Evoluindo Redes Neurais Convolucionais na Detecção de Emoções Usando Micro AGs

The Deep-emotive v.1 is a CNN that recognizes emotions by thehuman face’s pictures. In this context, the CNN’s structure creationdepends on several hyperparameters, which impact the resultspositively or negatively. The Genetic Algorithm implementationallows us to explore the search space of these hyperparametersto find the best architecture for solving the problem. The definedsearch space is formed by the combination of both the numberof convolutional layers and the fully connected ones, the numberof filters for each layer, the size of filters, the subsampling type,and the number of nodes in the fully connected layer. This paper proposes to improve the Deep-Emotive network with the imple-mentation of Convolutional Neural Networks (CNNs) architectures using Genetic Algorithms. The FER-2013 dataset was chosen toclassify seven emotions by images of facial expressions, as it had the worst performance in the first version of the network, reach-ing an accuracy of 60.71%. This dataset has images with common problems for computer vision algorithms, such as occlusion, im-balance, perspective, noises, as well as images that do not exist in the context of emotions. The experiment’s results indicate thatthe proposed approach can generate a CNN architecture with anaccuracy of 63,84% in the train set and 62,39% in the validationset. Despite a low-performance rate, the experiments indicate thatthe algorithm can generate more adapted individuals who havealready overcome the performance achieved by the first version ofthe network defined empirically. Thus, results show potential forexploitation in environments with more computational resources.