Different　algorithms　have　been　proposed　for　Compressed　Sensing　(CS).　One　of　the　most　popular　frameworks　is　Orthogonal　Matching　Pursuit　(OMP).　And　there　are　many　variants　of　it.　Among　various　versions,　a　family　of　algorithms　treats　the　distribution　over　an　original　signal　as　prior　knowledge　to　obtain　a　training　set　for　the　model,　and　it　achieves　a　good　performance.　However,　there　is　other　trivial　prior　knowledge　about　the　measurement　matrix　that　has　never　been　used　in　compressed　sensing　in　previous　work.　Hence,　we　propose　a　new　method　to　embed　the　prior　knowledge　about　the　measurement　matrix　and　distribution　over　the　original　signal　into　the　neural　networks　for　CS.　In　the　end,　the　empirical　support　shows　that　the　proposed　method　brings　out　a　significant　improvement.