Abstract
This paper present an alternative approach to PDLDA for incremental data which belong to old/known and new classes called as incremental PDLDA (IPDLDA). The IPDLDA not only can overcome the main problem of the conventional LDA in terms of large computational cost for retraining but also can provide almost the same optimum projection matrix (W) as that original LDA for each incremental data. The proposed method can be realized by redefining new formulation for updating the between class scatter (Sb) using constant global mean assignment and simplifying the equation for updating the within class scatter (Sw). These new updating algorithms make the IPDLDA require much less time complexity for retraining the incremental data. In addition, they also make the IPDLDA have almost the same properties as the original one in terms of the power discriminant and scattering matrix. To know the ability of the IPDLDA on features clustering, we implement it for face recognition with the DCT-based holistic features as the dimensional reduction of raw face image. The experimental results show the proposed method provides robust recognition rate and less processing time than that of GSVD-ILDA and SP-ILDA in several challenges databases when the experiments were done by retraining the system using two scenarios: the incremental data belonging to new and old classes.
