In SSPML( Semi-Supervised Partial Multi-Label Learning), complex correlations exist between the feature and label, and within them. Models leveraging label correlations are susceptible to interference from noisy labels in candidate label sets. T-SSPML ( Transformer Encoder-Based Semi-Supervised Partial Multi-Label Learning), a novel semi-supervised partial multi-label image classification algorithm is proposed based on Transformer. A backbone network is constructed using Transformer encoders, taking both image features and label features as joint inputs, and a new loss function is defined. The multi-head self-attention mechanism is employed to capture holistic structural information and label interdependencies. A label masking mechanism is

introduced to enhance the model’s capability in learning label correlation information. Extensive experiments on multiple datasets demonstrate that the T-SSPML algorithm achieves superior classification performance compared to other representative methods in most cases.

partial multi-label learning, semi-supervised learning, image classification, consistency regularization, pseudo-labeling methods, Transformer encoder