As　a　computer　vision　problem,　image　inpainting　aiming　at　restoring　the　missing　portions　within　an　image　has　been　widely　studied.　Motivated　by　inpainting,　we　begin　to　research　outpainting,　which　extrapolates　beyond　image　boundaries.　Then　an　image　can　be　expanded　arbitrarily　by　recursive　outpainting.　Image　outpainting　based　on　deep　learning　can　be　regarded　as　an　information　generation　problem　from　the　input　image　to　the　corresponding　output　image.　With　deep　convolutional　network　and　conditional　generative　adversarial　structure,　our　trained　model　generates　the　plausible　outpainting　contents　of　a　specific　image.　We　adopt　two　subnetworks　based　on　residual　net　and　U-net　to　replace　the　single　encoder-decoder　structure　in　the　most　existing　image　inpainting　networks.　The　first　half　of　the　generator　network　is　used　to　generate　rough　outline　images,　while　the　second　half　is　used　to　enhance　outline　images　to　clear　outpainted　images　with　pixel　to　pixel　mapping.　Experimental　results　show　that　the　generated　contents　are　smoothly　integrated　with　the　pixels　around　the　boundaries　and　express　real　semantics　together　with　the　input　image.　©　2019　IEEE.