Blind Image Quality Assessment With Hierarchical Feature Concatenation

In the era of big data,images have became the main information carrier in people's daily lives.However,during the imaging and transmission process,various distortions will inevitably be introduced,which will cause the degradation of image quality.As the main source of information for human subjective perception and machine vision,the quality of images directly determines the validity of the information obtained.Therefore,how to evaluate the image quality and whether the image quality meets the requirements of specific scenes needs to be solved urgently,and the solution of this problem requires the establishment of reasonable image quality assessment methods and evaluation standards.The development of objective image quality assessment methods cannot be separated from the support of the image quality assessment databases.The image quality assessment databases improve the existing methods through benchmark tests and provide a data source for the training and parameter fitting of new models.The image quality assessment databases usually follow standard procedures when they are established.Not only is the diversity of image contents limited,but artificially evaluating image quality will consume a lot of manpower and resources,resulting in existing small databases.The small-scale databases make the existing deep learning-based no-reference image quality assessment methods have various disadvantages,and existing deep learning-based methods usually have large model parameters and complex processes,which cannot meet the accuracy and real-time requirement in practical applications.In addition,image quality assessment usually uses the quality score to quantitatively analyze the degradation of image quality.In some cases,it may not be consistent with human subjective perception.Moreover,a single scalar value cannot reflect the damaged degree and which part of the image is damaged.In view of the above problems,the research contents of this article are as follows:Aiming at the small size of the existing databases,a large-scale image quality assessment database is established,and the validity of the database is further verified through experimental analysis.The database establishment process is as follows: First,in order to ensure the diversity of image contents,10340 high-definition images of typical scenes are selected from the MS COCO2014 dataset as the original images.Second,in order to ensure the diversity of distortion types,21 kinds of distortion types are simulated.Then,in order to solve the problem of artificially evaluating image quality,which is time-consuming and expensive,a hybrid full reference algorithm is used to evaluate the quality score of the distorted image.Finally,to solve the problem of inconsistent quality score scale,a non-linear model is used for normalization.This dataset greatly expands the diversity of visual contents and distortion types,solving the problem of the insufficient training data,providing support for the development of no-reference image quality assessment algorithms.Aiming at the problems of the existing deep learning-based no-reference image quality assessment methods and inspired by the hierarchical mechanism of the human visual system when processing visual signals,a cascaded hierarchical degradation fusion network is proposed,which is composed of cascaded feature extraction network,feature downsampling network and feature regression network.It not only learns end-to-end,but also fully considers the hierarchical characteristics of image quality degradation.Due to the ultra-lightweight network design,it is further integrated into the micro-processing system to achieve a noreference quality evaluation system for visible light images,which can accurately and quickly evaluate large amount of images.Aiming at the problem that the image quality is only evaluated by a single scalar value,a no-reference image quality evaluation method oriented to hierarchical semantic degradation is proposed.Firstly,the visual semantic degradation is described.The description standard of the image uses local detail semantics,regional contour semantics,and global concept semantics as the hierarchical semantics of the image.The degree of semantic degradation of the image at different levels is defined as slight distortion,significant distortion,and severe distortion.Then,a multi-level semantic quality degradation measurement model is designed,and the qualitative description of image quality degradation using text is realized.