Blind quality assessment of night-time photos: A region selective approach

Despite the emergence of low-light enhancement algorithms and the associated quality assessment metrics in literature, there are rare works considering the quality assessment of real night-time photos captured by mobile cameras. In this paper, we handle this task by first constructing a night-time photo database (NPHD), which consists of 510 photos captured by 30 mobile devices in 17 scenes. Their mean opinion scores are rated by 10 people using the anchor ruler method. Furthermore, we propose a region selective approach for the objective image quality assessment (RSIQA), based on which different feature sets are extracted. Specifically, the center and around regions are partitioned for the brightness, contrast, vignetting, saturation and shading. The brightest areas are located as the region where the highlight suppressing capability is qualified. Finally, we select the foreground and sharpest regions for the assessment of preserving details, naturalness, noises, and image structure. To map different/multiple quality attributes of the night-time photo into a single quality score, four regressors: support vector regression, decision tree, random forest or AdaBoost.R2 are chosen and compared. Experiments on NPHD demonstrate that the proposed RSIQA achieves superior result compared to 17 state-of-the-art, 4 types of quality metrics, including conventionally general-purpose, deep learning based, contrast oriented and night specific ones.