Review of Quality Assessment Algorithms on the Realistic Blurred Image Database (BID2011)

Accurate realistic blurred image quality assessment (RB-IQA) is challenging due to potential occurrence of blurring in image acquistion, compression and transmission. Derived from the Blurred Image Database (BID2011), a technical review is executed by screening literatures that cited the database for fully understanding the current achievement on realistic image sharpness estimation. By removing review and non-English-written papers and other irrelevant publications, 43 technical papers remain. Generally, the technical algorithms can grouped into shallow- and deep-learning categories. Notably, deep-learning-based RB-IQA algorithms via advanced learning strategies (transfer learning, rank learning, self-supvised learning, continual learning, meta-learning and domain adaptation) are predominantly developed, and remarkable progress has been made. RB-IQA is crucial for evaluating digital imaging devices and benchmarking restoration algorithms, and in future work, efforts should be made to improve the RB-IQA performance closer to human visual perception.

I. Introduction

Realistic blur is annoying yet common in digital imaging. It distorts image readability, hamper content understanding and poses many difficulties in automated image processing. On the other hand, it is prone to occur in image acquistion, data compression, signal transmission and user consumption [1]. When taking a picture, it is found that slight camera shake or out-of-focus causes unexpected image distortions.