No-reference sharpness metric based on local edge kurtosis

Sharpness metrics that use the frequency spectrum of the image cannot separate the sharpness information from scene content. Sharpness metrics that use spatial gradients of the edges of objects work only for comparisons among versions of the same image. We have developed a content independent, no-reference metric. In this approach, we create an edge profile by detecting edge pixels and enclosing them with 8×8 pixel blocks. For each block we compute sharpness using kurtosis of the DCT. The final metric is the average sharpness of the blocks in the edge profile. The advantages of this method include robust combination of spatial and frequency domain information, flexibility in the edge detection component, and a normalized sharpness scale. The results show high correlation with subjective quality for sharpness-enhanced video.