Label propagation via local geometry preserving for deep semi-supervised image recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep semi-supervised image recognition relies on pseudo-labeling for training with limited labeled data.
• Existing methods often suffer from noisy information flow and loss of local geometry during label propagation.
• This can lead to biased feature mapping and overfitting to noise in the learned representations.

Purpose of the Study:

• To propose a novel transductive pseudo-labeling method for improved deep semi-supervised image recognition.
• To address the issues of noisy information flow and loss of local geometry in current approaches.
• To enhance the accuracy and robustness of image recognition models trained with scarce labeled data.

Main Methods:

• Incorporating self-supervised learning into feature learning for cleaner information flow.
• Utilizing a reconstruction concept to measure pairwise similarity in feature space, preserving local geometry.
• Developing a transductive pseudo-labeling framework that minimizes information loss from labeled to unlabeled data.

Main Results:

• Self-supervised learning yields cleaner feature representations for subsequent label propagation.
• The reconstruction-based similarity measure effectively preserves local geometry in the feature space.
• Ablation studies confirm the synergistic benefits of the proposed self-supervised features and geometry-preserving similarity graph.
• Extensive experiments on benchmark datasets demonstrate the superior effectiveness of the proposed method.

Conclusions:

• The proposed method significantly improves deep semi-supervised image recognition by ensuring noiseless and low-loss information flow.
• Integrating self-supervised learning and reconstruction-based similarity enhances feature learning and preserves crucial local geometric information.
• The novel approach offers a more robust and accurate solution for image recognition tasks with limited labeled data.