HDR IMAGE QUALITY ASSESSMENT USING MACHINE-LEARNING BASED COMBINATION OF QUALITY METRICS

We present a Full-Reference Image Quality Assessment (FR-IQA) approach to improve High Dynamic Range (HDR) IQA by combining results from various quality metrics (HDR-CQM). To combine these results, we apply linear regression and various machine learning techniques such as multilayer perceptron, random forest, random trees, radial basis function network and support vector machine (SVM) regression. We found that using a non-linear combination of scores from different quality metrics using SVM is better at prediction than the other techniques. We use the Sequential Forward Floating Selection technique to select a subset of metrics from a list of quality metrics to improve performance and reduce complexity. We demonstrate improved performance using HDR-CQM as compared to a number of existing IQA metrics. We find that our HDR-CQM metric comprised of only four metrics can obtain statistically significant improvement over HDR video quality measure (HDR-VQM), the best performing individual IQA metric for HDR still images.