Composition-preserving deep approach to full-reference image quality assessment

Abstract

Image quality assessment is an important element of a broad spectrum of applications ranging from automatic video streaming to display technology. In this study, we present a novel full-reference image quality assessment algorithm relying on a Siamese layout of pretrained convolutional neural networks (CNNs), feature pooling, and a neural network. Unlike previous methods, our algorithm handles input images without resizing, cropping, or any modifications. As a consequence, it effectively learns the fine-grained, quality-aware features of images. The proposed model derives its core performance from pretrained CNNs, being trained at a higher resolution than that in previous works. The presented architecture was trained on the recently published KADID-10k, which is the largest image quality database and contains 10,125 digital images. Experimental results on KADID-10k demonstrate that the proposed method outperforms other state-of-the-art algorithms. These results are also confirmed with cross-database tests using other publicly available datasets.