A Novel Approach for Face Recognition Using Fused GMDH-Based Networks

This paper explores a novel approach for automatic human recognition from multi-view frontal facial images taken at different poses. The proposed computational model is based on fusion of the Group Method of Data Handling (GMDH) neural networks trained on different subsets of facial features and with different complexities. To demonstrate the effectiveness of this approach, the performance is evaluated and compared using eigen-decomposition for feature extraction and reduction with a variety of GMDH-based models. The experimental results show that high recognition rates, close to 98%, can be achieved with very low average false acceptance rates, less than 0.12%. Performance is further investigated on different feature set sizes and it is found that with smaller feature sets (as few as 8 features), the proposed GMDH-based models outperform other classifiers including those using radial-basis functions and support-vector machines. Additionally, the capability of the group method of data handling algorithm to select the most relevant features during the model construction makes it more attractive to build much simplified models of polynomial units.