It　is　prevailing　to　apply　the　Deep　learning　model　to　several　problems　related　to　autonomous　driving.　Usually,　these　solutions　depend　on　a　large　scale　of　networks,　especially　the　Convolution　Neutral　Network(CNN),　requiring　some　specific　databases　of　real　image　samples　targeted　at　the　certain　issue　for　some　proper　trainings.　The　pursue　for　the　recognition　itself　is　not　always　necessary　in　the　autonomous　driving　context　if　the　experiment　ignores　the　speed　and　the　accuracy　of　the　recognition　of　these　traffic　signs　since　these　autonomous　driving　vehicles　will　not　have　a　plenty　of　time　on　the　road,　just　as　in　the　laboratory,　to　recognize　what　the　traffic　sign　in　the　camera　is.　In　order　to　improve　the　efficiency　and　the　accuracy　of　the　recognition　of　these　traffic　signs,　this　paper　proposes　a　novel　CNN-based　algorithm,　which　contains　more　convolution　layers　and　fully-connected　layers.　In　this　case,　in　order　to　meet　the　demand　of　the　novel　CNN　structure　taken　in　the　experiment,　the　dataset　has　also　been　modified　ahead　of　the　experiment,　which　has　been　called　the　preprocessing　operation.　The　dataset　which　has　been　chosen　as　the　dataset　for　this　experiment　is　the　German　Traffic　Sign　Benchmarks　(GTSRB).　The　dataset　generation　method　has　also　been　used　in　the　experiment　which　requires　only　(i)　common　natural　images,　and　(ii)　templates　of　the　traffic　signs,　which　are　created　to　illustrate　the　meaning　and　the　category　of　a　traffic　sign.　The　generated　training　dataset　is　shown　to　be　effective　when　using　the　novel　CNN　structure　proposed　by　this　paper.　The　experiment　result　has　achieved　to　0.069783　oftrain　loss　and　0.030411　ofvalid　loss.　The　experiment　surprisingly　shows　that　CNN　can　be　trained　with　data　generation　methods　to　get　a　dramatically　more　efficient　and　accurate　result　which　is　usually　ignored　in　the　formal　mainstream　papers　©　2019　IEEE.