Convolutional neural networks (CNNs) are used in an increasingly systematic way in a great variety of computer vision applications, including image quality assessment. However, their application to evaluate the perceived quality of images is strongly limited by the lack of adequate
and consistent training data. A CNN-based framework for evaluating image quality of consumer photographs is made up of several building blocks that can be implemented in different ways. In this article, we schematically illustrate how these building blocks have been implemented and combined
so far to create feasible solutions that take the most positive characteristics of CNNs while mitigating their intrinsic limitations. Some experimental results are reported to show the effectiveness of CNN-based solutions on real-world image quality datasets.
