The effect of scene content on image quality

Device-dependent metrics attempt to predict image quality from an ‘average signal’, usually embodied on test targets. Consequently, the metrics perform well on individual ‘average looking’ scenes and test targets, but provide lower correlation with subjective assessments when working with a variety of scenes with different than ‘average signal’ characteristics. This study considers the issues of scene dependency on image quality. This study aims to quantify the change in quality with scene contents, to research the problem of scene dependency in relation to devicedependent image quality metrics and to provide a solution to it.

A novel subjective scaling method was developed in order to derive individual attribute scales, using the results from the overall image quality assessments. This was an analytical top-down approach, which does not require separate scaling of individual attributes and does not assume that the attribute is not independent from other attributes. From the measurements, interval scales were created and the effective scene dependency factor was calculated, for each attribute. Two device-dependent image quality metrics, the Effective Pictorial Information Capacity (EPIC) and the Perceived Information Capacity (PIC), were used to predict subjective image quality for a test set that varied in sharpness and noisiness. These metrics were found to be reliable predictors of image quality. However, they were not equally successful in predicting quality for different images with varying scene content.

Objective scene classification was thus considered and employed in order to deal with the problem of scene dependency in device-dependent metrics. It used objective scene descriptors, which correlated with subjective criteria on scene susceptibility. This process resulted in the development of a fully automatic classification of scenes into ‘standard’ and ‘non-standard’ groups, and the result allows the calculation of calibrated metric values for each group. The classification and metric calibration performance was quite encouraging, not only because it improved mean image quality predictions from all scenes, but also because it catered for nonstandard scenes, which originally produced low correlations. The findings indicate that the proposed automatic scene classification method has great potential for tackling the problem of scene dependency, when modelling device-dependent image quality. In addition, possible further studies of objective scene classification are discussed.