There has been an unprecedented explosion in creation and consumption of high definition videos and images. This has led to the development and adoption of several systems which manipulate graphic content. Most of these systems require a robust, no-reference, domain independent, and subjective quality metric. The state-of-the-art algorithms aimed at no-reference quality assessment of images and videos employ neural networks that are content domain dependent and require application specific training. Further, present systems are trained to estimate quality metric in accordance to objective metrics such as Peak Signal to Noise Ratio or Mean Squared Error. We propose a convolutional neural network architecture that leverages a variable length long short term memory. The model will also make use of the Laplacian as a fourth layer in the input to ensure content domain independence in quality estimation. The system has been trained on the images of KonIQ-10K dataset and tested on KonViD-1k video dataset and LIVE video dataset. It has also been tested on 10 0 0 frames and videos from the VIRAT dataset and images from BSD300 and Set5 datasets with artificial blur and distortion applied. The quality estimations were compared to the mean opinion square given by five human viewers and resulted in a 11% root mean squared error. Thus the proposed model provides a subjective quality estimation of graphic content without any dependence on the domain of the content.(c) 2021 Elsevier Ltd. All rights reserved.
展开▼
