Effects of luminance, color, and spatial frequency variations on perceived image quality
MetadataShow full item record
The primary objective of this dissertation was to investigate the effects of varying levels of luminance, color, and spatial frequency content on the perceived image quality of a soft-copy color image. A secondary objective was to test the robustness of selected image quality metrics (MTFA, SQRI, and ICS) to the color variations as measured by the change in correlations between the perceived quality ratings and the values of the image quality metrics. To accomplish these objectives, a color image was selected and its luminance, color, and spatial frequency components were attenuated systematically using image processing software. With the manipulated images, an experiment was conducted in which subjects were asked to rate, on a 0.0 - 9.0 continuous scale, the perceived quality of a displayed image in comparison to the original image. Results of the statistical analysis of the collected data were characterized by the highly significant main effects and interaction effects. However, the magnitudes of the interactions were small.
The effect of the luminance component on perceived quality was found to be dominant and consistent across all the levels of the other two variables. As the luminance increased, the perceived quality increased at a decreasing rate. The luminance main effect was modeled well (R2 = 0.9968) by the second-order polynomial of the luminance attenuation level, or, equivalently, by the relative amount of the luminance contained in the image. The range of variation of perceived quality produced by the six luminance levels was about five units on a 0.0 - 9.0 continuous scale. It was concluded that perceived quality of the color image was determined primarily by the luminance component of the image.
The effect of color on perceived quality was found to be smaller than expected. The range of variation in perceived quality produced by the six color levels was only a little over one unit on a 0.0 - 9.0 continuous scale. Perceived qualities increased at a decreasing rate as the level of color increased. However, the slope of the curve representing the color effect was smaller than that of the luminance effect The main effect of color was modeled well (R2 = 0.9972) by the second-order polynomial of the color attenuation level, or, equivalently, by the relative amount of color contained in the image. Based on the findings of the color effect, two different roles of color in image perception are suggested. At extremely low luminance, color acts primarily as a facilitator of the luminance by providing more cues on the content of the image. At sufficiently high luminance, the increased perceived quality stems from the aesthetic characteristics of the color.
Both highpass and lowpass filtering, on the average, caused about 1.5 units of degradation as compared to the unfiltered image in perceived image quality on a 0.0 - 9.0 continuous scale. The perceived quality of the unfiltered image was greater than that of the filtered images across all the levels of luminance and color attenuation except at a low luminance level. There was no significant difference between the perceived qualities of the highpass and lowpass filtered images.
The R2 of the second-order polynomial for image qUality metrics (MTFA, SQRI, and ICS) and the mean perceived qualities did not vary across the color variations in the image manipulations. That is, these image quality metrics were robust to the color variations when the relationship between the quality metric values and the actual perceived qualities was represented by the second-order polynomial. However, with the first-order model, the R2 increased as the color level increased. The SQRI yielded higher R2 values than did the MTFA and ICS metrics when the first-order model was used. Also, the range of variation of R2 for the SQRI was smaller than that for the other two metrics. Therefore, it appears that the robustness of an image quality metric to the color variation is affected by the degree of non-linearity correction in the metric if the robustness is tested in the context of the straight-line relationship.
- Doctoral Dissertations