Information: Theoretical Model for Saliency Prediction—Application to Attentive CBIR

Abstract

This work presents an original informational approach to extract visual information, model attention and evaluate the efficiency of the results. Even if the extraction of salient and useful information, i.e. observation, is an elementary task for human and animals, its simulation is still an open problem in computer vision. In this article, we define a process to derive optimal laws to extract visual information without any constraints or a priori. Starting from saliency definition and measure through the prism of information theory, we present a framework in which we develop an ecological inspired approach to model visual information extraction. We demonstrate that our approach provides a fast and highly configurable model, moreover it is as plausible as existing models designed for high biological fidelity. It proposes an adjustable trade-off between nondeterministic attentional behavior and properties of stability, reproducibility and reactiveness. We apply this approach to enhance the performance in an object recognition task. As a conclusion, this article proposes a theoretical framework to derive an optimal model validated by many experimentations.