PECA-Net: Pyramidal Attention Convolution Residual Network for Architectural Heritage Images Classification

Heritage buildings have essential cultural and economic values, with constant change and destruction. Existing technologies that classify architectural heritage images can promote the recording and protection of heritage buildings. However, the background information of architectural heritage images is complex and changeable, making it difficult to extract the key feature information. In order to address the above problem, this paper proposes a pyramidal attention convolution residual network (PECA-Net) based on ResNet18, in which pyramidal convolution (PyConv) and attention mechanism have been adopted. PyConv is introduced into residual blocks instead of standard convolution, which can extract detailed information on different scales without increasing the parameter space of the model. Then, we propose a dual-pooling channel attention mechanism (DP-ECA) that effectively improves the ability to extract key information. Compared with the original ResNet18, experimental results show that the accuracy of our model is increased by 1.86%, and the parameters and FLOPs (floating point operations) are reduced by 0.48M and 0.28G respectively.