Majority clustering for imbalanced image classification.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Class imbalance is a common problem in image classification, leading to biased models that underperform on minority classes.
• This imbalance negatively impacts the overall reliability and performance of image classification systems.
• Existing methods often struggle to effectively address imbalances both between and within classes.

Purpose of the Study:

• To introduce a novel under-sampling technique, Majority Clustering for Imbalanced Image Classification (MCIIC), to mitigate class imbalance in image datasets.
• To transform binary classification problems with imbalanced data into multi-class problems for a more balanced solution.
• To improve the detection of rare samples within datasets.

Main Methods:

• An under-sampling approach is employed, focusing on reducing majority class samples.
• Unsupervised clustering is used to partition the majority class into distinct clusters.
• The elbow method is utilized to determine the optimal number of clusters for the majority class, with each cluster assigned a new label.

Main Results:

• The MCIIC technique effectively creates a more balanced class distribution, addressing imbalances both between and within classes.
• Empirical evaluations on benchmark datasets demonstrate significant improvements in predictive performance for imbalanced image datasets.
• The study shows positive impacts on model accuracy, precision, recall, and F1-score.

Conclusions:

• MCIIC is a practical and effective pre-processing step for handling imbalanced image datasets.
• The proposed method offers a significant improvement over traditional approaches for imbalanced classification tasks.
• This technique enhances the reliability and performance of machine learning models dealing with skewed data distributions.