Markov random field based convolutional neural networks for image classification

Yao Peng, Hujun Yin

Research output: Contribution to conferencePaperpeer-review

Abstract

In image classification, deriving efficient image representations from raw data is a key focus as it can largely determine the performance of a vision system. Conventional methods extract low-level features based on experiments or certain theories, whilst deep learning approaches learn image representations hierarchically with multiple layers of abstraction from vast number of sample images. Markov random fields are generative, flexible and stochastic image texture models, in which global image representations can be obtained by means of local conditional probabilities. Texture has been strongly linked to human visual perception. The ability of deriving global description from local structure shares compatibility with convolutional neural networks. Inspired by this property, we investigate the combination of Markov random field models with deep convolutional neural networks for image classification. Various filters from Markov random field models are first derived to form the features maps. Then convolutional neural networks are trained with prefixed filter banks. Comprehensive experiments conducted on the MNIST dataset, EMNIST database and CIFAR-10 object database are reported.
Original languageEnglish
Pages387
Number of pages396
Publication statusPublished - 2017
EventInternational Conference on Intelligent Data Engineering and Automated Learning -
Duration: 1 Jan 1824 → …

Conference

ConferenceInternational Conference on Intelligent Data Engineering and Automated Learning
Period1/01/24 → …

Keywords

  • Image classification
  • Image representations
  • Markov random field
  • Convolutional Neural Networks (CNN)

Fingerprint

Dive into the research topics of 'Markov random field based convolutional neural networks for image classification'. Together they form a unique fingerprint.

Cite this