No-reference Image Quality Assessment And Its Application In Image Enhancement

With the popularity of mobile Internet and mobile devices,images have become one of the most important ways of information interaction in people's daily life.In the fields of defense military,industrial medical and urban construction,there are also a variety of image data,and high-quality images are more conducive to extract useful information from them.However,the commonly acquired images will cause a variety of complex distortions in the process of acquisition,transmission and storage,which seriously affect people's understanding of image information and the pursuit of aesthetics.This topic is derived from the Youth Science Fund project “Blind Image Quality Evaluation Based on Deep Convolutional Neural Networks”,which aims to improve the subjective and objective consistency of multi-type synthetic distortion and real scene distortion image quality assessment.As people's demand for photography increases,improving the image quality of mobile terminals has important practical value.This paper will also explore the use of image quality assessment to optimize the mobile-end image enhancement network model.The specific paper studies the following three parts:The first part proposes a no-reference image quality assessment method for color space statistical joint texture features.Firstly,the statistical characteristics of multi-scale MSCN coefficients and their weighted LBP histogram features are extracted.Then the statistical parameters of red-green channel and blue-yellow channel coefficient distribution in color space are extracted.Finally,the image quality assessment regression model is established by SVR.The statistical features are based on the overall structure.The LBP histogram features start from the local texture,and the joint statistics and texture features have stronger perceptual image quality.The experimental results show that the algorithm has good subjective and objective consistency and generalization on the standard synthetic distortion database,and the algorithm complexity is also low.The second part proposes a no-reference image quality assessment method based on ranking learning.The method uses the VGG16 as the backbone network,and firstly uses the unlabeled data set Waterloo in the real scene to generate images of different distortion types and distortion levels.Then the training of the model is divided into two stages.The first stage uses the image with relative quality as the training set,and uses the list ranking loss function to optimize the ranked learning model.The second stage uses the IQA database to fine tune the network model,and the network output is the histogram distribution of the quality score,and finally update the model parameters by optimizing the EMD distance loss.The experimental results show that the proposed algorithm can better evaluate the image quality of complex synthetic distortion and real scene distortion,and verify that the network after ranking learning can better perceive image quality.In the third part,aiming at the specific application of image quality assessment,a mobile image enhancement method guided by image quality assessment is proposed.Firstly,the U-Net network model RSGUNet with range scaling and global features is introduced,and the loss function used in training the image enhancement task with the network as an image generator is introduced in detail.Then based on the RSGUNet network model,the no-reference image quality assessment model based on ranking learning is used to extract the subjective representation loss and subjective score loss of the enhanced image.These two losses guide the training of the image generation network and improve the subjective effect of the algorithm.The experimental results show that the improved network structure and loss function can not only improve the PSNR and SSIM objective indicators,but also improve the subjective effect of the human eye.This also shows that the image quality assessment has practical application value in the mobile terminal image enhancement task.