| With the rapid development of modern information technology,digital images,as an important carrier of information,play an increasingly important role in daily life and production practice.On the one hand,digital images are accompanied by information loss during the process of collection,compression,transmission,and storage,which degrades image quality;on the other hand,with the development of self-media technology,a large number of images generated by non-professional users appear on social media,image quality cannot be guaranteed.Although the evaluation results of the traditional method of subjectively evaluating image quality based on human vision fully conform to human cognition,it cannot meet the processing needs of massive image information on the Internet due to its time-consuming and labor-intensive shortcomings.Therefore,it is necessary to design an objective image quality evaluation method that is consistent with the human visual perception.In this paper,based on the visual characteristics of the human eye and the relevant knowledge of visual psychology,three reference-free image quality evaluation models are designed from the statistics of image information,the visual reasoning process of the brain,and the multi-layer perception characteristics of human vision.It is verified by experiments that the model shows good performance on different image quality evaluation datasets,and has a high consistency with the subjective evaluation results of human eyes.The main work of this paper is as follows:First,a no-reference image quality evaluation model based on image information statistics(SC-BRISQUE)is proposed.In view of the fact that the existing image quality evaluation model does not consider the change of the degraded image information,combined with the Mean subtracted contrast normalized(MSCN)and the image entropy matrix,a method that can reflect the image information size and the image entropy matrix is designed.The fused feature vector of the texture information is used to calculate the objective quality score of the image.The experimental results show that the distribution characteristics of the fusion vector can reflect the degree of image distortion,and it is of practical significance to incorporate the image information feature into the image quality evaluation algorithm.Second,an active reasoning-based no-reference image quality assessment algorithm model(AR-IQA)is proposed.The active reasoning module of the human brain is constructed by generating an adversarial network structure,and the image in the distorted data set is used as the input of the generated adversarial network,and the training network outputs the main content of the image.Then,a regression map from image features to image quality scores is built using the original image of the distorted image,the main content of the image,and the residual map calculated from the distorted image and the main content of the image.Experimental results show that the algorithm has achieved better results in the evaluation of images with specific distortion types,and its comprehensive performance is at the forefront of existing algorithms.Finally,a multi-layer visual perception based image quality assessment model(MVP-IQA)is proposed.First,a multi-layer convolutional neural network is used to extract image semantic information of different dimensions,and the feature maps obtained by different convolutional layers are fused.Then,the Transformer model is used to perform an attention mechanism operation on the fusion features output by the convolutional neural network to obtain the intrinsic relationship of image features.Finally,the fully connected layer is used as a fusion layer to combine image local(convolution)features and global(self-attention mechanism)features,and predict the quality score of the image.Experimental results show that the model shows good performance on three image quality evaluation datasets. |
