Influence of Positive Additive Noise on Classification Performance of Convolutional Neural Networks
Jakub Hrabovský, Martin Kontšek, Pavel Segeč, Ondrej Šuch · 2018
Convolutional neural networks have emerged as a leading architecture in computer vision tasks. In practical applications, the input layer of the network may need to process images with added noise. In some applications, such as speech recognition, medical imaging or network intrusion detection systems, the noise will be positive and additive. We evaluate performance of convolutional neural networks on recognition of MNIST and CIFAR datasets with such noise added. Our preliminary findings indicate that convolutional networks are resilient to the noise. However, there appears to be benefit to train the network in matched setting that is with such signal-to-noise ratio in training set that will be encountered in practice.