| As an important presentation method in virtual reality applications,omnidirectional images have become an important indicator of immersion in the perceptual experience.However,due to the limitations of existing technologies,distortion is inevitably introduced in the projection and transmission stages of omnidirectional images,thus degrading the quality of the images.At present,most of the deep learning-based quality assessment algorithms for no-reference omnidirectional images use the method of directly mapping the input image to a quality score,i.e.the end-to-end learning method,which can reflect the quality of the image to a certain extent,but the quality prediction ability decreases significantly when processing richer data.Based on the above problems,this paper proposes a no-reference quality assessment method for omnidirectional images.Firstly,in order to capture the omnidirectional image features comprehensively,a multi-channel convolutional neural network model is proposed,and a local distortion perception module is introduced in each backbone network to realise multi-scale features learning of the image;secondly,starting from the subjective way of human visual perception,the hypernetwork idea is introduced into the omnidirectional image quality evaluation,and the image quality assessment rules are generated adaptively;finally,the extracted features and rules are input into the quality prediction network to obtain quality prediction scores.Through experiments on OIQA,CVIQ and self-built datasets,it is shown that the proposed method in this paper is significantly better than existing mainstream methods,has good consistency with human subjective assessment,and has good generalization.The main contributions and work of this paper are as follows:(1)A novel omnidirectional image quality evaluation dataset is constructed and used in the comparison of mainstream quality evaluation algorithms and cross-dataset validation experiments.The dataset contains 60 reference omnidirectional images and720 distorted images with three types of distortion: JPEG compression,Gaussian blur and Gaussian noise,with subjective scores provided in the form of mean opinion score(MOS),taking values in the range [0,100];(2)A multi-channel convolutional neural network model is proposed for image feature learning.The network model consists of six residual networks,and the local distortion-aware module is introduced by modifying the residual structure to fully learn the distortion features of the six projected images obtained through the transformation of the cubic projection format;(3)Implementing an adaptive quality evaluation model based on a hypernetwork.The hypernetwork consists of three sets of weights and deviation branches,which enables the model to have a rule-aware function,effectively simulating the human subjective assessment process and improving the accuracy of omnidirectional image quality assessment;(4)Based on the above model,an omnidirectional image quality assessment system is developed,which contains functions such as image type recognition,image quality assessment,image quality ranking and statistical data visualization. |
