In-the-wild No-Reference Image Quality Assessment using Deep Convolutional Neural Networks

With the ever-growing portion of internet traffic associated with multimedia data and the existence of multiple copies of the same content in various forms, it has become vital to measure the quality of the image and video files. In most cases, without access to the original file, the quality shall be assessed solely based on the available file. Specifically, the challenge of no-reference image quality assessment (NR-IQA) is to predict a quality measure for given images in a consistent manner with human perception of quality. Conventional NR-IQA methods try to fit certain distortion models to a given image and quantify the quality. In practice, however, an image is affected by a combination of multiple distortion types. In this paper, we propose a deep convolutional neural network (CNN) for in-the-wild (no specific distortion model) NR-IQA challenge, which learns the quality measure without classifying the distortion type. Unlike most methods which extract features, the proposed CNN structure receives the full image as the input and estimates the distribution of human opinion score for the quality.