Deep Neural Networks with Markov Random Field Models for Image Classification

Abstract

As one of the most intensively researched topics, image classification has attracted significant attention in recent years. Numerous approaches have been proposed to derive robust and effective image representations and to counter the intra-class variability. Conventional feature extraction and the recent deep neural networks are two common methods to find a good set of features for image description and recognition. Apart from these features-based approach, Markov random fields (MRFs) are generative, probabilistic image texture models, in which global model can be obtained by means of local relations and neighbourhood dependencies. This kind of property shares compatibility with convolutional neural networks (CNNs) and enables combination of CNNs and model-based MRF features. In this work, we propose an MRF loss function in CNNs to minimise modelling errors and estimate parameters. Incorporated with CNNs, these estimated parameters are utilised as the initialised weights in the first convolutional layer. Then the networks are trained. Comprehensive experiments conducted on the MNIST, CIFAR-10 and CIFAR-100 datasets are reported to verify the proposed approach.