Mural classification network based on the fusion of shallow feature enhancement and dense attention aggregation

This paper proposes an enhanced DenseNet-based strategy to improve mural classification performance. Briefly, a new independent component layer structure separates the larger 7 × 7 convolution kernels in the input layer into the trunk of three series of 3 × 3 small convolution kernels. The original ReLU activation is replaced with Mish activation. An attention-based feature filtering aggregation bottleneck is designed and the maximum pooling layer is used instead of global average pooling. A fully connected layer is added for richer contour and shape feature extraction. The accuracy, precision, recall and F1 value of the model are 93.07%, 93.86%, 93.08% and 93.37%, respectively, whose accuracy increased by 7.26% compared with the original DenseNet-121 model. The proposed model outperforms mainstream classification algorithms to a certain extent in terms of classification accuracy, generalization ability and stability, which is highly important for the large-scale screening and protection of mural images and the construction of digital museums.