A Hybrid Approach for Blur Detection Using Naïve Bayes Nearest Neighbor Classifier

Blur detection of the partially blurred image is challenging because in this case blur varies spatially. In this paper, we propose a blurred-image detection framework for automaticallQy detecting blurred and nonblurred regions of the image. We propose a new feature vector that consists of the informat ion of an image patch as well as blur kernel. That is why it is called kernelspecific feature vector. The informat ion extracted about an image patch is based on blurred pixel behavior on local power spectrum slope, gradient h istogram span, and maximum saturation methods. To make the features vector useful for real applications, kernels consisting of motion-b lur kernels, defocus-blur kernels, and their combinations are used. Gaussian filters are used for filtering process of extracted features and kernels. Construction of kernel-specific feature vector is followed by the proposed Naïve Bayes Classifier based on Nearest Neighbor classification method (NBNN). The proposed algorithm outperforms the up-to-date blur detection method. Because blur detection is an initial step for the de-blurring process of partially blurred images , our results also demonstrate the effectiveness of the proposed method in deblurring process.