Blind　image　deconvolution　aims　to　recover　the　sharp　image　from　a　blurred　one　when　the　blur　kernel　is　unknown.　To　solve　this　underdetermined　inverse　problem,　most　existing　methods　exploit　various　image　priors　to　constrain　the　solution.　Our　work　is　inspired　by　the　observation　that　the　cross-scale　self-similarity　of　the　sharp　image　will　diminish　after　blurring,　and　the　down-sampled　blurry　image　has　stronger　similarity　with　the　sharp　image　than　the　blurry　image.　In　this　paper,　we　propose　a　blind　deconvolution　method　based　on　cross-scale　low　rank　prior,　in　which　the　similar　image　patch　group　is　formed　from　sharper　patches　sampled　from　the　down-sampled　image,　and　the　low　rank　matrix　approximation　is　used　to　explore　the　low　rank　structure　of　this　group.　By　introducing　the　cross-scale　low　rank　prior　as　the　regularization　constraint,　the　intermediate　latent　image　is　enforced　to　contain　the　sharp　edges　and　fine　details.　The　low　rank　matrix　approximation　elegantly　indicates　the　global　structure　of　data,　allowing　for　the　noise-avoiding　kernel　estimation　without　acquiring　any　additional　handling　of　noise.　Experimental　results　on　blurry　images　and　blurred-noisy　images　demonstrate　that　our　method　can　estimate　accurate　large　blur　kernels,　meanwhile,　it　has　good　robustness　to　noise.　©　2022　Science　Press.　All　rights　reserved.