Probabilistic-based semantic image feature using visual words

This paper presents a new image feature that is based on a semantic-level perspective in order to bridge the semantic gap between low-level features of images and high-level concepts of human perception. In this work, low-level image features are first quantized into a set of visual words, and then we apply probabilistic Latent Semantic Analysis model to automatically analyze what kinds of hidden concepts between visual words and images are involved. Therefore, we collect discovered concepts of an image and filter a part of unreliable concepts out to build a semantic-based image feature. We also discuss in detail how to define parameters for extracting the proposed feature. Several experiments are presented to show the efficiency of this work.