Focused pooling for image fusion evaluation

Abstract

This paper presents the results of an investigation into optimal spatial pooling of localised quality scores for use in objective evaluation of multisensor image fusion. We propose and evaluate a two stage focused pooling method with a localised aggregation of pixel-level performance estimates into regional fusion performance scores as the first step followed by a global pooling of regional scores into a global fusion performance score. We investigate a selection of linear and non-linear global pooling methods and show that quality driven methods which take into account regional fusion performance levels exhibit optimal performance. The proposed pooling algorithm is general and applicable to any fusion performance and quality metric based on structural preservation estimates, local differences between input and fused images. Specifically, we evaluate the proposed method in conjunction with three well-known structural preservation fusion metrics against their baseline pooling methods. We show, through correlation with an extensive subjectively annotated dataset of fused images, that regional aggregation of local performance scores over 3–6° of visual angle with selection of the worst performing region as the global score can improve performance for all the structural fusion metrics tested.