Fig 1.
Flowchart of the proposed approach.
In an offline training process, the proposed approach is obtained using some of images from a benchmark dataset. Images are assessed by full-reference IQA measures. Then, a genetic algorithm selects IQA measures and assigns weights to them. Obtained weights for linear combination of selected measures are used in image quality assessment tasks.
Table 1.
IQA benchmark image datasets.
Table 2.
Performance comparison of resulted fusion measures with IQA models that were used in optimisation.
Table 3.
Statistical significance tests.
Fig 2.
Scatter plots of subjective opinion scores against scores obtained by the two best IQA measures and LCSIM3 on used datasets.
Different types of distortions are represented by different colours; the set of colours is coherent within a dataset. Curves fitted with logistic functions are also shown.
Fig 3.
Absolute values of the difference between objective scores and nonlinearly fitted subjective scores for 50 exemplary images from LIVE dataset.
For each image, a smaller value denotes objective assessment which is closer to human evaluation.
Table 4.
Time and memory costs of IQA measures used in the optimisation.
Table 5.
Comparison of the approach with other fusion IQA measures based on SRCC values.
Table 6.
SROCC values of IQA measures for each distortion type.
Table 7.
SRCC between objective scores of IQA measures on CSIQ.
Table 8.
Results of the experiment with predefined number of aggregated IQA models on CSIQ dataset.
Table 9.
Influence of the β on obtained IQA fusion measures.