Blind　image　deconvolution　recovers　a　sharp　image　from　a　blurred　image　when　the　blur　kernel　is　unknown.　To　solve　this　underdetermined　inverse　problem,　most　existing　methods　exploit　various　image　priors　to　constrain　the　solution.　In　this　study,　we　propose　a　blind　deconvolution　method　based　on　cross-scale　dictionary　learning,　in　which　the　down-sampled　blurry　image　is　used　to　learn　a　dictionary　as　training　samples　and　the　texture　region　is　represented　sparsely　over　the　dictionary　as　the　regularization　term.　Because　the　down-sampling　process　weakens　the　blur　of　the　image,　it　will　result　in　the　formation　of　redundant　cross-scale　similar　patches.　To　ensure　that　a　sharp　image　is　represented　sparsely,　sharper　image　patches　from　the　down-sampled　image　in　this　study　were　used　to　learn　the　dictionary　as　training　samples.　The　results　showed　that　the　sparse　representation　error　of　the　texture　patch　from　the　sharp　image　was　less　than　that　from　the　blurred　image,　further　diminishing　the　sparse　representation　error　over　the　dictionary,　and　the　intermediate　latent　image　approached　the　sharp　image.　The　mean　peak　signal-to-noise　ratio　of　the　results　by　our　method　on　the　dataset　of　Kohler　et　al.　is　29.54　dB.　Experimental　results　on　blurry　images　demonstrated　that　our　method　can　estimate　large　blur　kernels　accurately　and　that　it　has　good　robustness.　©　2021,　Science　Press.　All　right　reserved.