Scene　classification　is　an　important　task　for　computer　vision,　and　Convolutional　Neural　Networks,　a　model　of　deep　learning,　is　widely　used　for　object　classification.　However,　they　rely　on　pooling　and　large　fully　connected　layers　to　combine　information　from　spatially　disparate　regions;　these　operations　can　throw　away　useful　fine-grained　information,　and　in　natural　scenes,　there　are　many　useless　information　which　will　increase　computation　cost.　In　this　paper,　mid-level　discriminative　patches　are　utilized　to　pre-process　the　full　images.　The　proposed　method　which　combines　mid-level　discriminative　patches　for　preprocessing　with　CNN　for　feature　extraction　improved　the　efficiency　of　computation　and　are　more　suitable　for　classifying　scenes.　Firstly,　full　images　are　divided　into　discriminative　parts.　Then　utilize　these　patches　to　go　through　CNN　for　feature　extraction.　Finally,　a　support　vector　machine　will　be　used　to　classify　the　scenes.　Experimental　evaluations　using　MIT　67　indoor　dataset　performs　well　and　proved　that　proposed　method　can　be　applied　to　scene　classification.