Assessment of multi-exposure HDR image deghosting methods

Abstract

To avoid motion artefacts when merging multiple exposures into a high dynamic range image, a number of HDR deghosting algorithms have been proposed. However, these algorithms do not work equally well on all types of scenes, and some may even introduce additional artefacts. As the number of proposed deghosting methods is increasing rapidly, there is an immediate need to evaluate them and compare their results. Even though subjective methods of evaluation provide reliable means of testing, they are often cumbersome and need to be repeated for each new proposed method or even its slight modification. Because of that, there is a need for objective quality metrics that will provide automatic means of evaluation of HDR deghosting algorithms. In this work, we explore several computational approaches of quantitative evaluation of multiexposure HDR deghosting algorithms and demonstrate their results on five state-of-the-art algorithms. In order to perform a comprehensive evaluation, a new dataset consisting of 36 scenes has been created, where each scene provides a different challenge for a deghosting algorithm. The quality of HDR images produced by deghosting method is measured in a subjective experiment and then evaluated using objective metrics. As this paper is an extension of our conference paper, we add one more objective quality metric, UDQM, as an additional metric in the evaluation. Furthermore, analysis of objective and subjective experiments is performed and explained more extensively in this work. By testing correlation between objective metric and subjective scores, the results show that from the tested metrics, that HDR-VDP-2 is the most reliable metric for evaluating HDR deghosting algorithms. The results also show that for most of the tested scenes, Sen et al.'s deghosting method outperforms other evaluated deghosting methods. The observations based on the obtained results can be used as a vital guide in the development of new HDR deghosting algorithms, which would be robust to a variety of scenes and could produce high quality results.

Subjective and Objective Evaluation of Local Dimming Algorithms for HDR Images

“…In order to evaluate HDR images, several image quality metrics have been developed and adapted to the HDR domain. These include puPSNR [29], puSSIM [30] and HDR-VDP-2 [31], which have been adopted for comparing HDR images for a number of applications [32], [33] [35]. These metrics are capable of addressing a wide range luminance and are used widely.…”

Dataset and Metrics for Predicting Local Visible Differences

“…This set is an important source of near-threshold distortions. deghosting (12 images) contains artifacts due to HDR merging, which exposes shortcomings of popular deghosting methods [Karađuzović-Hadžiabdić et al 2017]. ibr (36 images), and cgibr (6 images) contain artifacts produced by viewinterpolation and image-based rendering methods, which come from [Adhikarla et al 2017].…”

Overview of State-of-the-Art Algorithms for Stack-Based High-Dynamic Range (HDR) Imaging

“…For example, Tursun et al [45] conducted a user study to examine HDR quality and found that the methods of Sen et al [39] and Hu et al [19] ranked first and second, respectively, with a sizeable margin over other state-of-the-art methods. More recently, Karaduzovic-Hadziabdic et al [25] also found that the method of Sen et al [39] outperformed others for most scenes. Finally, while most algorithms require that the image stack be captured only by varying the exposure time between images (changing settings like the aperture affects the depth-of-field, which makes images difficult to align or deghost), Sen et al [39] showed that their patchbased method is able to handle changes in depth-of-field automatically, thereby enabling longer exposures than could be done by simply changing the exposure time.…”