Font Size: a A A

Similar Quality Map Generation Based On Deep Learning

Posted on:2022-03-26Degree:MasterType:Thesis
Country:ChinaCandidate:Z Y ChenFull Text:PDF
GTID:2518306338490644Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
In image applications,it is highly desirable to obtain an image quality map that reveals the image quality degradations in a localized manner.Such goal can be realized by the fullreference image quality assessment(FR-IQA)methods which compare the distorted image and its original undistorted image pixel by pixel,leading to a localized similarity or discrepancy map that can serve as the image quality map.However,in such manner,the original image must be available in advance,which is quite strict for practical applications.In this paper,we propose a novel no-reference approach to predict the local image quality map without referring to the original undistorted image.Unlike traditional methods of predicting the quality map,we train an end-to-end(graph to graph)deep learning neural network frame that is the first neural network framework in the world to yields the image quality map through the promising generative adversarial learning methodology.Specifically,the proposed scheme contains two deep neural networks module,one is the generative network module for producing the quality map,and the other one is the discriminative network module for judging the predicted results.By resorting to the annotations of the FR-IQA methods,we train the generative and discriminative networks defiantly on a large-scale image database and obtain a robust model for quality map prediction.Thorough experiments performed on four standard image databases in the direction of image quality evaluation confirm the high effectiveness and good generalization ability of the proposed method.The experimental results are internationally advanced.
Keywords/Search Tags:image quality map, image quality assessment, generative adversarial learning, deep learning
PDF Full Text Request
Related items