Markov Random Field Based Convolutional Neural Networks for Image Classification

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.