Research On The Method Of Image Quality Assessment For Cross-screen Visual Content

Nowadays,intelligent visual products are deeply embedded in people’s lives,whether it is video conferencing,games or new media,and a large amount of cross-screen visual content is disseminated in communication and network systems every day,for which image quality assessment becomes necessary to meet users’ perceptual quality-of-experience requirements.Image quality assessment provides guidance for downstream tasks by establishing certain models for automated quality assessment of images.Most of the existing image quality methods are based on natural images and cannot be directly applied to crossscreen visual content images with good results.Cross-screen visual content images,i.e.images on a variety of different screens(mobile phones,computers,projectors,etc.),are usually generated by computers and generally have many different characteristics from images of natural scenes captured by cameras.These feature differences can pose significant challenges to their quality evaluation.In recent years,the task of screen quality evaluation has attracted a lot of attention from researchers and a lot of work has been carried out to address the problems faced by screen content quality evaluation.On the one hand,the existing publicly labeled datasets for screen content image quality assessment are scarce due to high construction costs,resulting in learning-based methods that cannot be pre-trained on a large scale and are ineffective? on the other hand,the types of image distortion are complex and variable,and unknown distortions can occur in real scenarios requiring rapid model adaptation.In addition,the distortion of images in practical application scenarios is dynamically changing,and it is also worthwhile to conduct research on how to perform dynamic quality evaluation to meet the stability.In this thesis,we address the strengths,weaknesses,and limitations of existing methods and conduct research on the image quality assessment problem of cross-screen visual content,starting from three problems: scarcity of data sets,complexity of distortion types,and stability of dynamic distortion,respectively.The specific research includes the following three aspects.(1)To address the problem of lacking large-scale labeled datasets for cross-screen visual content image quality assessment,a contrast learning-based screen content quality evaluation method is proposed,inspired by unsupervised contrast learning methods.Two datasets containing natural and screen content images are firstly used for self-supervised pre-training with the goal of an auxiliary task,and then fine-tuned on the publicly available screen content image quality assessment dataset.For the extracted quality-related features,a contrast learning operation is performed with ”positive samples” at a closer distance and”negative samples” at a farther distance,so that the self-supervised learning idea is used to better learn the quality features based on the inability to accurately grasp the quality features.In order to achieve better learning ability of the model.(2)To address the problem of rich distortion types and large content span of crossscreen visual content images,a quality evaluation method using meta-learning is proposed from the perspective of rapid model adaptation to new distortion and image content.The cross-content image quality assessment dataset is used as the support set and query set,and the model learns the commonality of image contents as well as distortions through two stages of meta-learning to improve the generalization performance,and fine-tunes the publicly available screen content image quality assessment dataset to adapt to unknown distortions and contents.(3)To address the problem of dynamic distortion of projectors,a lightweight dynamicaware quality ranking-based autofocus method for projector autofocus is designed and implemented from the perspective of dynamic perception.The model adopts the idea of ranking and takes image pairs as input.Based on the extracted feature maps,the model uses focal distance to generate class activation maps for better learning of quality-related features.