With　the　development　of　visual　social　networks,　the　sentiment　analysis　of　images　has　quickly　emerged　for　opinion　mining.　Based　on　the　observation　that　the　sentiments　conveyed　by　some　images　are　related　to　salient　objects　in　them,　we　propose　a　scheme　for　visual　sentiment　analysis　that　combines　global　and　local　information.　First,　the　sentiment　is　predicted　from　the　entire　images.　Second,　it　is　judged　whether　there　are　salient　objects　in　an　image　or　not.　If　there　are,　sub-images　are　cropped　from　the　entire　image　based　on　the　detection　window　of　the　salient　objects.　Moreover,　a　CNN　model　is　trained　for　the　set　of　sub-images.　Predictions　of　sentiments　from　entire　images　and　sub-images　are　then　fused　together　to　obtain　the　final　results.　If　no　salient　object　is　detected　in　the　images,　the　sentiment　predicted　directly　from　entire　images　is　used　as　the　final　result.　The　compared　experimental　results　show　that　the　proposed　approach　is　superior　to　state-of-the-art　algorithms.　It　also　demonstrates　that　reasonably　utilizing　the　local　information　could　improve　the　performance　for　visual　sentiment　analysis.