Image quality assessment with lasso regression and pairwise score differences

Abstract

The reception of multimedia applications often depends on the quality of processed and displayed visual content. This is the main reason for the development of automatic image quality assessment (IQA) techniques which try to mimic properties of human visual system and produce objective scores for evaluated images. Most of them require a training step in which subjective scores, obtained in tests with human subjects, are used for parameters tuning. In this paper, it is shown that pairwise score differences (PSD) can be successfully used for training a full-reference hybrid IQA measure based on the least absolute shrinkage and selection operator (lasso) regression. The results of extensive experimental evaluation on four largest IQA benchmarks show that the proposed IQA technique is statistically better than its version trained using raw scores, and both approaches are statistically better than state-of-the-art full-reference IQA measures. They are also better than other hybrid approaches. In the paper, the evaluation protocol is extended with tests using PSD.

Pairwise Learning to Rank for Image Quality Assessment

“…Gao et al [38] presented an NR-IQA framework which formulated the problem of mapping difference feature vectors to preference labels as a classification problem and used a multiple kernel learning algorithm to learn a classification model. Mariusz [39] introduced a hybrid FR-IQA index which used the lasso regression and pairwise score differences. Ma et al [40] employed a neural network-based PLR algorithm to learn an opinion-unaware NR-IQA model.…”

“…To validate the performance of the proposed PLRA index, we compared it with 22 state-of-the-art IQA metrics. Table 5 shows the comparisons of these 22 state-of-the-art IQA metrics on four databases: four of them are machine learningbased metrics, namely SVDR [29], CD-MMF [30], NMF [31], and IrSIM [39]; two of them are CNN-based metrics, namely, WaDIQaM-FR [32] and DeepQA [33]. The three best assessment results for each database are highlighted in red, green, and blue and are formatted in boldface.…”

“…We compared the proposed PLRA with four training-based IQA metrics, namely SVDR [29], CD-MMF [30], NMF [31], and WaDIQaM-FR [32]. Because the codes and cross-database test performance values of the other training-based metrics (IrSIM [39] and DeepQA [33]) that were compared in Table 5 were not available, we did not compare these in Table 8.…”

Current Trends and Advances in Image Quality Assessment

“…The application of neural networks and some other machine learning algorithms for the fusion of IQA metrics was discussed by Barri et al [41], whereas the use of pairwise score differences to obtain the lasso regression Similarity measures (lrSIMs) was examined by Oszust [42]. Nevertheless, the obtained results were slightly worse than those presented in [40].…”

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

“…In [32], in turn, image blocks were first classified using decision trees and then FSIM [13], mean squared error, and different variations of PSNR [33] were combined. In [34], in turn, lasso regression models were obtained using pairwise scores differences. Six IQA measures were fused using neural network in [35].…”