Self-Paced Balance Learning for Clinical Skin Disease Recognition

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Class imbalance in datasets leads to biased classification results, particularly affecting minority classes with fewer training samples.
• Existing methods primarily focus on resampling instances to balance class distribution, often neglecting the varying recognition difficulty across categories.

Purpose of the Study:

• To propose a novel algorithm, self-paced balance learning (SPBL), that addresses class imbalance by considering both sample quantity and recognition difficulty.
• To introduce a comprehensive metric, the complexity of image category, integrating sample number and recognition difficulty for a more nuanced approach to imbalance.

Main Methods:

• The SPBL algorithm initializes class complexity using a self-paced learning model.
• Penalty weights are assigned to classes, with higher weights for more complex categories and lower weights for easier ones.
• The training curriculum is reconstructed by rearranging samples based on assigned weights, enabling iterative learning of discriminative representations.

Main Results:

• Experimental results on SD-198 and SD-260 datasets show SPBL outperforms state-of-the-art methods in classification tasks.
• The algorithm demonstrates effective generalization capabilities across diverse tasks, including clinical skin disease recognition, indoor scene recognition, and object classification.

Conclusions:

• The proposed SPBL algorithm effectively mitigates class imbalance by balancing sample number and recognition difficulty.
• SPBL offers a promising approach for improving classification performance in scenarios with imbalanced and complex datasets.