How does the perceived quality of compressed images depend on image content?

The impact of image compression algorithms varies significantly across image contents in a way that is challenging to predict. The ongoing trend towards richer visual content, e.g. High Dynamic Range and Wide Color Gamut, increases both the relevance and complexity of this issue. This study analyzes first grades of perceived quality of compressed images to determine in which proportion their variance is due to compression levels, image content and compression type, respectively. An ANOVA analysis on 3 HDR datasets indicates that the variance of the subjective evaluations is due for 45-62% to the compression level and for 7-10% to the image content. Secondly, we present a framework for identifying which features calculated on the source images are efficient to predict the part of image content in grades of perceived quality of compressed images. We build on traditional regression analysis by adding an adaptation of the recent Model Class Reliance approach. In an experiment on 6 published datasets of subjective quality grades of compressed images, OLS-R and KNN models predicting the grades are built using two input variables: the compression level and one feature characterizing the original content. The Empirical Model Reliance is then calculated to measure the importance of the content feature in the regression model as well as the Model Class Reliance to bound the impact of a reduced fit to the training data, i.e. indicating robustness towards generalization. Results show that traditional regression analysis alone is not robust for identifying the most relevant features and confirms that when the most useful features for SDR are SI/block contrast measures, other features characterize HDR content best, such as DR and color features (colorfulness or saturation).