Unsafe image classification using convolutional neural network for brand safety
Quan Minh Vo, Nhan Thi Cao, An Hoa Ton-That · 2020 IEEE Asia-Pacific Conference on Computer Science and Data Engineering (CSDE) · 2020
Contextual advertising is an advertising form in which the advertisements are served by automated systems based on the context of the host websites. In contextual advertising, deciding the matching level between a brand and a website context is crucial. Brand safety is a strategy to keep a brand's reputation safe by avoiding advertisement placements on inappropriate websites. In this paper, we approach brand safety as an image classification issue to classify and filter websites containing inappropriate images. We propose an image multi-class classification method in order to recognize some categories using for brand safety purpose. We also apply pruning techniques in order to reduce model complexity. In addition, we create a dataset including 56,501 images in six categories: adult, gambling chip, gore, gun, knife and safe. Experimental results show that our proposed method achieves better performance compared to some original methods.