Abstract
A modified version of Boosted Mixture of Experts (BME) for low-resolution face recognition is presented in this paper. Most of the methods developed for low-resolution face recognition focused on improving the resolution of face images and/or special feature extraction methods that can deal effectively with low-resolution problem. However, we focused on the classification step of face recognition process in this paper. Using Neural Networks (NN) combinations is an efficient approach to deal with complex classification problems, such as the low-resolution face recognition which involves high-dimensional feature sets and highly overlapped classes. Mixture of Experts (ME) and boosting methods are two of the most popular and interesting NN combining methods, which have great potential for improving performance in classification. A modified combining approach based on both features of ME and boosting is presented in order to deal with this complex classification problem efficiently. Previous works [1,2] made attempts to incorporate the complementary features of boosting method in ME training algorithm to boost the performance. These approaches called Boosted Mixture of Experts (BME) have some drawbacks. Based on the analysis of the problems of previous approaches, some modifications are suggested in this paper. A modification in the pre-loading (initialization) procedure of ME is proposed to address the limitations of previous approaches and overcome them using a two stages pre-loading procedure. In our suggested approach, both the error and confidence measures are used as the difficulty criteria in boosting-based partitioning of the problem space. Regarding the nature of this approach, we call the proposed method Boosted Pre-loaded Mixture of Experts (BPME). The proposed method is tested in a low-resolution face recognition problem and compared to the other variations of ME and boosting method. The experiments are conducted using low-resolution variations of two common face databases including the ORL and Yale databases. The experimental results show that BPME method has significant better recognition rates against the other compared combining methods in various tested conditions including different quality grades of face images and different sizes of the training set.