Towards Graffiti Classification in Weakly Labeled Images Using Convolutional Neural, Networks

Abstract

Graffiti is an urban phenomenon that can be useful as an indicator of social and economic factors of a geographic region or community. Automatically identifying this urban writings can be useful for understanding cities and their communities. In this paper we investigate the use of Convolutional Neural Networks aiming at classifying weakly labeled images to identify the presence or absence of graffiti art in images. We propose the use of a VGG-16 architecture pre-trained on the ImageNet dataset and show a novel approach to fine-tuning the network over graffiti examples extracted from Flickr. Experiments using this approach show accuracy comparable to that of ImageNet classes.