No-reference Image Quality Assessment Based On Deep Learning

With the continuous development of Internet technology,there are a huge number of multimedia resources,especially images,in the network.Accurate prediction of image quality is particularly important for the efficiency of information dissemination.No-reference image quality assessment(NR-IQA)based on deep learning attracts a great research attention recently.Especially,the introduction of convolutional neural network(CNN)in no-reference image quality assessment(NR-IQA)gains great success in improving its prediction accuracy.However,there exists something which can still be optimized.Two NR-IQA methods based on CNN are proposed in this thesis.On the one hand,the research of these methods can promote the development of image quality assessment;on the other hand,it can make us better grasp the application of CNN in the field of computer vision,and constantly improve the network.Firstly,this thesis introduces the background and significance of the research,and then makes a brief introduction of the theoretical basis about machine learning and deep learning,image processing and image quality assessment.Then two new methods of image quality assessment based on deep learning are proposed.In the first method,they propose a quality-distinguished and region-weighted NR-IQA approach based on convolutional neural network(QDRW-CNN).They improve the predicting accuracy by two proposed mechanisms: quality-distinguished adaption and region-wise regression.The former trains multiple models from different subsets of a dataset and adaptively selects one for predicting quality score of a test image according to its quality level,and the latter divides an image into different regions and assigns different weights to them for calculating image quality score.They further improve the efficiency of network training by a new patch sampling way based on the visual importance of each patch.In the second method,because the performance of CNN relies on the magnitude of training samples.However,many widely-used existing image databases cannot provide adequate samples for CNN training.In this paper,They propose a pair-comparing based convolutional neural network(PC-CNN)for blind image quality assessment.By taking reference images into consideration,they generate more training samples of patch pairs by different combination of distorted images and reference image.They build a new CNN network which has two inputs for patch pairs and two outputs predicting the scores of patches.They conduct extensive experiments on several public databases and compare our proposed QDRW-CNN and PC-CNN with existing state-of-the-art methods.Our experimental results demonstrate that our proposed two methods outperform the others both in terms of accuracy.