Script identification of ancient books by Chinese ethnic minorities using multi-branch DCNN and SPP

Automatic classification of ancient books is an important component of the digital platform of ancient books, while automatic classification of ancient books is more challenging. In view of the ancient books script identification task of different ethnic minorities in China, this paper proposes a deep convolutional neural network (CNN) ancient books script identification method with multi-branch structure and spatial pyramid pooling (SPP), called MbSPPVGG. We build a dataset of Chinese ethnic ancient handwritten books, and crop and standardize preprocessing images of ancient books. In order to improve the identification accuracy of ancient books and ability of CNN to perceive multi-scale changes in image, bottom-level and high-level features of CNN are merged by multi-branch structure to enhance the networks expression ability, and then use SPP to multi-scale de-dimensionality of convolutional features, increase the spatial scale invariance of CNN. The introduction of multi-branch structure and SPP in the CNN model constitutes a new ancient books identification model. The experimental results show that the precision, recall and F1-score of MbSPPVGG model are all 99.94%. As demonstrated by comparison experiments, the classification accuracy of MbSPPVGG model is better than that of state-of-the-art GhostNet, CSPDenseNet, MixNet and other deep learning methods, and its effectiveness is verified on multiple datasets.