Statistical modeling of perceptual blur degradation in the wavelet domain

To automatically detect blur in images, without needing to perform blur kernel estimation, we develop a new blur descriptor. It is modeled by image perceptual gradient statistics. As blurring affects especially edges, the proposed idea turns on extract specific statistical features from the perceptual edge map in the wavelet domain using the just noticeable blur concept (JNB). Extracted statistical features are used to robustly classify images as perceptually blurred or sharp using the support vector machines (SVM). The proposed descriptor performance is evaluated in terms of classification accuracy across different datasets. Obtained results revealed high correlation values of the proposed perceptual statistical features against subjective scores.