An Improved Convolutional Neural Network Architecture for Image Classification

This manuscript presents the design and implementation of an improved convolutional neural network (CNN) for image classification which was carefully crafted to avoid overfitting. Contrary to most CNNs which apply normalization before pooling, our proposed architecture reverse the order of such tasks. The performance of the proposed architecture, named ACEnet, was evaluated using a hold-out method over five selected databases: Olivia, Paris, Oxford Buildings, Caltech-101, and Caltech-256. We present three main results: processing time, training performance and testing performance for each database. Also, we present a comparison versus the well-known Alexnet architecture, where our CNN proposal improves 5.11% the mean testing performance over the selected databases.