Semi-Supervised Multimodal Fusion for Lightweight Multi-Label Image Classification
A multi-label image classification system that fuses a lightweight CNN image branch (EfficientNet-family backbones) with a compact pretrained language-model branch (BERT-Tiny/ALBERT) over image captions, concatenating the two feature streams in a fusion module trained with per-label sigmoid outputs. Training data are expanded by iterative semi-supervised self-training: ensembled model predictions on unlabeled test images become pseudo-labels that are merged back into the training set over repeated rounds until validation performance plateaus. Ablations attribute the gains to the caption branch, the fusion module, and the pseudo-label cycle.
Google is all you need: Semi-Supervised Transfer Learning Strategy For Light Multimodal Multi-Task
As the volume of digital image data increases, the effectiveness of image classification intensifies. This study introduces a robust multi-label classification system designed to assign multiple labe…