A Teaching Figure Classification Based on the Fusion of Multiple Modals

Deep learning-based classification methods can help the teaching system automatically collect figure samples while the samples from different sources cause the difficulty of the classification. To solve this problem, we use multiple models to increase the accuracy of the classification. In more details, we pretrain the deep learning models on some public datasets. Then, we continuously train these models on teaching figure datasets. On a figure sample, we firstly classify the dataset that may contain this sample. After we predicted the dataset, we apply a weighted fusion method to classify the label of this sample. As the experimental results shows, our methods can achieve higher accuracy than the existing ones.