Abstract
The difficulty and manual nature of procuring human evaluators for ranking the quality of images affected by various types of degradations, and of those cleaned up by developed algorithms, has lead to the widespread adoption of automated metrics, like the Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity Index Metric (SSIM). However, disparities between rankings given by these metrics and those given by human evaluators have encouraged the development of improved image quality assessment (IQA) metrics that are a better fit for this purpose. These methods have been previously used solely for quality assessments and not as objectives in the training of neural networks for high-level vision tasks, despite the potential improvements that may come about by directly optimizing for desired metrics. This paper examines the adequacy of ten recent IQA metrics, compared with standard loss functions, within two trained dehazing neural networks, with observed broad improvement in their performance.