Image　super-resolution　is　widely　used　and　research　on　its　algorithm　are　also　developed　rapidly.　In　recent　years,　deep　learning　has　been　introduced　into　the　process　of　image　super-resolution,　and　the　output　image　has　been　improved　effectively.　On　this　foundation,　this　paper　proposes　to　reconstruct　the　mapping　layer　with　a　method　that　reduces　the　dimension　of　extracted　features　at　first　and　then　extends　them　at　last.　Also,　a　deconvolution　layer　is　used　at　the　end　of　the　network　to　map　an　uninterpolated　low-resolution　image　to　a　high-resolution　image　directly.　Besides,　smaller　convolution　kernels　and　more　mapping　layers　are　used　in　this　algorithm.　Comparative　experiments　demonstrate　that　the　above　methods　can　accelerate　the　speed　and　increase　the　effectiveness　of　the　image　reconstruction　by　optimizing　the　network　structure　and　reducing　the　computational　complexity.