Research On Blind Image Quality Assessment Of Realistic Distorted Image

With the advent of the mobile digital age,various social media applications generate hundreds of millions of images every day.Digital images fill all aspect of our lives,most of which were taken by amateur users in various non-professional environments.Unlike photos taken by professional photographers,images generated by amateur users are usually degraded in quality due to many factors such as environmental factors,hardware equipment parameters and personnel photography techniques that lead to overexposure or underexposure,low visibility,motion blur,out-of-focus,ghosting and other distortions.In addition,there is an inevitable degradation of image quality in tasks such as image acquisition,compression,transmission and storage,leading to an undesirable visual experience.On the one hand,high-quality images can improve the quality of the viewer’s experience,and on the other hand,many computer vision algorithms can benefit from them.Therefore,the image quality is a very important indicator for service providers to deliver high-quality content to users,and an extremely important metric for visible camera-based systems to filter low-quality images to avoid decision errors.With massive images being generated by millions of cameras every moment,there is an urgent need to develop an effective realistic distortion image quality assessment model to help improve the perceived quality of real images.In recent years,some progress has been made in image quality assessment based on realistic distortion.However,realistic distortion is different from artificially synthetic distortion,the factors leading to real distortion are very complex and the types of real distortion produced are also very diverse.It is extremely difficult to accurately evaluate the perceived quality of realistic distorted images.How to accurately and efficiently evaluate the perceptual quality of realistic distorted images and obtain the prediction results that are consistent with human subjective opinions is an important subject to be studied urgently in the field of image quality assessment.Exposure,hue,contrast distortion and motion blur are the most common types of realistic distortion in images taken by amateur users on a daily basis,and the most intuitive stimuli brought by these distortions to the human visual system are the stimuli of color and texture information.Moreover,deep learning methods have achieved excellent performance in many vision tasks due to their powerful image characterization capabilities,while existing deep learning methods for realistic distortion image quality assessment mix semantic learning and quality assessment in a single network,ignoring the impact of image semantic features on perceptual quality.This thesis is dedicated to solving the realistic distortion image quality assessment problem by fusing various visual features sensitive to realistic distortion.The research of this thesis consists of two main aspects as follows.(1)In this thesis proposes a realistic distortion image quality assessment model based on color and texture features.First,a given distorted image is mapped from RGB color space to HSV color space to extract luminance,hue and contrast information more effectively.For each color channel,the model extracts the first 3-dimensional color moments as the color features of the image.Second,we divide the image into four layers based on the real-valued response of the logarithmic Gabor filter by thresholding,and each layer corresponds to a region of the image with different texture characteristics,and calculate its mean subtracted contrast normalized(MSCN)coefficient in each layer separately.Again,the distribution of MSCN coefficients in each layer is fitted using generalized Gaussian distribution and asymmetric generalized Gaussian distribution,and the fitted parameters of the two generalized Gaussian distributions are constructed as the texture features of the image.Then,the two features are fused to form a 49-dimensional image feature.Finally,a support vector machine regression was used to learn the mapping relationship from the image features to the final perceptual quality score.We compare the performance of the proposed model with nine state-of-the-art general-purpose no-reference image quality assessment methods on two publicly available realistic distortion image databases,i.e.,CID2013(Camera Image Database)and BID(realistic-Blur Image Database),and the experimental results show that the proposed model achieves better performance on both databases.Also,the experimental results demonstrate that fusing color features and texture features has an important impact on evaluating the perceived quality of realistic distortion.(2)In this thesis,a realistic distorted image quality assessment model based on multi-task convolutional neural networks is proposed.The model decomposes the quality assessment task into three subtasks,i.e.,image attribute prediction subtask,image scene classification subtask,and image quality assessment subtask.The image attribute prediction subtask predicts five image attributes,i.e.,brightness,colorfulness,contrast,noisiness,and sharpness,based on the shallow features of the input image.The image scene classification subtask classifies the given input image into 9 predefined scene categories,i.e.,animal,plant,human,indoor scene,cityscape,night scene,landscape,still-life,and others.The image quality assessment subtask,on the other hand,predicts the perceptual quality score of a given image after fusing the image high-level semantic features with the output of the image scene classification subtask and the image attribute prediction subtask.We perform pre-training and end-to-end joint optimization of the three subtasks to predict the perceptual quality of realistic distorted images by fusing image attribute prediction results,image scene classification results,and low-level and high-level,global and local features of the images.We compare the proposed model with 10 advanced Blind Image Quality Assessment algorithms on the SPAQ(Smartphone Photography Attribute and Quality)database,and the proposed model achieves significant advantages and good agreement with human subjective ratings.The effectiveness of the proposed modules was also verified by ablation experiments.