2026-05-30T04:09:31.610Z http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai

ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai:1:7019 2019-05-30T11:33:29Z 1

Image tagging with an ensemble of deep convolutional neural networks 2017 http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai/record/1/7019 М. Јоцић Đ. Obradović В. Малбаша З. Коњовић We present a method for image tagging, i.e. assigning a set of tags/labels to an image. Three popular architectures of deep convolutional neural networks were used: VGG16, Inception-V3, and ResNet-50, which were pretrained on the ImageNet data set for a classification problem, and then fine-tuned on the HARRISON data set for the image tagging problem. The final model consists of an ensemble of these three convolutional neural networks, whose outputs were combined by different methods: averaging, voting, union, intersection and by two-layer feedforward neural network. We verified these models on the hashtag recommendation for images from social network task, with a predefined set of 50 possible hashtags. All of the models were evaluated using the following metrics: precision, recall and F1-measure. conferenceObject 13 17 7th International Conference on Information Society Technology and Management, ICIST 2017