This paper presents a novel speckle reduction algorithm based on sparse representation and structure characteristics of PolSAR image. First, each pixel in original image is classified into bright point or line targets, dark point or line targets and others to form a classification map. Second, sparse decomposition and reconstruction is performed on PolSAR image by OMP and K-SVD methods to filter speckle. Finally, the blurred point and line targets in filtered image are enhanced with the classification map. Experimental results with the data of Hayward area from AIRSAR system show that the proposed method is effective both on speckle reduction and scattering characteristics preservation.