Image Background Noise Impact on Convolutional Neural Network Training

Small size dataset is general issue that we may encounter when training neural networks for analysis of image data. There are many cases when networks can not start training even with data augmentation. This paper proposes a new method how to allow training of image classification even when traditional approaches fail. It presents an experiment, which shows that subtraction of redundant background from images can significantly improve convergence of neural network training. Improvement is not in accuracy matter but it means that neural network is able to train and to start convergence. For experimental evaluation, person binary classification was used and compared to experiments, where the background was removed.