How do you choose the best image quality metrics for different applications and domains?

1 What is IQA and why is it important?

IQA is the process of measuring and evaluating how well an image preserves the information and appearance of the original scene or object. IQA can be used for various purposes, such as compression, enhancement, restoration, segmentation, recognition, and analysis. IQA is important because it can help improve the quality and efficiency of image processing systems, as well as the user experience and satisfaction of the end-users.

2 How to classify IQA metrics?

IQA metrics can be broadly classified into three categories: full-reference (FR), reduced-reference (RR), and no-reference (NR) metrics. FR metrics compare the distorted image with the reference image, which is assumed to be of high quality and fully available. RR metrics use some features or information extracted from the reference image, but not the whole image. NR metrics do not require any reference image, and rely on some prior knowledge or assumptions about the image content or distortion.

3 What are the advantages and disadvantages of each category?

FR metrics are usually more accurate and reliable, as they have access to the complete reference image. However, they are also more demanding in terms of storage, bandwidth, and computation, and may not be applicable in situations where the reference image is unavailable or inaccessible. RR metrics are less demanding than FR metrics, but still require some communication or synchronization between the reference and distorted images. They are also less accurate and reliable than FR metrics, as they may lose some information or introduce some errors during feature extraction or transmission. NR metrics are the most flexible and practical, as they do not need any reference image or feature. However, they are also the most challenging and complex, as they have to deal with the uncertainty and variability of the image content and distortion.

4 How to evaluate IQA metrics?

IQA metrics can be evaluated by comparing their results with the subjective ratings of human observers, who are considered to be the ultimate judges of image quality. The subjective ratings can be obtained by conducting experiments with a large and diverse group of participants, who are asked to rate the quality of a set of images on a numerical or categorical scale. The objective results of IQA metrics can then be correlated with the subjective ratings, using some statistical measures such as Pearson correlation coefficient (PCC), Spearman rank-order correlation coefficient (SROCC), or root mean squared error (RMSE).

5 How to choose IQA metrics for different applications and domains?

When choosing IQA metrics, there is no single or universal answer since different applications and domains may have different requirements and preferences. Factors that may influence the selection of IQA metrics include the type and level of distortion affecting images, the availability and quality of the reference image, the computational complexity and efficiency of the metrics, and their compatibility and consistency with human perception of image quality. Consequently, it is important to conduct a careful analysis and comparison of these factors, as well as the specific objectives and constraints of each application and domain, when choosing IQA metrics.