Analysis of Structural Similarity for Image Quality Assessment

Objective. The aim of the work is to study the shortcomings of currently widespread objective mathematical criteria based on the principle of measuring the introduced error, as well as to demonstrate the advantages of criteria that take into account the features of the human visual system. Method . The study is based on modeling the introduction of distortions by impulse noise. The resulting image quality is assessed using the above metrics. Quality assessment metrics should be consistent with subjective assessment results for a wide range of images without requiring complex computations; they should have a simple analytical form and be applicable as optimality criteria for optimizing or selecting parameters for image processing systems, including machine learning systems. Result . The experimental results show the weakness of traditional approaches to image quality assessment based on the use of mathematical measures that assess the error introduced by the processing algorithm. Conclusion. These metrics are currently widespread objective criteria that can be used to measure deviations, but they correlate poorly with subjective quality indicators. Metrics that take into account the peculiarities of the human visual system allow us to evaluate image quality as a measurement of structural distortions, rather than introduced errors. Our experiments have shown that the improved metric has a non-linear but exponential dependence on the probability of pixel value distortion, and correlates much better with the subjective visual perception of the quality of the image that has been processed than the standard deviation.