Deep hierarchical network for full-reference omnidirectional image quality assessment

Youzhi Zhang, Lifei Wan, Yiren Chen, Xiaofei Zhou, Xiaolin Zhang, Deyang Liu

Abstract: Virtual Reality(VR) can provide users immersive experience, which is considered as an innovative teaching tool for education. The rapid advancement of VR technology has brought a substantial increase in the availability of omnidirectional images on the internet, which highlights the urgent need for omnidirectional image quality assessment. Assessing the quality of omnidirectional images poses a significant challenge due to their spherical representation. Unlike traditional 2D images, omnidirectional images record \({360^{\circ }\times 180^{\circ }}\) panoramic content with extremely high resolution, which brings more difficulties in handling the complicated scene content during quality assessment procedure. To relieve this problem, in this paper, we propose to learn a deep hierarchical network for full-reference omnidirectional image quality assessment. Motivated by the characteristic of the human visual system (HVS), the proposed method utilizes intermediate layers of a convolutional neural network (CNN) to extract multi-level information. These layers capture different semantic representations, encompassing both low-level and high-level details. The shallow layers capture low-level information, such as edges and corners, which are highly sensitive to quality degradation. To effectively model the impact of distortions, we further introduce a local distortion-aware (LDA) block which can accurately capture the degradation of image quality. The proposed method can achieve promising performances in experiments conducted on two publicly available databases, which demonstrates the effectiveness of our method.

External IDs:dblp:journals/mms/ZhangWCZZL25