Research On Deep Learning Based Image Quality Assessment

Image Quality Assessment(IQA)refers to firstly extract image quality related features,and then the method of machine learning is used to evaluate the degree of image distortion..The quality score is an important index for evaluating the degree of image and video distortion,which seriously affects the accuracy,stability,and robustness of the image processing system.At the same time,the computer vision application system increases rapidly with the great development of deep learning.Accurate and effective evaluation system quality algorithm of input image is imminent.Therefore,image quality evaluation has become one of the key technologies for computer vision in recent years.This paper explores image quality assessment algorithms that uses combined features or deep learning.Firstly,this paper introduces a method using the combined artificial features to complete the image quality evaluation.The algorithm extracts image quality related features in different transform domains and multiple scales.This framework uses K-means to generate the image block dictionary,and then the distorted image patches are encoded in this dictionary.In the experiment,it is found that the parameters of different distributions used to fit the encoding coefficient can reflect the distortion type and the degree of the image.Moreover,quality evaluation algorithms based on combined features stand out in many distortion types.Then,this paper explores an image quality assessment algorithm that uses deep learning.The algorithm generally models IQA as a standard regression problem.Convolutional neural networks can directly map image data to quality scores.However,the literature rarely explores the effect of neural network structure on the performance of image quality evaluation;in addition,the introduction of attention mechanisms in the traditional image quality assessment task is generally defined by human.Most of the custom saliency maps are analogous to those in traditional computer vision tasks and are not accurate in image quality evaluation.Therefore,we first parameterize the depth and width of the neural network structure,determine the optimal network structure through comparison experiments,and then compare the impact of the linear regression and support vector regression on the experiment.At the same time,the saliency map is automatically generated using the full convolutional neural network,which implies the importance of different regions in the image to full image quality score.Finally,a two-step training method is used to train the multi-tasking network composed of image quality score regression tasks and weight graph generation tasks,and the training network is more accurate.The final Spearman correlation coefficient in the LIVE dataset reached 0.958.