Image Quality Assessment Systems Based On Deep Learning

Image Quality Assessment(IQA)provides fast,accurate,objective,and automatic assessment of image quality,and provides evaluation criteria for related computer vision tasks.Therefore,IQA turn into one of the research hotspots in the field of computer vision.Capturing image quality related features plays a critical role of the IQA mission.Deep learning is the mainstream approach in the current IQA field because of its ability to adaptively extract features.There are several reasons for us to study deep learning based IQA approach in an profound way.First,most of the current IQA methods consider the feature distance of the distorted image and the reference image to assess the image quality while ignored their divergence between the feature distributions.Secondly,many IQA approaches apply image patches as quality assessment objects,while less research on how to obtain the scores of the entire image from image patch scores.In addition,the no-reference IQA lacks effective reference information to guide the network learning distortion characteristics.Thus,it is difficult to make its evaluation results highly consistent with the results of the subjective quality assessment.In response to the above problems,this paper proposes the following solutions:Firstly,in this article,a novel full reference IQA model is proposed based on the idea of combining moment estimation in deep convolutional neural networks.We not only applying the divergence of the features between the reference image and its distorted image to assess the image quality,but also employing their difference of distribution to evaluate the image quality in our full reference IQA model.Secondly,in some no-reference IQA methods employ the weighted average method to get the score of entire image.We proposed a simplified weighted average method which indicates that weighted average in no-reference IQA task is workless.What's more we finds that the take the average value of the image patches as the whole image score in the no-reference IQA task has more robust.Finally,for most no-reference IQA approaches,it is usually to directly learn the image-to-score prediction or to obtain the quality of the distorted image by comparing the distorted image with the generated hallucinated reference image.However,we employs the low-pass filtering method and the adaptive method to explore the information of the reference image from the distorted image.The no-reference IQA method can be used to find a relative "reference image" similar to the full reference IQA method,which in turn improves the performance of the no-reference IQA method.The experiment indicates that the proposed algorithm achieves high consistency between the prediction results and the subjective scores in the datasets commonly used in IQA.It also proves that the IQA algorithm proposed in this paper has good robustness.