This　paper　proposes　a　new　method　of　inter　prediction　based　on　low-rank　matrix　completion.　By　collection　and　rearrangement,　image　regions　with　high　correlations　can　be　used　to　generate　a　low-rank　or　approximately　low-rank　matrix.　We　view　prediction　values　as　the　missing　part　in　an　incomplete　low-rank　matrix,　and　obtain　the　prediction　by　recovering　the　generated　low-rank　matrix.　Taking　advantage　of　exact　recovery　of　incomplete　matrix,　the　low-rank　based　prediction　can　exploit　temporal　correlation　better.　Our　proposed　prediction　has　the　advantage　of　higher　accuracy　and　less　extra　information,　as　the　motion　vector　doesn＇t　need　to　be　encoded.　Simulation　results　show　that　the　bit-rate　saving　of　the　proposed　scheme　can　reach　up　to　9.91%　compared　with　H.　264/AVC.　Our　scheme　also　outperforms　the　counterpart　of　the　Template　Matching　Averaging　(TMA)　prediction　by　8.06%　at　most.