"Human perception can distinguish between visual qualities of images through psycho-visual (subjective)
assessment. However, to impart this quality to a machine is a challenging task, and therefore,
objective image quality assessment plays a vital role in quantifying the visual quality of an image.
Image quality assessment (IQA) aims at designing a mathematical model to gauge the overall
perceptual quality of an image based on the characteristics that are consistent with the subjective
evaluations. IQA plays an important role in various applications such as image compression,
restoration, enhancement, and video streaming. IQA serves the purpose of monitoring image quality,
benchmarking enhancement algorithms, and optimizing parameters in restoration algorithms.
Early research, focused in the area of full-reference methods (where a reference image is
available for comparison), presented a good understanding of characterizing visual quality in terms
error sensitivity, structural similarity, contrast and luminance similarity. As a good reference image
may not be available in most practical applications, this project focuses on no-reference image quality
assessment (NR-IQA). This project proposes a no-reference image quality metric of naturalness
based on the hypothesis that every image has latent additive white Gaussian noise (AWGN). The
quality score consists of four contributing parameters: gradient-based structural integrity, contrast
deviation, and Gabor-based smoothness and naturalness. The final score is a weighted summation
of each of these individual quality factors. A mathematical model was developed on a dataset of
about hundred test images by computing gradient-based structural similarity of corrupted images
w.r.t. the original. Statistical modeling of the observations was found to fit an exponential parametric
model. The standard deviation of the latent (or apparent) AWGN present in any image
was estimated using an SVD-based approach. The factor of no-reference gradient-based structural
integrity (NRGSI) is then computed by a simple back-projection of the estimated noise deviation
on the exponential model. This serves as the first contributing factor. Other contributory factors
are similarly defined based on naturalness, contrast and texture effects of degradation in an image.
Mathematical modeling was performed using four hundred and fifteen natural/noisy images
from standard databases. The proposed approach was further tested on a total of one hundred and
sixty-two images. The overall performance of the proposed no-reference image quality metric is
quantified by calculating the Spearman Rank Order Correlation Coefficient (SRCC) that denotes
the accuracy of objective evaluation with respect to subjective (human visual) evaluation. The
proposed IQA metric displays noteworthy performance (with an accuracy of 80% in terms of SRCC).
In comparison with other state-of- the-art metrics, the proposed metric displays better, and in some
cases, comparable performance.
i"