Abstarct: One of the most common image degradation is blur problem. If Camera moves in the exposure time, whole image becomes blurry. This kind of blur called global motion blur. When an object in an image moves in the exposure time, some parts of the image become blurry. This degradation called local motion blur. In the other hand, if we deblur whole image, sharp parts of the image degrade.so first the blur parts of the image must be detected. In this thesis, three features is introduced to detect the blur parts of image. These features are: discrete cosine transform, singular value decomposition and cepstrum. In the proposed method, at first image divided into square overlapping blocks, the size of these blocks are 3×3. Then the value of features are computed for each block and compared with a suitable threshold. Each feature separately decides that the block is blur or not, and the final result voted up. Final decision is pixel wise, if at least 2 features marked the pixel as blur, in final result pixel considered as blur, too. After detecting process, morphology is used for eradicating some small parts of non-blurry in the blur parts and vice versa. Then blur kernel is estimated and image reconstructed with sparse non-blind deconvolution technique. Running fast and not-existence of block effects are the benefits of this method