A Deep Learning Approach to Identify Not Suitable for Work Images

• Main Author: Bicho, Daniel
• Publication Date: 2020
• Other Authors: Ferreira, Artur, Datia, Nuno
• Format: Article
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: https://doi.org/10.34629/ipl.isel.i-ETC.80
• topic: Computers; Informatics; Multimedia; Deep Learning; Deep Neural Networks; Image Classification; Not Suitable for Work Images; ResNet; SqueezeNet; Redis Message Queue

Web Archiving (WA) deals with the preservation of portions of the World Wide Web (WWW) allowing their availability for future access. Arquivo.pt is a WA initiativeholding a huge amount of content, including image files.However, some of these images contain nudity and pornography, that can be offensive for the users, and thus being Not SuitableFor Work (NSFW). This work proposes a methodology to classify NSFW images available at Arquivo.pt, using deep neural network approaches. A large dataset of images is built using Arquivo.pt data and two pre-trained neural network models, namely ResNet and SqueezeNet, are evaluated and improved for the NSFW classification task, using the dataset.The evaluation of these models reported an accuracy of 93% and 72%, respectively. After a fine tuning stage, the accuracy of these models improved to 94% and 89%, respectively.The proposed solution is integrated into the Arquivo.pt Image Search System, enabling the filtering of the problematic NSFW images. At the time of this writing, the proposed solution is in production at https://arquivo.pt/images.jsp