Abstract | This study aims to introduce improvements in the predictions of device-dependent image quality metrics (IQMs). A validation experiment was first carried out to test the success of such a metric, the Effective Pictorial Information Capacity (EPIC), using results from subjective tests involving 32 test scenes replicated with various degrees of sharpness and noisiness. The metric was found to be a good predictor when tested against average ratings but, as expected by device-dependent metrics, it predicted less successfully the perceived quality of individual, non-standard scenes with atypical spatial and structural content. Improvement in predictions was attempted by using a modular image quality framework and its implementation with the EPIC metric. It involves modeling a complicated set of conditions, including classifying scenes into a small number of groups. The scene classification employed for the purpose uses objective scene descriptors which correlate with subjective criteria on scene susceptibility to sharpness and noisiness. The implementation thus allows automatic grouping of scenes and calculation of the metric values. Results indicate that model predictions were improved. Most importantly, they were shown to correlate equally well with subjective quality scales of standard and non-standard scenes. The findings indicate that a device-dependent, scene-dependent image quality model can be achieved. |
---|