Establishing　reliable　correspondences　is　often　challenging　for　retinal　image　registration　with　poor　quality　and　low　overlap.　The　conventional　local　search　methods　usually　find　many　incorrect　correspondences　by　feature　descriptor,　which　would　degrade　the　accuracy　of　image　registration.　In　this　paper,　we　propose　a　robust　correspondence　detection　framework　for　low　overlap　and　poor　quality　retinal　image　registration.　Specifically,　coherent　spatial　criterion　is　utilized　to　remove　the　false　correspondences　based　on　the　initial　matches　in　the　first　hierarchical.　And　the　L2-minimizing　estimator　is　used　for　further　hierarchical　discarding　significant　fraction　of　outliers　and　estimating　the　transformation　parameters　to　align　the　retinal　images　by　affine　model　and　quadratic　model.　Since　such　fitting　inliers　can　be　used　to　preserve　the　significant　correspondences　between　the　fixed　image　and　to-be-aligned　image　with　low　overlap,　it　becomes　more　efficient　to　obtain　accurate　transformation.　Through　quantitative　measurements　and　visual　inspect,　our　proposed　method　shows　the　superior　robustness　and　accuracy　to　the　state-of-the-art　methods.　©　2017　IEEE.