Image Quality Assessment Based On Convolutional Neural Networks

Image quality assessment(IQA)has been widely used in image processing tasks,which aims to simulate human vision perception to predict image quality.In the process of designing IQA metrics,researchers need to extract image features not only relate to image quality closely but also have no relationship with image content.Deep convolutional network,which is a global optimization algorithm,extracts features automatically and a lot of advanced training techniques can be used when training them.Because of these reasons,convolutional neural networks has been introduced to IQA problems.In recent deep learning based IQA metrics,the problems about the lack of training data and the inaccurate quality scores of local images.Starting from these problems,a series of improvements about the original method are proposed based on the characteristics of the human vision system,and they get better performance than the original method.The contributions of this paper are as follows:The method which trains CNN with local variance weight was proposed.The effect of different small patches on the quality of the whole image can be roughly measured with the local variance of small patches.The weight loss function can be used,and this change to the loss function can get better performance.The training method based on the density estimation was proposed.The density estimation can be used to automatically find small patches which can represent the whole image in quality prediction.And after many times of iterations,this method can do better than local-variance based methods.Considering that different image scales when human observe have an effect on the image quality,the method combining CNN and image scales was proposed.It is found that this improved method have very good performance.