Research On Neural Network Based No-Reference Image Quality Assessment

Images,as the efficient and intuitionistic carrier,have an important position in modern life.In practice,digital images are available to consumers which usually undergo several stages of processing including acquisition,compression,transmission,and presentation.Unfortunately,each stage will introduce certain type of distortion,such as white noise,Gaussian blur and compression distortion.All of them may degrade the received images perceptual quality and cannot meet the requirements of the final viewers.The research goal of no-reference image quality assessment is to establish a machine learning model related with the human visual system and evaluate the quality of the distorted image,without any information of the original reference image.Those methods are close to real life applications,have a certain practical significance.By studying the methods of no-reference image quality evaluation,this paper proposes the following two research results:1)This paper proposed a No-reference image quality assessment using joint multiple edge detection,which uses the similarity between the first-order edge information and the second-order edge information under distortion to predict the quality.The method extracts the gradient,relative gradient and LOG features of the image,then calculates the conditional entropy between the gradient and LOG and the conditional entropy between the relative gradient and LOG,and finally uses Adaboost neural network for regression prediction.Experiments on the public database LIVE and TID2008 show that the proposed algorithm has high consistency with the subjective DMOS value.2)Convolutional neural networks are widely used in the research of computer vision systems due to their powerful feature expression capabilities.In this paper,a noreference image quality evaluation algorithm based on convolutional neural network is proposed.The main innovation of this model is changing the structure of the fully connected layer in the traditional neural network,that solves the problem of insufficient training samples.Compared with other methods,this method does not require additional pretreatment,is a real end-to-end model.Experimental results on a generalpurpose database show that the performance of this method is superior to some stateof-the-art no-reference quality assessment algorithms.