Research On Universal And Unsupervised No-Reference Image Assessment Algorithm

With the rise of mobile internet revolution, demands for image and video services also increased. However, due to the non-ideal characteristics of the channel environment, the quality of transmitted images and videos is always damaged, that is, image quality appears distortion. In order to improve users’ experience, we often need to evaluate the quality of distorted images, and then send scores of distorted images to the source end, which can use some adaptive algorithms to slow down the non-ideal characteristics of the channel environment. Therefore, with the increase in demand for image or video services, making the image quality evaluation has become inevitable, and humans can benefit a lot from an excellent image assessment algorithm without doubt. Based on such background, this article proposed several relatively effective image quality assessment algorithms.This paper studies the existing no-reference image quality assessment methods, and made the following findings:(1) Proposed a general-purpose and unsupervised no-reference algorithm based on feature pool and feature extraction of interested regions and edge regions. The algorithm achieved quality metric by measuring the words distribution between test images and pristine images.(2) Proposed a general-purpose and unsupervised no-reference algorithm based on LDA topic model, Firstly, the algorithm trained the LDA model by some image samples. Secondly, using the LDA model to mine topic distribution of pristine and distorted images. Finally, get quality metric by measuring the differences of topic distributions. (3) Proposed a completely blind image quality assessment algorithm, the algorithm used GGD distribution to model the sub-band wavelet coefficients to form feature vectors, and then using MVG model to fit the feature vectors, and finally scoring distorted images by measuring the fitting parameters differences between pristine and distorted images.In the end, the above three algorithms were thought to have better consistence with people’s subjective perception by experimental demonstration.