Predicting visual difference maps for computer‐generated images by integrating human visual system model and deep learning

The quality of images generated by computer graphics rendering algorithms is mainly affected visible distortion at some pixel locations. Image assessment (IQA) metrics are commonly utilized to assess the rendered images, but their results a global difference value, which does not provide pixel-wise differences optimize renderings. In contrast, visibility models including visual perception and deep learning can calculate between distorted reference images. However, they either only applied single type or seriously dependent on datasets. To this end, authors propose novel model, dubbed Human Visual Perception Deep Learning Difference Metric (HPDL-IDM), combines System (HVS) model learning. HPDL-IDM primarily consists two modules: (i) feature calculation module, calculates maps various kinds features extracted from image according characteristics human eyes concatenates them, (ii) utilizes neural network encoder–decoder structure train LocvisVC VisTexRes datasets whose input output these concatenated final map respectively. Additionally, pool into value 0 1 apply many processing tasks related (IQMs). Experimental show that HPDL-IDM's generalization capacity accuracy improved large margin compared other models.