Blurred image regions detection using wavelet-based histograms and SVM

This paper presents an algorithm for detection and localization of blurred regions in images. The algorithm is based on discrimination of the gradient distributions between blurred and non-blurred image regions. For this purpose, global wavelet transform of Y component of the image is applied, and the obtained wavelet map is divided into overlapping patches. Then a trained probabilistic SVM classifier estimates the blur level of the patches on their wavelet gradient histograms and thereby probability map is constructed. Finally, we perform a more precise determination of borders of blur region based on estimated Laplace distribution of its wavelet coefficients.