Event recognition on images by fine-tuning of deep neural networks

Dmitry Yudin, Bassel Zeno

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

5 Citations (Scopus)

Abstract

The paper considers usage of fine-tuning of the deep neural network ensemble for recognition of 60 event types in the set of 60,000 images from WIDER database. The applied ensemble consists of two deep convolutional neural networks (CNN) using the GoogLeNet architecture, previously trained on other image bases: ImageNet and Places. Separately the accuracy of recognition of 10 events was analyzed: “Car Racing”, “Ceremony”, “Concert”, “Demonstration”, “Football”, “Meeting”, “Picnic”, “Swimming”, “Tennis” and “Traffic”. During the ensemble training output layer in the each of deep CNN is replaced to the layer with respectively 10 and 60 neurons and we tune only weights which connect output layer with previous one. The classification accuracy of 10 event classes from the WIDER image database averages 83.22%, for 60 event classes accuracy is 50.4%. In addition, the approach based on the automatic features formation using deep CNN provided a much better recognition quality of social events compared to the choice of features manually (LBP, LDP or HOG) and their further classification by support vector machine. The testing time of the developed ensemble provides the possibility of using the classifier in practical applications of event recognition with a processing speed up to 20 frames per second.

Original languageEnglish
Title of host publicationProceedings of the 2nd International Scientific Conference on Intelligent Information Technologies for Industry, IITI 2017
EditorsSergey Kovalev, Andrey Sukhanov, Margreta Vasileva, Valery Tarassov, Vaclav Snasel, Ajith Abraham
PublisherSpringer Verlag
Pages479-487
Number of pages9
ISBN (Print)9783319683201
DOIs
Publication statusPublished - 2018
Externally publishedYes
Event2nd International Conference on Intelligent Information Technologies for Industry, IITI 2017 - Varna, Bulgaria
Duration: 14 Sep 201716 Sep 2017

Publication series

NameAdvances in Intelligent Systems and Computing
Volume679
ISSN (Print)2194-5357

Conference

Conference2nd International Conference on Intelligent Information Technologies for Industry, IITI 2017
Country/TerritoryBulgaria
CityVarna
Period14/09/1716/09/17

Keywords

  • Deep learning
  • Fine-tuning
  • Image recognition
  • Neural network
  • Social event

Fingerprint

Dive into the research topics of 'Event recognition on images by fine-tuning of deep neural networks'. Together they form a unique fingerprint.

Cite this