No-Reference Image Quality Assessment Method For Distorted And Restored Images

No-reference image quality assessment can effectively improve the ability of information acquisition and application for images.In this paper,three kinds of blind image quality assessment methods are proposed for distorted images and restored images to identify and quantify the quality of the images.The proposed method can also assist the design of restoration algorithm in order to ensure the image quality of restored images is higher than those of original images.(1)A no-reference image quality assessment method fusing deep learning and statictical visual features for distorted images is proposed.The statistical visual features of the image are extracted,and according to their characteristics,the network structure with dual-channel convolutional layer is designed.The localized normalized luminance image patches and the localized normalized luminance LBP image patches are respectively input into the dual-channel network,and fusing the statistical visual features,the no-reference image quality assessment is achieved.The comparative experiments between the FDSVDIQA and other state-of-the-art methods were conducted in the LIVE,LIVEMD and MDORSID respectively.The experimental results show that the FDSVDIQA method has good subjective consistency and is robust both for natural distorted images and remote sensing distorted images.The database independence experiment was carried out on the LIVEMD and MDID databases.According to the experimental results,it can be concluded that the FDSVDIQA has database independence.(2)A no-reference image quality assessment method based on fusion of multiple features for recovered remote sensing images is proposed.The features of gradient?entropy and difference images are extracted from the recovered images to construct the feature vector which can represent image quality.AdaBoost_BP neural network is trained to learn the relationship between image features and image quality scores,thus no-reference quality assessment for restored images is achieved.A restored blurred optical remote sensing image database RBORSID is established.The comparative experiments between the GEDIQA and other state-of-the-art methods were conducted in the RBORSID.The experimental results show that GEDIQA algorithm has good subjective consistency for restored remote sensing images,and its performance is better than other algorithms.(3)A no-reference image quality assessment method fusing deep learning and spatial & frequency domain features for restored images is proposed.Firstly,the spatial domain features of gradient?differential images and frequency domain features of local two-dimensional entropy are extracted.Thus,spatial and frequency domain features of the images can be obtained by connecting these features;Then,the normalized luminance patches as well as the spatial and frequency domain features of the restored image are input into the CNN network;Finally,the quality of the restored images is achieved by learning and training.The experimental results show that the FDSFRIQA has good subjective and objective consistency for restored images,and its performance is superior to other algorithms.The experiments of FDSFRIQA were conducted in different types of databases,including single-distorted natural image database LIVE,multiply distorted natural image database LIVEMD and multiply distorted remote sensing image database MDORSID.The experimental results show the good subjective consistency and robustness for different types of images of FDSFRIQA.