Selecting Low-level Features for Image Quality Assessment by Statistical Methods

Abstract

Image quality assessment is an important component in every image processing system where the last link of the chain is the human observer. This domain is of increasing interest, in particular in the context of image compression where coding scheme optimization is based on the distortion measure. Many objective image quality measures have been proposed in the literature and validated by comparing them to the Mean Opinion Score (MOS). We propose in this paper an empirical study of several indicators and show how one can improve the performances by combining them. We learn a regularized regression model and apply variable selection techniques to automatically find the most relevant indicators. Our technique enhances the state of the art results on two publicly available databases.

A Regression-Based Family of Measures for Full-Reference Image Quality Assessment

“…In order to show dataset independent results, each introduced multimeasure was evaluated on all datasets. In the literature, some authors used different numbers of images from benchmarks for this purpose, e.g., 30% [13], [15], a one dataset [29], [24], [30], [18], [35], or even several datasets jointly, as in [20]. The following four IQA benchmarks were used: TID2013 [2], TID2008 [43], CSIQ [23], and LIVE [8], they are characterised in Table 1.…”

“…In the table, the best three results for a given benchmark are written in boldface, results not reported are denoted by "-". Some measures are not benchmark independent, i.e., their authors reported evaluation results on the benchmark that took part in the development of the measure without providing results for other benchmarks, or prepared a one IQA measure for each benchmark, also without cross-benchmark tests, e.g., [24], [26], [27], [29], [30], [31], [35], [36]. Results for approaches that are not dataset independent were excluded from comparison, they are written in italics in the table.…”

“…A linear combination of IQA measures can be found in [28]. In [29], a preliminary approach was presented with a regularized multi-dimensional polynomial estimator for combining seven IQA measures. A conditional Bayesian mixture of experts model with a support vector machines classifier was used in [30] for combining SSIM, VSNR, and VIF using k-nearest-neighbour regression.…”

Objective No-Reference Video Quality Assessment Method Based on Spatio-Temporal Pixel Analysis

“…, K different empirical sequences, whose performance is analyzed by PLCC distribution using the box-plot statistical distribution. Literature on image and video quality evaluation does not adopt the box-plot for performance analysis and it uses a small K-value for cross-validation process, for instance, in [8], [30], [31]- [33], [34], and [35]- [37] the K-value is equal to 1, 2, 5, 6, and 10, respectively, while we use a large random permutation of training-test pairs sets with K = 1, 000, i.e., one thousand distinct training-test set partitions evaluated in the cross-validation process using the box-plot statistical tool to measure the PLCC distribution.…”

Image De-Noising and Compression Using Statistical based Thresholding in 2-D Discrete Wavelet Transform

“…Wavelet transform domain shows sparse property and wavelet maps noise from image domain to wavelet domain thus, the energy of the image is concentrated in high coefficients while noise energy is mostly in low coefficients values. This principle enables the separation of image important features from noise [13]. Now the procedure in which small coefficient values are dropped in large coefficients values left is known as hard thresholding but the drawback it produces visual artifacts.…”