Deep blind quality evaluator for multiply distorted images based on monogenic binary coding

Frequency-domain blind quality assessment of blurred and blocking-artefact images using Gaussian Process Regression model

“…The methods considered for comparisons, include distortion specific such as Ferzli [9], EZMGM [15], Golestaneh [22], Zhan [23], Zhu [24], H. Liu [16] as well as general purpose methods such as DIIVINE [31], BLIINDS2 [32], CORNIA [38], BRISQUE [29], GMLoG [33], NFERM [34], FRIQUEE [37], BJLC [39], NIQE [41] and IL-NIQE [42]. Some methods designed for multiply distorted images namely GWH-GLBP [44], Miao [45], Zhou [47] and SISBLIM [46] specifically, are also included for fair comparison. These methods can be grouped into two classes: ML-based and non-ML based.…”

“…These methods can be grouped into two classes: ML-based and non-ML based. The IQA methods DIIVINE [31], BLIINDS2 [32], CORNIA [38], BRISQUE [29], GMLoG [33], NFERM [34], FRIQUEE [37], BJLC [39], H. Liu [16], GWH-GLBP [44], Miao [45] and Zhou [47] are ML-based methods, whereas NIQE [41], IL-NIQE [42], Golestaneh [22], Zhan [23], Zhu [24], Ferzli [9], Bahrami [13], EZMGM [15] and SISBLIM [46] are non-ML-based methods.…”

“…have been tested for multiple distortions, but found to have poor performance for these distortions as well as for unseen data. Additionally, few multiple distortion methods [44]- [47] have also been proposed to assess the image quality in presence of two or more commingled type of distortions. GWH-GLBP [44] was proposed by Li et al to assess multiple distortions in images.…”

“…For training-free methods like NIQE [41] and IL-NIQE [42], 1000 trials were performed on the test set for consistency and the corresponding median value are considered. For performance evaluation of other methods namely BJLC [39], Zhou [47], Golestaneh [22], Zhan [23], Zhu [24], Ferzli [9], EZMGM [15], Bahrami [13], and H. Liu [16] their results were taken either from corresponding original papers or most related articles. Since, Golestaneh [22], Zhan [23], Zhu [24] methods were developed for JPEG distortion, whereas Ferzli [9], EZMGM [15], Bahrami [13] were designed for GBLUR type of distortions, hence the scores for these methods will be reported for corresponding specific distortion type only.…”

“…Since, Golestaneh [22], Zhan [23], Zhu [24] methods were developed for JPEG distortion, whereas Ferzli [9], EZMGM [15], Bahrami [13] were designed for GBLUR type of distortions, hence the scores for these methods will be reported for corresponding specific distortion type only. For method proposed by Zhou [47], scores are reported only on MDID2013 [46] and LIVEMD [56] as the original paper mentions results only on these two databases.…”