Image Multi-label classification (MLC) datasets are often partially labeled due to expensive annotation costs.

In the context of supervised learning, the lack of known labels raises a significant difficulty for training deep neural network MLC models. Most existing MLC approaches assume that datasets are fully labeled, which limits research on the effective training of MLC models using partially labeled datasets—those more closely aligned with real-world scenarios.

 

This thesis aims to address such issues by developing approaches that can effectively train MLC models with various partially labeled datasets:

(1) The performance of models trained with Assume Negative-based and pseudo-labeling approaches is inevitably harmed by wrong pseudo-labels.

To this end, the proposed post-trained method called Category-wise Fine-Tuning employs clean known labels to fine-tune the logistic regressions of trained models individually to precisely calibrate predictions of each category.

(2) Existing pseudo-labeling approaches usually produce poorly calibrated models. To this end, the proposed framework called calibration-aware Thresholding-free Pseudo-Labeling produces pseudo-labels based on probabilities calibrated by an interactively learned individual graph neural network calibrator, plus an innovative re-parameterization method to reduce calibration errors without introducing extra inference costs.

(3) Due to the existence of unknown labels, most sample-mixing data augmentation methods cannot be effectively utilized to enlarge the magnitude of training samples to avoid overfitting. To this end, the proposed new method called LogicMix naturally mixes unknown labels by OR's logical equivalences, without replacement with constants.

 

The proposed three approaches are comprehensively evaluated on at least 3 public benchmarking datasets. They present competitive results to other state-of-the-art approaches.

In particular, state-of-the-art performance on 5 datasets is achieved, including CheXpert, MS-COCO, OpenImages-v3, VG-200, and Pascal VOC 2007.