DeepSMOTE: Fusing Deep Learning and SMOTE for Imbalanced Data.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Imbalanced data presents a significant challenge for machine learning models, particularly in deep learning for image analysis.
• Existing oversampling methods often struggle to preserve image properties or generate high-quality synthetic data suitable for deep learning.

Purpose of the Study:

• To introduce Deep synthetic minority oversampling technique (DeepSMOTE), a novel oversampling algorithm tailored for deep learning models.
• To address the need for an oversampling method that works on raw images, preserves their characteristics, and enhances minority classes.

Main Methods:

• DeepSMOTE utilizes an encoder/decoder framework combined with SMOTE-based oversampling.
• A dedicated loss function, enhanced with a penalty term, is employed for effective training.
• The algorithm operates directly on raw images, maintaining their inherent properties.

Main Results:

• DeepSMOTE generates high-quality, information-rich artificial images suitable for visual inspection.
• The method effectively enhances minority classes and balances training datasets for deep learning models.
• Compared to Generative Adversarial Network (GAN)-based methods, DeepSMOTE does not require a discriminator.

Conclusions:

• DeepSMOTE offers a simple yet effective solution for handling imbalanced data in deep learning image tasks.
• The generated synthetic images are valuable for improving model robustness and performance on underrepresented classes.
• This approach provides an advantageous alternative to GAN-based oversampling techniques.