Doklady Belorusskogo gosudarstvennogo universiteta informatiki i radioèlektroniki (Dec 2019)

COMPARATIVE ANALYSIS OF NO-REFERENCE MEASURES FOR DIGITAL IMAGE SHARPNESS ASSESSMENT

  • Y. I. Golub,
  • F. V. Starovoitov,
  • V. V. Starovoitov

DOI
https://doi.org/10.35596/1729-7648-2019-125-7-113-120
Journal volume & issue
Vol. 0, no. 7 (125)
pp. 113 – 120

Abstract

Read online

Recently, problems of digital image sharpness determination are becoming more relevant and significant. The number of digital images used in many fields of science and technology is growing. Images obtained in various ways may have unsatisfactory quality; therefore, an important step in image processing and analysis algorithms is a quality control stage of the received data. Poor quality images can be automatically deleted. In this article we study the problem of the automatic sharpness evaluation of digital images. As a result of the scientific literature analysis, 28 functions were selected that are used to analyze the clarity of digital images by calculation local estimates. All the functions first calculate local estimates in the neighborhood of every pixel, and then use the arithmetic mean as a generalized quality index. Testing have demonstrated that many estimates of local sharpness of the image often have abnormal distribution of the data. Therefore, some modified versions of the studied functions were additionally evaluated, instead of the average of local estimates, we studied the Weibull distribution parameters (FORM, SCALE, MEAN weib, MEDIAN weib). We evaluated three variants of the correlation of quantitative sharpness assessments with the subjective assessments of human experts. Since distribution of local features is abnormal, Spearman and Kendall rank correlation coefficients were used. Correlation above 0.7 means good agreement between quantitative and visual estimates. The experiments were carried out on digital images of various quality and clarity: artificially blurred images and blurred during shooting. Summing up results of the experiments, we propose to use seven functions for automatic analysis of the digital image sharpness, which are fast calculated and better correlated with the subjective sharpness evaluation.

Keywords