Optimizing visual attention models for predicting human fixations using Genetic Algorithms

Abstract

Predicting where humans look in a scene is crucial in tasks like human-computer interaction, design, graphics, image and video compression, and gaze animation. This work proposes the use of a mixed-integer constraint Genetic Algorithm (GA) for searching the optimal parameters of a bio-inspired visual saliency model for accurate prediction of human eye fixations. Bioinspired visual saliency models are complex models, mimicking the primate visual system with a vast choice of design parameters that can be tuned to achieve optimal performance. The bottom-up visual attention model used in this study was trained on three challenging image datasets from the ImgSal database using a standard performance metric (area under Receiver Operating Characteristic curve) as the fitness. To compensate for any bias of the optimized model towards the standard metric, we use two other scoring metrics to assess performance. Performance comparisons with eight state-of-the-art models have been presented for all three scoring metrics. Results show that the proposed GA optimized visual attention model provides better prediction performance than several state-of-the-art models of visual attention.