Research On Objective Quality Evaluation Of Multi-Exposure Fusion Images Based On Visual Perception

With the development of digital media and network technology,people are stricter for visual experience,and traditional imaging systems can hardly satisfy the requirements of the visual experience.However,limited by the image acquisition equipment,the camera sensors span a limited dynamic range,which is much smaller than the dynamic range of luminance in the real world.This leads to lose many details in over-exposed and under-exposed areas of the images.As a result,high dynamic range(HDR)imaging technology has emerged,which fuses multiple low dynamic range images of the same scene at different exposure levels into a single detail-rich high quality fused image.This technique makes up the lack of hardware devices and has become a popular research in academia and industry.In recent years,numerous multi-exposure fusion(MEF)researches have been proposed.For the same scene,the quality of fused images generated by different MEF algorithms varies greatly.Therefore,how to accurately and efficiently evaluate the perceptual quality of fused images has become an important research in the field of image fusion technology,which can promote MEF researches forward in a right direction.This paper focuses on studying special distortion types of MEF images,and extracting effective visual features and constructing image quality assessment(IQA)models to accurately predict the quality of MEF images.The detail research content is as follows.(1)The MEF algorithms require the source image sequences to be perfectly aligned.The small displacement between image sequences may cause distortion in the fused image,which is called ghosting artifacts.Many researchers have worked on MEF algorithms to suppress ghosting artifacts.However,research on the quality assessment of MEF images is limited.Therefore,this paper proposes a MEF image quality evaluation algorithm based on superpixel segmentation.First,with the help of superpixels,we divide the fused images into large-and small-changed regions using the structural inconsistency map between the fused images and each source exposure image.Subsequently,the quality maps of large-and small-changed regions are computed based on Laplacian pyramid.Finally,considering that the human visual system focuses more attention on the regions with much complex changes,we propose an information theory induced adaptive pooling strategy to calculate the quality score of the fused image.A large number of comparative experiments demonstrate that the proposed IQA model outperforms the art-of-state IQA models.(2)Convolutional neural networks(CNN)have been successfully applied to many image processing and computer vision tasks,such as image recognition,semantic segmentation and face recognition and other computer vision tasks.Researchers have also tried to introduce CNN into IQA model by combining the features of the human visual system.In this paper,we propose a no-reference IQA model for MEF images based on CNN.The network contains three parts: global feature extraction,local feature extraction and quality prediction.The global feature extraction sub-network mainly extracts the overall image content of the fused image.Since there is local ghost artifacts in the fused image,we apply the multi-scale local distortion perception module to extract local features of the fused image.Finally,the global features and local features are connected together and fed into multi-layer perceptron to obtain the quality score.The experimental results show that the proposed network has excellent performance,and we demonstrate the potential application for parameter tuning of MEF algorithms.In summary,the objective IQA models for MEF images proposed in this paper achieve promising performance,which has certain research significance and practical application value for MEF algorithms.