Multi-level Semantic Fusion Network For Few-shot Multimedia Image Recognition In Education Management

Conventional few-shot multimedia image recognition methods in education management ignore the important semantic information in the training samples, resulting in insufficient feature learning, which is difficult to solve the problem of large intra-class variation. In this paper, we propose a global feature learning method based on multi-level semantic fusion. Specifically, according to the characteristics of different semantic levels of training samples, we design and implement different semantic learning tasks at the sample level, class level and task level, respectively. The semantic learning task and few-shot image classification task are integrated into the same architecture through the multi-task learning framework, which fuses the multi-level semantic information of categories and the discriminative information between classes. Therefore, the model can learn the category features from multiple perspectives, better find the commonality between samples with large differences, enhance the representativeness of the features. Compared with the baseline method, the large accuracy improvement is obtained on three datasets.