No-Reference Image Sharpness Assessment via Multi-scale Texture Naturalness
Zhongting Sun, Gang Hua, Yonggang Xu · 2023
One of the common types of distortion in image acquisition is blurring. The influence of the multi-scale characteristics of the human visual system makes the perception of blur/clarity more complex. In addition, blurring can lead to a loss of texture naturalness. Based on these considerations, this article proposes an image clarity evaluation method based on multi-scale texture naturalness, which has the characteristics of fast and effective. This method first uses a Gaussian filter to passively further blur the input image, considering that the degree of blur will affect the similarity between the image and its further blurred version. In order to adapt to the impact of viewing distance on image rotation perception, the input image and its further blurred versions are downsampled to different scales. At each scale, statistical texture features based on grayscale co-occurrence matrix (GLCM) are extracted for subsequent clarity evaluation. At each scale, the similarity of texture features between the original image and its re blurred image is calculated as sharpness features. Finally, by learning texture naturalness features, a Support Vector Regression (SVR) model is generated. The experimental results on five image databases show that the metrics proposed in this paper outperform any existing clarity evaluation metrics and blind universal quality metrics in terms of performance. In addition, it has good real-time performance and encouraging generalization ability.