Merging of MOS of Large Image Databases for No-reference Image Visual Quality Assessment

Abstract

For training of no-reference image visual quality metrics large specialized image databases are used. For images of the databases mean opinion scores (MOS) are experimentally obtained collecting judgments of many observers. MOS of a given image reflects an averaged human perception of visual quality of the image. Each database has its own unknown scale of MOS values depending on unique content of the database. For training of no-reference metrics based on convolutional networks usually only one selected database is used, because all MOS values on input of training loss function should be in the same scale. In this paper, a simple and effective method of merging of several large databases into one database with transforming of their MOS into one scale is proposed. Accuracy of the proposed method is analyzed. Merged MOS is used for practical training of no-reference metric. Better effectiveness of the training is shown in comparative analysis.