An　improved　K-SVD　method　based　on　non-noisy　pixel　reconstruction　(PK-SVD)　is　proposed　to　filter　impulse　noise.　In　the　phase　of　image　reconstruction,　non-noisy　pixels　are　applied　in　the　construction　of　optimal　function　to　obtain　the　reconstructed　image　and　improve　the　filtering　performance,　and　the　optimal　function　is　solved　by　integrating　the　hierarchical　property　into　the　OMP　algorithm.　In　the　phase　of　dictionary　training,　PK-SVD　uses　the　iterant　K-singular　value　decomposition　to　renovate　both　atoms　and　their　coefficients　rather　than　fixes　the　coefficients.　The　simulation　results　show　that　compared　with　the　other　three　methods,　PK-SVD　obtains　the　sparsest　dictionary　and　the　clearest　image　with　higher　peak　signal　to　noise　ratio.　©,　2014,　Journal　of　Pattern　Recognition　and　Artificial　Intelligence.　All　right　reserved.