Image Classification by Optimized Convolution Neural Networks

Abstract

Considering the fact that digital images are used in almost all scientific areas and they are a big part of everyday life, it is obvious that the importance of good methods for processing and analyzing them is great. One of the most frequent tasks in various applications that use digital images is image classification. A revolutionized improvement in this area was achieved with convolutional neural networks (CNN). The convolutional neural networks managed to achieve classification accuracy significantly better compared to previously proposed and used methods. Even better results can be obtained by tuning CNN hyperparameters. Since this is a hard optimization problem, swarm intelligence algorithms can be successfully used. In this paper, we propose bare bones fireworks algorithm for tuning a selected subset of hyperparameters and it was tested on the benchmark dataset for handwritten digit recognition, MNIST. The proposed method achieved higher classification accuracy compared to the methods from the literature.