A hybrid indicator for realistic blurred image quality assessment

Blurriness is annoying yet common in digital images. Many sharpness assessment indicators using handcrafted features achieve impressive results on synthesized blurring images, while room exists for improvement on realistic datasets. This study presents a hybrid indicator in which no-reference indicators perform as mid-level feature extractors and their outputs are selected using a consensus-based method for discriminative ones. On realistic image datasets, 15 off-the-shelf indicators are explored, and experimental results reveal that the hybrid indicator obtains considerable improvement (>= 21.5%, BID2011; >= 11.6%, CID2013; >= 7.1%, LIVE Challenge; and >= 11.6%, KonIQ-10k) compared to the baseline indicator. Meanwhile, the indicator requires more features for representation of diverse distortions (CID2013, LIVE Challenge and KonIQ-10k) than different blurriness (BID2011). Four regression models are investigated, and fitting neural network leads to overall better results. Realistic image quality assessment is challenging, fusion of existing indicators improves the performance, while to develop advanced indicators remains desirable.