[go: up one dir, main page]

Czech Technical University in Prague
Faculty of Electrical Engineering


Perceptually Based Image Quality Assessment and Image Transformations

PhD Thesis


Author: Martin Čadík, cadikm@fel.cvut.cz
Tutor: Prof. Ing. Pavel Slavík, CSc.

Oponents:
Karol Myszkowski
Mateu Sbert
Pavel Zemčík

Date: January 2008

Download:
Martin Čadík's [PhD thesis (web optimized, compressed images, 10MB, PDF)]
[PhD thesis statement (1.6MB, PDF)]
[bibTeX entry (bib)]


ABSTRACT

Computational approaches mimicking perceptual properties of the human visual system (HVS) have been successfully applied in various fields of computer graphics and digital imaging. The knowledge about the HVS is continually developing, however there are still many open questions and hypotheses. Naturally, we are far from having an accurate computational model of the HVS. It is therefore of particular importance and necessity to conduct experimental subjective analyses of the methods that incorporate HVS models to validate and evaluate them properly. Besides exposing strengths and weaknesses of inquired methods, the experimental evaluations on human subjects also attain a deeper knowledge of the examined field which can advance current state of the art and which can furthermore result in proposals of new approaches.

The fields that traditionally utilize computational models of the HVS and which are concerned in this thesis are a) image quality assessment, b) high dynamic range tone mapping, and c) color image to grayscale conversions.

a) Image quality assessment metrics aim to predict the difference between images as perceived by human subjects. We present results of an experimental subjective evaluation of two principal approaches to image quality assessment (traditional error sensitivity based approach, and structural similarity based approach). The results show that the structural similarity outperforms the traditional approach for involved input stimuli.

b) The field of tone mapping has witnessed a lot of research effort to solve the problem of displaying images with high dynamic range (HDR) of luminance on ordinary output devices. We present a study about the effect of image attributes (contrast, brightness, etc.) in the HDR tone mapping. Furthermore, we present experimental subjective evaluations of global and local tone mapping approaches. Our results imply that proper global part of a tone mapping method is essential to obtain good perceptual results for typical real world scenes.

Motivated by implications of conducted studies, we propose a novel simple yet powerful general hybrid approach to tone mapping. In our approach, we apply a global tone mapping method first to reproduce overall image attributes correctly and we construct an enhancement map to guide a local operator to the critical areas that deserve an enhancement. The new approach is general and can be easily tailored to miscellaneous goals of tone mapping. An implementation of the proposed approach produces plausible results, it is easy to implement, and fast to compute.

c) Simple color image to grayscale conversions potentially disrupt chromatic informations contained in color images, and existing advanced conversions are eminently computationally intensive. We propose novel perceptually convincing local color to grayscale conversion techniques that operate in gradient domain. The novel methods produce perceptually justifiable results, and they have linear complexity in the number of pixels, which makes them suitable for high-resolution images.

Keywords:

human perception, human visual system, experimental evaluation, image quality assessment, image comparison, HDR, tone mapping, color to gray


ADDITIONAL MATERIALS

Evaluation of Image Quality Metrics Perceptual evaluation of image quality metrics
Experimental Data and Input images
Evaluation of Tone Mapping Operators Perceptual evaluation of tone mapping methods
Experimental Data and Input images
Hybrid tone mapping Perception Motivated Hybrid Approach to Tone Mapping
Color-to-gray conversion Conversion of color images to grayscale



Last changes: 05/2008 - by Martin Čadík     Valid XHTML 1.0 Transitional