Image Quality Assessment Based On Visual Perception Modeling And Representation Learning

Digital images are playing an increasingly important role in our daily life and work,with the growing popularity of Internet services, mobile devices, and mass storage de-vices. However, digital images unavoidably suffer some distortions during: acquisition,compression, transmission, and storage. Such distortions may degrade the viewing ex-perience, and also deteriorate the performance of image processing and understanding algorithms. With the explosive growth of digital visual data, it is of increasingly signifi-cance and value to predict the quality of images. The purpose of objective image quality assessment (IQA) is to automatically predict the perceived image that is consistent with human. In this dissertation, we are focusing on no-reference (NR) IQA which is aimed at measuring the visual quality of a distorted image without any information of its high-quality reference image. The main contributions and innovations of this dissertation are as follows:(1) This dissertation proposes a visual perception based NR IQA model. Human visual system (HVS) perceives visual signals with a so-called internal generative mech-anism (IGM). According to the IGM, we decompose the input image into an orderly part and a disorderly one. We extract the gradient magnitude (GM) map and the Laplacian of Gaussian (LOG) response map from the orderly part, and integrate the two maps into joint features. For the disorderly one, local binary pattern (LBP) distribution histograms are extracted. Those two groups of features are proved by experiments to be comple-mentary for predicting image quality. So, they are concatenated for further regression model training. Experiments on three widely used IQA databases show the promising performance of the new model.(2) In order to tackle with the authentically distorted IQA problem which is more challenging, this dissertation proposes a novel and effective NR-IQA framework based on convolutional neural network (CNN). The CNN model is trained with image patches.Those image patches are not simply assigned with the same quality label of the corre-sponding whole image. Instead, we propose a so-called noisy label, which is a com-bination of the quality label of the whole image and a small random noise value. Ex-periments on LIVE In the Wild Image Quality Challenge Database, which is a newly created image quality assessment database of authentically distorted images, validate the effectiveness of the proposed model and the regularization effect of the noisy label.