Bronze Inscriptions Classification Algorithm On Imbalanced Dataset

Bronze inscriptions are important data for the study of ancient Chinese history. We obtained a dataset included about 100000 images of bronze inscriptions organized by professionals. However, the data distribution is imbalanced--minority classes occupy most of the sample. In this paper, we designed a variety of models based on Convolutional Neural Network(CNN) to classify the dataset. First of all, we added Convolutional Block Attention Module(CBAM) to ResNet. Secondly, we added the sample weight when calculating the train loss. At last, we used Two-Phases-Training method in order to get better results. It has two different ways-only fine-tuning the fully connected(FC) layer and adjusting all layers. According to experimental results, the Two-Phases-Training method not only improves the average accuracy(Avg_acc), but also reduces the probability that the model's output is biased towards major classes. Among them, fine-tuning the FC layer has the best performance during the second phase. Not only the training time was the shortest, but also the Avg_acc and Fl_score are the best.