Generating Robust Convolutional Networks by Injecting Partial Noise in the Training Data
Artem Pilzak, Jean‐Philippe Thivierge · 2022
Convolutional Neural Networks (CNN) have emerged as a highly efficient deep learning algorithm for processing, classifying, and analyzing images. Currently, there is a high demand for networks that are robust to noise in the training and testing data. However, the presence of noise in images has been an ongoing problem when it comes to deploying CNN for real-world applications. In this paper, we propose a solution to this problem by incorporating noisy and noise-free images in the training set of a CNN network. Using the Modified National Institute of Standards and Technology database (MNIST), we demonstrate that marked improvements to classification accuracy only require a small proportion of noisy images in the training set. This result generalized across test sets with different amounts of noise and types of noise. In sum, our novel training method increases the generalization of a CNN model when processing noisy data.