No-Reference Image Sharpness Assessment Based on Rank Learning

To address the label shortage problem and the fixed-size input constraint of CNN models, we propose a no-reference image sharpness assessment method based on rank learning and effective patch extraction. First, we train a Siamese mobilenet network by learning quality ranks among the synthetically blurred and unsharpen seed images without any human label, which provides effective prior knowledge about the appropriate image sharpness. The extracted single branch finetuned on benchmark datasets is used to predict the subjective rating regarding image sharpness. While the performance of CNN based IQA metrics is compromised due to their fixed-size input constraint, we design a multi-scale gradient guided patch extraction method to increase the quality-sensitive input information and boost the performance. Extensive experimental results on the six public datasets demonstrate that our approach outperforms other state-of-the-art no-reference image sharpness assessment metrics.