Face recognition using modular Neural Networks

Monolithic Neural Networks are generally prone to sub-optimal performance in highly complex and dimensional problems that hinders learning. Modular Neural Networks employ a divide and conquer strategy to convert a complex problem into a set of simpler problems. In classification this means focus upon local features and making of simpler feature space. The simpler problems in a modular architecture are solved by different modules or experts, each of whose outputs are integrated to give the final output. Each module is a combination of feature extraction technique and classifier, and returns a matching score as output. The paper presents a two step modular architecture. At the first step the facial image is decomposed into 3 sub-images. At the second stage each sub-image is solved redundantly by two different neural network models and features extraction techniques. Two step integration is performed using probabilistic sum, min, max, product and polling integration techniques. The proposed modular architecture gives improvised matching score with all integration techniques.