It　is　important　to　determine　the　stable　keypoints　and　select　transformation　models　for　image　registration　and　mosaic.　In　this　paper　a　method　is　presented　for　retinal　image　mosaic.　Central　to　the　new　method　is　to　detect　the　PCA-SIFT　(Principal　Components　Analysis-Scale　Invariant　Feature　Transform)　feature　and　estimate　the　quadratic　transformation　model　which　is　employed　to　simulate　the　anatomy　of　human　eyes.　The　transformations　models　are　estimated　by　matching　PCA-SIFT　landmarks.　The　hierachical　notion　is　used　to　map　the　Inter-Image.　The　random　sample　consensus　(RANSAC)　is　used　to　estimate　the　affine　transformation　model　and　remove　exterior　point.　The　quadratic　is　estimated　by　M-estimator.　And　the　weighted　mean　is　used　to　Stitch　retinal　images.　The　proposed　approach　can　effectively　realize　the　retinal　image　mosaic.