Patch-based image classification through conditional random field model

We present an image classification system based on a Conditional Random Field (CRF) model trained on simple features obtained from a small number of semantically representative image patches. The CRFs are very powerful to handle complex parts dependencies due to their approach based on the effective modelling of the source probability conditioned on the evidence data, and they have been applied successfully to image classification and segmentation tasks in presence of a large number of low level features. In this paper an agile system based on the application of CRFs to images coarsely segmented is introduced. The main advantage of the system is a reduction in the required training time, slightly sacrificing the classification accuracy. The model implementation is described, experimental results are presented and conclusions are drawn.