Image Quality Assessment Based On Human Eyes Perception Analysis Of Visual Scene

The development of objective Image Quality Assessment(IQA)methods that are highly relevant to human perception is of great significance to the development of image processing,communications,and other fields.In recent years,IQA algorithms based on human eyes perception and effective analysis and modeling of real visual scenes have become a hot research focus in this field.However,most existing IQA models have poor generalization and can only achieve good results in some image databases,but not all databases.In addition,most deep learning-based methods predict image quality through fine-tuning or pre-trained convolutional neural network models,while ignoring the important visual information on each convolutional layer,which contains the deep features of images from low dimension to high dimension.In view of the above problems,two Full-Reference IQA methods are proposed as follows:One image quality assessment method via spatial-transform domains multi-feature fusion is proposed.This method simulates that the Human Visual System(HVS)can simultaneously process the spatial and transform domains feature of the image.Based on the traditional gradient algorithm,the fusion process of gradient components is improved to take the relationship between the central pixel and more neighboring points into consideration,so as to better describe the rich edge changes of the image.Four complementary features including color features,visual saliency features,gradient features and contrast sensitivity functions extracted from the image in the spatial-transform domain are fused.Finally,the quality evaluation model is established by using the Random Forest(RF)regression strategyAnother Full-Reference image quality evaluation model based on deep features and structure-weighted LBP feature is established.This method simulates the feature of HVS sensing image edge information from coarse to fine,and extracts the Local Binary Patterns(LBP)features based on structure weighting at multiple scales from the brightness channel of the image.The gradient enhancement method is adopted to extract the information of each pooling layer of the image in the VGG16 network to capture the deep features from low dimension to high dimension.These deep features cover from the original data at the pixel level to the abstract semantic concepts.The multi-layer network data is extracted to achieve more effective feature expression.Such a model can be used to simulate the complex distortion patterns caused by multiple distortionsThe experimental results show that the PLCC index of the proposed method on TID2013 and LIVE database reaches 0.9452 and 0.9779 respectively,and its predictive power is significantly higher than other mainstream evaluation algorithms.In addition,both two methods have a strong ability to run across data,achieving a good balance between prediction performance and computational complexity.