Improved PCA based face recognition using similarity measurement fusion

This paper proposes a technique for human face recognition based on improved Principal Component Analysis (PCA). Specifically the proposed method apprehends the unsupervised PCA technique and transforms it into a supervised classification approach by imparting feedback of the classification correctness of each of the principal components obtained. The proposed method trounces the disadvantages of the conventional PCA based methods because the optimality criteria of the classification are related directly and unique to the training dataset. Moreover, the proposed method consists of four different similarity measurement techniques and the final decision on the identified face is developed based on a proposed probability based decision fusion method. An added intermediate training phase is utilized to come up with the Bayesian probability model used in the fusion. The tests were carried out using the Extended Yale B face database with multiple test cases in order to overcome the effect of small sample size problem on the results. In addition to the standard face database which has controlled formatting on head position and illumination, the proffered approach has proven to produce high rates of recognition for non-standard face database created by us, notably with successful identification of lookalike twins.