Architectural style classification based on CNN and channel-spatial attention

The accurate classification of architectural styles is of great significance to the study of architectural culture and human historical civilization. Models based on convolutional neural network (CNN) have achieved highly competitive results in the field of architectural style classification owing to its more powerful capability of feature expression. However, most of the CNN models to date only extract the global features of architecture facade or focus on some regions of architecture and fail to extract the spatial features of different components. To improve the accuracy of architectural style classification, we propose an architectural style classification method based on CNN and channel-spatial attention. Firstly, we add a preprocessing operation before CNN feature extraction to select main building candidate region in architectural image and then use CNN feature extractor for deep feature extraction. Secondly, channel-spatial attention module is introduced to generate an attention map, which can not only enhance the texture feature representation of architectural images but also focus on the spatial features of different architectural elements. Finally, the Softmax classifier is used to predict the score of the target class. The experimental results on the Architectural Style Dataset and AHE_Dataset have achieved satisfactory performance.