Finding effective information in scene is a key problem in scene recognition.In order to construct an effective representation of scene image

a scene recognition method based on multi-scale attention network is proposed.The improved channel attention structure is integrated into the module

the local features and global features worthy of attention are obtained.To solve the problem of information loss in the process of spatial attention calculation

a spatial attention structure based on different scales is proposed

and the final representation of scene image is obtained by using complementary features.Furthermore

the center loss function joint supervision strategy is introduced to further reduce the misjudgment of intra-class differences.Experiments on Indoor 67 and SUN 397 datasets demonstrate that the proposed model has better identification performance than other classification methods in scene recognition.