| Abstract | This thesis is concerned with the measurement of spatial imaging performance and the modelling of spatial image quality in digital capturing systems. Spatial imaging performance and image quality relate to the objective and subjective reproduction of luminance contrast signals by the system, respectively; they are critical to overall perceived image quality.
The Modulation Transfer Function (MTF) and Noise Power Spectrum (NPS) describe the signal (contrast) transfer and noise characteristics of a system, respectively, with respect to spatial frequency. They are both, strictly speaking, only applicable to linear systems since they are founded upon linear system theory. Many contemporary capture systems use adaptive image signal processing, such as denoising and sharpening, to optimise output image quality. These non-linear processes change their behaviour according to characteristics of the input signal (i.e. the scene being captured). This behaviour renders system performance “scene-dependent” and difficult to measure accurately. The MTF and NPS are traditionally measured from test charts containing suitable predefined signals (e.g. edges, sinusoidal exposures, noise or uniform luminance patches). These signals trigger adaptive processes at uncharacteristic levels since they are unrepresentative of natural scene content. Thus, for systems using adaptive processes, the resultant MTFs and NPSs are not representative of performance “in the field” (i.e. capturing real scenes).
Spatial image quality metrics for capturing systems aim to predict the relationship between MTF and NPS measurements and subjective ratings of image quality. They cascade both measures with contrast sensitivity functions that describe human visual sensitivity with respect to spatial frequency. The most recent metrics designed for adaptive systems use MTFs measured using the dead leaves test chart that is more representative of natural scene content than the abovementioned test charts. This marks a step toward modelling image quality with respect to real scene signals.
This thesis presents novel scene-and-process-dependent MTFs (SPD-MTF) and NPSs (SPDNPS). They are measured from imaged pictorial scene (or dead leaves target) signals to account for system scene-dependency. Further, a number of spatial image quality metrics are revised to account for capture system and visual scene-dependency. Their MTF and NPS parameters were substituted for SPD-MTFs and SPD-NPSs. Likewise, their standard visual functions were substituted for contextual detection (cCSF) or discrimination (cVPF) functions. In addition, two novel spatial image quality metrics are presented (the log Noise Equivalent Quanta (NEQ) and Visual log NEQ) that implement SPD-MTFs and SPD-NPSs.
The metrics, SPD-MTFs and SPD-NPSs were validated by analysing measurements from simulated image capture pipelines that applied either linear or adaptive image signal processing. The SPD-NPS measures displayed little evidence of measurement error, and the metrics performed most accurately when they used SPD-NPSs measured from images of scenes. The benefit of deriving SPD-MTFs from images of scenes was traded-off, however, against measurement bias. Most metrics performed most accurately with SPD-MTFs derived from dead leaves signals. Implementing the cCSF or cVPF did not increase metric accuracy.
The log NEQ and Visual log NEQ metrics proposed in this thesis were highly competitive, outperforming metrics of the same genre. They were also more consistent than the IEEE P1858 Camera Phone Image Quality (CPIQ) metric when their input parameters were modified. The advantages and limitations of all performance measures and metrics were discussed, as well as their practical implementation and relevant applications. |
|---|