Impact of imbalanced features on large datasets.

Area of Science:

• Computer Vision and Machine Learning
• Data Science and Artificial Intelligence

Background:

• Exponential growth in image and video data necessitates efficient content-based searching.
• Feature-based image classification is vital but challenged by imbalanced class distributions.
• Real-world datasets often exhibit uneven class instances, hindering classifier performance.

Purpose of the Study:

• To explore image classification frameworks considering balanced and imbalanced feature distributions.
• To analyze the impact of class imbalance on image classification performance, particularly on large datasets.
• To identify effective techniques for mitigating class imbalance in image classification.

Main Methods:

• Evaluation of classification frameworks using image features under balanced and imbalanced conditions.
• Extensive experimentation on large-scale datasets to assess performance variations.
• Comparison of different balancing techniques, including Distributed Gaussian (D-GA) and Distributed Poisson (D-PO).

Main Results:

• All tested models demonstrated superior performance with dataset balancing compared to imbalanced datasets.
• Distributed Gaussian (D-GA) and Distributed Poisson (D-PO) emerged as highly effective balancing techniques.
• Significant performance improvements were observed for Random Forest (RF), Support Vector Machine (SVM), and deep learning models.

Conclusions:

• Dataset balancing is crucial for enhancing the accuracy of image classification models.
• D-GA and D-PO techniques are particularly effective in improving performance for RF and SVM classifiers.
• Balancing strategies positively impact deep learning models, contributing to better overall image analysis.