Constructive Autoassociative Neural Network for Facial Recognition

Bruno J. T. Fernandes, George D. C. Cavalcanti, Tsang I. Ren, Gennady Cymbalyuk
2014 PLoS ONE  
Autoassociative artificial neural networks have been used in many different computer vision applications. However, it is difficult to define the most suitable neural network architecture because this definition is based on previous knowledge and depends on the problem domain. To address this problem, we propose a constructive autoassociative neural network called CANet (Constructive Autoassociative Neural Network). CANet integrates the concepts of receptive fields and autoassociative memory in
more » ... dynamic architecture that changes the configuration of the receptive fields by adding new neurons in the hidden layer, while a pruning algorithm removes neurons from the output layer. Neurons in the CANet output layer present lateral inhibitory connections that improve the recognition rate. Experiments in face recognition and facial expression recognition show that the CANet outperforms other methods presented in the literature. Autoassociative memory inspired the development of neural networks for oneclass classification tasks, enabling the models to be suitable for learning with only positive patterns. Consequently, this establish closed decision boundaries in the input space. In previous works, we proposed two neural networks for visual pattern classification: LIPNet [7] and AAPNet [15] . The former is a pyramidal neural network (i.e. a neural network with 2-D layers connected in cascade where each layer is smaller than the previous layer) with lateral inhibition that showed good results for dichotomous problems, such as face detection and forest detection. The latter is an autoassociative pyramidal neural network for one-class classification without lateral inhibition that achieved good recognition rates in the multiclass task of object categorization. LIPNet, AAPNet, and other neural networks presented in the literature that use at least one of the discussed biological concepts (receptive fields, lateral inhibition, autoassociative memory) share the same problem: the configuration of the model have to be defined prior to the learning step generating a neural network with a static structure. The learning procedure of such neural networks are then restricted to the changing of the synaptic connections weights and dependent on knowledge from a specialist for the prior architecture configuration. Constructive learning [16] algorithm can overcome these limitations including new neurons or connections in the neural network topology as a function of the learning process. Ma and Khorasani [17] proposed a model with a constructive feedforward neural network for facial expression recognition. Additionally, a pruning step [18] was performed to reduce the size of the architecture without sacrificing the performance of the neural network. The technique proposed by Ma and Khorasani outperforms the methods using fixed neural network structure in computational efficiency, generalization and recognition performance capability. In this paper, we propose the Constructive Autoassociative Neural Network (CANet) that incorporates the concepts of receptive fields, autoassociative memory, and lateral inhibition. The receptive field concept is used for implicit feature extraction preserving the spatial topology of the extracted features. Implicit feature extraction [19] enables the neural network to learn patterns from raw data integrating feature extraction and pattern classification in the same structure. On the other hand, the autoassociative memory is applied to reconstruct the input image with the features implicitly extracted from the raw image by CANet. A constructive-pruning algorithm dynamically updates the neural network architecture during the learning process in order to reduce the difference between the input image and the output of the CANet. The lateral inhibition in the output layer of CANet improves the recognition capability of the neural network [7, 13] . The parameters for the lateral inhibition are experimentally defined in this work.
doi:10.1371/journal.pone.0115967 pmid:25542018 pmcid:PMC4277427 fatcat:pq4fplb4dbcsznwr54u6u6tahm