best performance when the histogram equalisation is applied or no photometric normalisation is performed. The Euclidean distance measure slightly outperformed the other two measures when the zero mean and unit variance photometric normalisation was employed.