A Lightweight Inference Method for Image Classification

We demonstrate a two phase classification method, first of individual pixels, then of fixed regions of pixels for scene classification—the task of assigning posteriors that characterize an entire image. This can be realized with a probabilistic graphical model (PGM), without the characteristic segmentation and aggregation tasks characteristic of visual object recognition. Instead the spatial aspects of the reasoning task are determined separately by a segmented partition of the image that is fixed before feature extraction. The partition generates histograms of pixel classifications treated as virtual evidence to the PGM. We implement a sampling method to learn the PGM using virtual evidence. Tests on a provisional dataset show good (+70%) classification accuracy among most all classes.