Research On No-Reference Quality Assessment Of HDR Images Based On Convolutional Neural Network

With the rapid development of High Dynamic Range(HDR)technology in the field of Ultra High-Definition,HDR images have been widely utilized in various scenes.Currently,the quality assessment of HDR image tone mapping algorithms based on user visual experience and image quality assessment based on convolutional neural network have become a popular research trend.Since HDR images can show rich luminance and chrominance information,but using different tone mapping algorithms will result in different tone mapped images.Therefore,a reliable quality evaluation algorithm is needed to provide realistic data for the selection and improvement of the tone-mapping operator,thus improving the image quality and providing a better experience for practical applications.In this paper,as a result,it is proposed a HDR image no-reference quality evaluation method based on convolutional neural network focusing on HDR images after tone mapping,in order to effectively tackle the problems of multiple sample selection constraints,tedious evaluation process and high comprehensive cost in the existing HDR image quality evaluation.The main contributions and research results of this study are summarized as follows:(1)A no-reference assessment algorithm for HDR images based on conventional convolutional neural networks is proposed.Through extracting luminance features and chromaticity features from YUV color space respectively,where the U and V chromaticity channels are pre-processed,weights and thresholds are assigned for the pixels of different channels,subsequently feature fusion is achieved by convolutional stitching.The convolutional neural network converts the pixel values of the input images into the output feature vectors,which are converted into predictions by a linear layer,and the loss function is calculated with the ground truth of MOS in the datasets,updating the network weight parameters iteratively by backward.The weight parameters of the network are finally determined after several experiments,thus achieving end-to-end no-reference quality evaluation.The method was experimented on the ESPL-LIVE HDR datasets and compared with existing evaluation algorithms.The results showed that the algorithm achieved higher accuracy.(2)A lightweight,no-reference quality assessment algorithm model is proposed,which effectively addresses the problems of traditional convolutional neural networks in terms of the number of parameters and computational resources and time through simplifying the number of model parameters and takes into account the speed and accuracy,enabling low latency and high response and increasing the accuracy of the evaluation model.The experimental results show that our method achieves efficient end-to-end HDR image no-reference evaluation,with the subjective and objective performance evaluation scores of 0.8013,0.8322 for PLCC and SROCC,respectively.(3)In this paper,an exclusive image dataset was constructed by acquiring multi-exposure image sequences,using hue mapping algorithm,multi-exposure fusion algorithm and software post-processing,and completed the display chromaticity characterization in a darkroom laboratory scene to verify the consistency of the effect of the subjective and objective HDR image no-reference quality assessment algorithm based on the designed psychophysical experimental scheme,which scientifically standardized the experimental scene and achieved the overall consistency between the predicted value of the algorithm and the subjective assessment result,which has excellent practical application value.