Objective and subjective quality assessment of 360-degree images
Doctoral thesis
Permanent lenke
https://hdl.handle.net/11250/3045020Utgivelsesdato
2023Metadata
Vis full innførselSamlinger
Sammendrag
360-degree images, a.k.a. omnidirectional images, are in the center of immersive media. With the increase in demands of the latter, mainly thanks to the offered interactive and immersive experience, it is paramount to provide good quality of experience (QoE). This QoE is significantly impacted by the quality of the content. Like any type of visual signal, 360-degree images go through a sequence of processes including encoding, transmission, decoding, and rendering. Each of these processes has the potential to introduce distortions to the content. To improve the QoE, image quality assessment (IQA) is one of the strategies to be followed. This thesis addresses the quality evaluation of 360-degree images from the objective and subjective perspectives. By focusing on the influence of Head Mounted Displays (HMDs) on the perceived quality of 360-degree images, a psycho-visual study is designed and carried out using four different devices. For this purpose, a 360-degree image datasets is created and a panel of observers is involved. The impact of HMDs on the quality ratings is identified and highlighted as an important factor to consider when conducting subjective experiments for 360-degree images. From the objective perspective, we first comprehensively benchmarked several convolutional neural network (CNN) models under various configurations. Then, the processing chain of CNN-based 360-IQA is improved at different scales, from input sampling and representation to aggregating quality scores. Based on the observations of the above studies as well as the benchmark, two 360-IQA models based on CNNs are proposed to accurately predict the quality of 360-degree images. The obtained observations and conclusions from the various contributions shall bring insights for assessing the quality of 360-degree images.