No-Reference Image Quality Assessment Based on Localized Gradient Statistics: Application to JPEG and JPEG2000

This paper presents a novel system that employs an adaptive neural network for the no-reference assessment of perceived quality of JPEG/JPEG2000 coded images. The adaptive neural network simulates the human visual system as a black box, avoiding its explicit modeling. It uses image features and the corresponding subjective quality score to learn the unknown relationship between an image and its perceived quality. Related approaches in literature extract a considerable number of features to form the input to the neural network. This potentially increases the system's complexity, and consequently, may affect its prediction accuracy. Our proposed method optimizes the feature-extraction stage by selecting the most relevant features. It shows that one can largely reduce the number of features needed for the neural network when using gradient-based information. Additionally, the proposed method demonstrates that a common adaptive framework can be used to support the quality estimation for both compression methods. The performance of the method is evaluated with a publicly available database of images and their quality score. The results show that our proposed no-reference method for the quality prediction of JPEG and JPEG2000 coded images has a comparable performance to the leading metrics available in literature, but at a considerably lower complexity.