Characters Recognition based on CNN-RNN architecture and Metaheuristic

Abstract

Convolutional neural networks (CNN) are composed of multiple convolutional layers and a fully connected layer(s) (FC). In most of CNN models, the memory needed only for the weights of FC layers exceeds the total required by the rest of the layers. Consequently, for decreasing memory size needed and the acceleration of the inference, it obvious to focus on the an FC layer optimization method. In this paper, we propose a hybrid neural network architecture to perform image classification that combines CNN and the recurrent neural networks (RNN) to deal with the presented problem. To do so, a pretrained CNN model is used for features extraction (without FC Layers), then plugged into a parallel architecture of a RNN. In this work the Hopfield is considered. The obtained results on the Noisy MNIST Dataset have exceeded the state of the art for this problem.