Prediction of visual attention with deep CNN on artificially degraded videos for studies of attention of patients with Dementia