Research On No-reference Image Quality Assessment Method Based On Convolutional Neural Network

With the continuous development of multimedia,network and information technology,and the fast popularization of diverse clever terminal devices,large-scale digital images are extensively used in lots of fields which includes information communication,public safety,biomedicine,business production,etc.,and play an important role.Digital images can record and transmit beneficial data,and their quality has a big effect at the application value.Therefore,the study of reliable objective image quality assessment methods and the effective monitoring and measuring image quality change are of super importance to the research and application of digital images.The no-reference image quality assessment approach independently measures the quality of the distorted image primarily based totally on the features of the distorted image without relying on the reference image.In this thesis,the following researches and explorations are carried out for the no-reference image quality assessment method combined with the deep learning technology.In view of the trouble that the present technique based on deep learning extracts image spatial domain features and ignores other domain features,an image quality assessment method primarily based on dual-channel convolutional neural network is proposed.Using gradient images to seize the edge changes due to distortion,this method constructs two CNN branch networks to extract features from the image space domain and the gradient domain respectively,and merge them in an end-to-end framework to evaluate image quality.Through experimental verification on two databases,the features extracted from the image gradient domain may be used as an powerful complement to the spatial domain features,and richer image features are useful to enhance the performance of the no-reference image quality assessment method.Existing studies have proven that when an image is distorted,its visual saliency areas commonly are changed,which influences human's visual perception and judgment of image quality.Aiming at the problem that the present no-reference methods ignore the human visual perception features,this paper proposes an image quality assessment method based on visual saliency.By combining the superior perception characteristics of the human visual system with neural networks,a lightweight multitasking model is constructed.Visual saliency maps and quality score prediction tasks are used to guide the neural network to higher learn features related to image quality perception.The monotonicity indexes on the LIVE and CSIQ databases are 0.964 and 0.915,respectively,indicating that the model has good subjective and objective consistency.