New Face Recognition Algorithm Adopting Wide Fast Embedded Capsule Networks with Reduced Complexity and Preserved Accuracy

Computer Vision has come a long way after the introduction of Convolutional Neural Networks, that simulated the first perception layers in the human vision, specially in classification, and segmentation tasks. With Convolutional layers came maximum pooling that is not natural and is the reason for information loss and the lack of preserving spatial information of the patterns, that is why Capsule Networks were introduced. Capsule Networks handle patterns as vectors preserving spatial information of the patterns along with their pose but at the cost of having slow processing and high complexity. Wide Fast Embedded Capsule Networks were introduced as the faster and simpler version of Capsule Networks. However, they could not handle complex datasets like Labeled Faces in the Wild (LFW). That is the reason Wide Fast Embedded Capsule Networks are proposed in this paper to handle intermediate complex datasets like LFW boosting the speed boost, reducing complexity and preserving accuracy. Experimental results show that the speed is tripled, the complexity is reduced by 80.6% and the accuracy is preserved at 93.7%.