Efficient neural-network-based no-reference approach to an overall quality metric for JPEG and JPEG2000 compressed images

Reliably assessing overall quality of JPEG/JPEG2000 coded images without having the original image as a reference is still challenging, mainly due to our limited understanding of how humans combine the various perceived artifacts to an overall quality judgment. A known approach to avoid the explicit simulation of human assessment of overall quality is the use of a neural network. Neural network approaches usually start by selecting active features from a set of generic image characteristics, a process that is, to some extent, rather ad hoc and computationally extensive. This paper shows that the complexity of the feature selection procedure can be considerably reduced by using dedicated features that describe a given artifact. The adaptive neural network is then used to learn the highly nonlinear relationship between the features describing an artifact and the overall quality rating. Experimental results show that the simplified feature selection procedure, in combination with the neural network, indeed are able to accurately predict perceived image quality of JPEG/JPEG2000 coded images