The Impact of Padding on Image Classification by Using Pre-trained Convolutional Neural Networks

The work presented in this paper aims to investigate the effect of pre-processing on image classification by using CNN pre-trained models. By considering how different quality factors of the input images affect the performances of a CNN based classifier, we propose a preprocessing pipeline (i.e., padding) that is able to improve the classification of the model on challenging images. The presented study allows to improve the performances by only acting on the input images, instead of re-training the model or augmenting the number of CNN's parameters. This finds very practical applications, since such model adaptation requires high amounts of labelled data and computational costs.