Image　super-resolution　reconstruction　is　to　use　a　single　or　a　set　of　degraded　images　to　produce　a　high　resolution　image,　to　overcome　the　limitation　or　ill-posed　conditions　of　the　image　acquisition　process　to　achieve　better　content　visualization　and　scene　recognition.　This　paper　proposes　a　super　resolution　reconstruction　algorithm　based　on　the　combination　of　compressed　sensing　and　depth　perception　neural　networks.　The　algorithm　originally　makes　use　of　a　double　pyramid　of　images,　built　starting　from　the　input　image　itself,　to　extract　the　dictionary　patches,　and　employs　a　regression　based　method　to　directly　map　the　low-resolution　(LR)　input　patches　into　their　related　high-resolution　(HR)　output　patches.　With　the　integration　of　deep　neural　network　architecture　and　the　compressive　sensing　theory,　the　robustness　will　be　enhanced.　Experiments　on　natural　images　show　that　the　proposed　algorithm　outperforms　some　of　the　state-of-the-art　algorithm　in　terms　of　peak　signal　to　noise　ratio,　mean　square　error　and　structural　similarity　index.