A Hypernetwork-Based Method for Omnidirectional Image Quality Assessment
Jie Liu, Jinhong Li · 2023
As the primary carrier of virtual reality, omnidirectional images have become a crucial metric for evaluating immersion in perceptual experience. Due to existing technological limitations, omnidirectional images are susceptible to various types of distortions during stages such as creation and transmission, which can ultimately impact the user experience. The investigation of omnidirectional image quality assessment (IQA) carries significant academic and practical significance, as it serves as a guide for enhancing the quality of images. In this paper, we first establish a multi-scene omnidirectional IQA (MOIQA) dataset. Then, we propose a no-reference omnidirectional IQA method based on hypernetwork (OIQA-Hyper). OIQA-Hyper consists of three parts: feature extractor network, hypernetwork, and quality prediction network. In terms of feature learning, we propose an improved multi-channel convolutional neural network (CNN) by introducing local feature extractor modules. The hypernetwork has the ability to adaptively generate assessment rules, which not only enhances the predictive performance of the model, but also conforms to the human subjective perception mode. The experimental results shows that OIQA-Hyper possesses excellent quality prediction capability.