No Reference Image Quality Assessment Based On Convolutional Neural Network And Its Application

Currently,image applications are developing rapidly,and many related image technologies and services have been introduced into our daily lives and many professional fields.Image quality assessment is one of the most challenging directions in the fields of image processing,machine learning and computer vision,and has received more and more attention.Many distortions can be occur when collecting,transmitting,processing and displaying images.Manual image quality supervision cost is too high,and the current image quality automatic detection technology has not achieved the desired effect in real-time and effectiveness.Therefore,the establishment of a real-time and effective image quality assessment method has become an increasingly hot research direction.This paper mainly proposes a planar image quality evaluation method and three stereo image quality assessment methods.The main work is summarized as follows:Firstly,a planar no-reference image quality assessment based on full convolution network is proposed.The network framework predicts the quality map and then weighted pooling to obtain the final predicted image quality score.Then,based on the previous planar image assessment method,extend a stereo image quality assessment.A stereo image quality assessment algorithm based on convolutional network and saliencyweighting is proposed.The main model of the framework is the quality map generation network.Another stereo assessment method is based on a deep fusion network.It mainly combines the shallow convolution network feature map and the deep convolution network graph feature of the stereo image,and adds the contrast feature at the same time,then enters the deep neural network.The last layer is the quality score prediction layer,forming a deep fusion network model that predicts stereo image quality scores.The last stereo image quality assessment based on weight initialization.First,a deep network model is trained using a public planar image database to obtain optimal weight parameters.Then,the image blocks extracted by the left and right viewpoint images and the fused viewpoint image are respectively used as input of the model,and the optimal weight parameter is used as the initialization weight value of the three-way network.The three are combined to get the final predicted quality score.