Quality Assessment Of Tone-Mapped Image Using Deep Neural Network

To ensure that high dynamic range images(HDRI)can be reproduced on traditional displays,and to ensure good visual perception quality,scholars have proposed a large number of tone-mapped operators(TMOs)recently.With different TMOs,various distortions are generated in the tone-mapped images(TMIs),and the visual perception quality deteriorates.Hence,it is crucial to assess the quality of TMIs.This thesis takes TMIs as the research object,and conducts the following research work for the Blind Image Quality Assessment(BIQA).(1)The existing deep learning-based BIQA methods for TMIs are two-stage tasks of feature extraction and quality prediction that cannot optimize the whole process together,this paper proposes a BIQA method based on multi-scale visual features under the framework of deep neural network for TMIs.First,inspired by the hierarchical perception mechanism of the human visual system,the model combines with the image hierarchical degradation,and extracts and fuses multi-scale features for quality prediction.Second,the method jointly optimizes the process of feature extraction,multi-scale feature fusion and quality prediction in an end-to-end manner.It is based on data-driven,and can automatically capture statistical information in the data.Experiments show that the SROCC values of the algorithm and the human subjective evaluation reach 0.8915 and0.8245 on the TMID and ESPL-LIVE HDR datasets,respectively.(2)In the process of HDRI mapping to traditional displays,the distortion such as halo phenomenon will occur,this paper proposes a blind image quality assessment algorithm for tone-mapped image based on gradient feature enhancement.The algorithm combines the extraction of human primary visual perception features with a deep learning framework,and uses a two-stream neural network model to extract multi-scale content features and primary visual features from the original distorted image and gradient image,respectively,and merge them into hybrid visual features to capture the local structural changes in the TMI to better represent the image content.The improved TMIQA method can extract richer image perception features.Experiments show that the model is effective.The SROCC values on the TMID and ESPL-LIVE HDR datasets reach 0.9176 and 0.9357,respectively,and the accuracy has been further improved.(3)In view of the large amount of parameters and high engineering complexity of the existing BIQA model based on deep learning,an efficient and simple optimization network is proposed for the blind image quality assessment algorithm of tone-mapped.The algorithm makes full use of multi-scale features,and modulates multi-scale features through feature modulation module.In addition,the gradient domain features of the image are used to reflect the degree of visual degradation of the underlying TMI.Experiments show that the proposed model parameters are less than the existing deep learning-based IQA method,and it has achieved better results on two public tone-mapped image datasets.