CNN-based Cross-dataset No-reference Image Quality Assessment

Recent works on no-reference image quality assessment (NR-IQA) have reported good performance for various datasets. However, they suffer from significant performance drops in cross-dataset evaluations which indicates poor generalization power. We propose a Siamese architecture and training procedures for cross-dataset deep NRIQA that achieves clearly better performance. Moreover, we show that the architecture can be further boosted by i) pre-training with a large aesthetics dataset and ii) adding low-level quality cues, sharpness, tone and colourfulness, as additional features.