Thesis title: Study of the subjective effects of blur on the vision of natural images: an abstract, physical parametric model for Image Quality Assessment.
Looking at a link between blur and visual discomfort, in the present thesis, blur is viewed as a cause of a cognitive loss, and the discomfort as the immediate consequence of this loss. Among the basic cognitive functions of the Human Visual System (HVS), detection, recognition, and coarse localization functions are strongly conditioned by the individual experience. Conversely, it seems plausible that the fine localization function is committed to stabler and inter-subjective functions of the HVS. After a preliminary discussion of the operators and the ML model used (Part II), the approach presented in Part III of this thesis starts from postulating that, in the absence of vision problems, the HVS performs the fine localization of the observed objects with the best accuracy allowed by its physical macro-structure. This is a fundamental assumption, because it is known from the estimation theory that the maximum accuracy attainable when measuring the fine position of patterns in background noise is obtained by the Fisher Information about positional parameters. In fact, the Fisher Information inverse yields the minimum estimation variance. The proposed approach is based on an abstract, functional model of the Receptive Fields (RF) of the HVS, referred to as Virtual Receptive Field (VRF) and it is tuned to statistical features of natural scenes. It is a complex-valued operator, orientation-selective both in the space domain and in the spatial frequency domain. The role of the VRF model is to extract the Positional Fisher Information (PFI) as a measure of the pattern localizability loss.
In the Image Quality Assessment (IQA) Full Reference (FR) environment, subjective assessments refer to the retinal image and lead to the MOS/DMOS values (Difference of Mean Opinion Score). The quality calculated by the IQA metrics is objective and refers to the image reproduced on the display. A parametric scoring function maps these metrics onto the MOS/DMOS values and depends critically on the Viewing Distance (VD) of the subject from the monitor in which the image is reproduced. When objective quality estimates for different VDs are required, as in the case of auditoria, cinemas, classrooms, a re-training procedure must be repeated for each different VDs. In the final part of this thesis (Part IV), the problem of VD is dealt with from a theoretical point of view and a model of the scoring function is defined for the case of blurred images where image degradation substantially depends on the VD. Starting from a Fisher Information loss model applied to the Gaussian distortion case in natural images, we see that the VD is estimated from the data themselves. Several maps are given with the aim of obtaining a DMOS prediction at different distances starting from the data available for a specific distance, without performing new experiments. Moreover, the theoretical results are verified on some most popular IQA FR methods and the problem of VD correction is generalized to the other distortions.
Finally, the impact of isolated, long, strong, unidirectional edges on early vision is shown. As for the VD correction, an a-priori linear estimator is presented. It does not require rectification through a re-training procedure. Useful maps for detecting the position and the intensity of the PFI losses in an image are given, and the isoluminance colors allow to highlight strong and isolated edges, maintaining a constant intensity at the same edge level. We have an easy visual feedback on the images themselves to see where the greatest loss of information and the greatest discomfort due to blur are.