To　obtain　the　sparse　property　of　signals　better,　a　mliti-directional　adaptive　sparse　model　and　recovery　algorithm　for　it　in　compressive　sensing　were　proposed.　The　mliti-directional　autoregressive　model　could　use　the　local　statistical　correlation　and　texture　directions　of　image　to　represent　signal　sparsely.　In　a　transform　based　codec　framework,　the　transform　matrix　was　regarded　as　a　measurement　matrix.　The　traditional　inverse　transform　in　decoder　is　replaced　by　the　multidirectional　adaptive　sparse　model.　Simulation　results　over　a　wide　range　of　images　show　that　the　proposed　technique　can　improve　the　reconstruction　quality　of　JPEG.