Camera based Display Image Quality Assessment
MetadataShow full item record
This thesis presents the outcomes of research carried out by the PhD candidate Ping Zhao during 2012 to 2015 in Gjøvik University College. The underlying research was a part of the HyPerCept project, in the program of Strategic Projects for University Colleges, which was funded by The Research Council of Norway. The research was engaged under the supervision of Professor Jon Yngve Hardeberg and co-supervision of Associate Professor Marius Pedersen, from The Norwegian Colour and Visual Computing Laboratory, in the Faculty of Computer Science and Media Technology of Gjøvik University College; as well as the co-supervision of Associate Professor Jean-Baptiste Thomas, from The Laboratoire Electronique, Informatique et Image, in the Faculty of Computer Science of Universit´e de Bourgogne. The main goal of this research was to develop a fast and an inexpensive camera based display image quality assessment framework. Due to the limited time frame, we decided to focus only on projection displays with static images displayed on them. However, the proposed methods were not limited to projection displays, and they were expected to work with other types of displays, such as desktop monitors, laptop screens, smart phone screens, etc., with limited modifications. The primary contributions from this research can be summarized as follows: 1. We proposed a camera based display image quality assessment framework, which was originally designed for projection displays but it can be used for other types of displays with limited modifications. 2. We proposed a method to calibrate the camera in order to eliminate unwanted vignetting artifact, which is mainly introduced by the camera lens. 3. We proposed a method to optimize the camera’s exposure with respect to the measured luminance of incident light, so that after the calibration all camera sensors share a common linear response region. 4. We proposed a marker-less and view-independent method to register one captured image with its original at a sub-pixel level, so that we can incorporate existing full reference image quality metrics without modifying them. 5. We identified spatial uniformity, contrast and sharpness as the most important image quality attributes for projection displays, and we used the proposed framework to evaluate the prediction performance of the state-of-the-art image quality metrics regarding these attributes. The proposed image quality assessment framework is the core contribution of this research. Comparing to conventional image quality assessment approaches, which were largely based on the measurements of colorimeter or spectroradiometer, using camera as the acquisition device has the advantages of quickly recording all displayed pixels in one shot, relatively inexpensive to purchase the instrument. Therefore, the consumption of time and resources for image quality assessment can be largely reduced. We proposed a method to calibrate the camera in order to eliminate unwanted vignetting artifact primarily introduced by the camera lens. We used a hazy sky as a closely uniform light source, and the vignetting mask was generated with respect to the median sensor responses over i only a few rotated shots of the same spot on the sky. We also proposed a method to quickly determine whether all camera sensors were sharing a common linear response region. In order to incorporate existing full reference image quality metrics without modifying them, an accurate registration of pairs of pixels between one captured image and its original is required. We proposed a marker-less and view-independent image registration method to solve this problem. The experimental results proved that the proposed method worked well in the viewing conditions with a low ambient light. We further identified spatial uniformity, contrast and sharpness as the most important image quality attributes for projection displays. Subsequently, we used the developed framework to objectively evaluate the prediction performance of the state-of-art image quality metrics regarding these attributes in a robust manner. In this process, the metrics were benchmarked with respect to the correlations between the prediction results and the perceptual ratings collected from subjective experiments. The analysis of the experimental results indicated that our proposed methods were effective and efficient. Subjective experiment is an essential component for image quality assessment; however it can be time and resource consuming, especially in the cases that additional image distortion levels are required to extend the existing subjective experimental results. For this reason, we investigated the possibility of extending subjective experiments with baseline adjustment method, and we found that the method could work well if appropriate strategies were applied. The underlying strategies referred to the best distortion levels to be included in the baseline, as well as the number of them.