Fuzzy Overclustering: Semi-Supervised Classification of Fuzzy Labels with Overclustering and Inverse Cross-Entropy

Area of Science:

• Computer Science
• Machine Learning
• Image Classification

Background:

• Deep learning requires large labeled datasets, a challenge for underwater image classification.
• Existing semi-supervised methods struggle with fuzzy labels common in uncurated real-world data due to ambiguous class boundaries.
• Fuzzy labels arise from limited image information and transitional object stages, leading to expert disagreement.

Purpose of the Study:

• To propose a novel framework for semi-supervised classification that effectively handles fuzzy labels.
• To introduce a new loss function that enhances the overclustering capability for improved fuzzy label classification.
• To demonstrate the superiority of the proposed framework over state-of-the-art methods for real-world fuzzy-labeled datasets.

Main Methods:

• Developed a novel semi-supervised classification framework utilizing overclustering to identify substructures within fuzzy labels.
• Introduced a new loss function designed to enhance the overclustering performance of the framework.
• Evaluated the framework on real-world plankton datasets exhibiting fuzzy labels.

Main Results:

• The proposed framework demonstrated superior performance compared to existing state-of-the-art semi-supervised methods.
• Overclustering proved beneficial for handling fuzzy labels in underwater image classification tasks.
• Achieved 5-10% more consistent predictions of substructures within the fuzzy-labeled data.

Conclusions:

• The novel framework effectively addresses the challenge of semi-supervised classification with fuzzy labels.
• Overclustering is a viable strategy for improving classification accuracy and consistency in ambiguous datasets.
• The method shows significant promise for applications in underwater image analysis and other domains with uncurated data.