Multi-level Feature Learning for Multimedia Pattern Recognition in English Education

In multimedia pattern recognition, particularly in English education, the sharing of local features between classes and their varying classification reliability are often overlooked in existing methods, which diminishes feature discrimination and complicates the handling of small inter-class variations. In this paper, a multi-level feature learning method based on enhanced local descriptors is proposed for mining multimedia patterns in English educations (MFL). Specifically, multi-scale global information is extracted using the pyramid aggregation network and fused with local features to enhance inter-class uniqueness. During classification, local descriptors that better distinguish between classes are emphasized, resulting in improved inter-class discrimination. MFL achieves a significant accuracy improvement over baseline methods across three datasets, offering a new perspective for analyzing multimedia content in English education.