DeepSMOTE

DeepSMOTE generates synthetic minority-class images in a latent space by integrating SMOTE interpolation with an encoder/decoder to balance imbalanced image datasets for deep learning model training.

Key Features:

• Encoder/Decoder Framework: Employs an encoder to compress raw image data into a latent space and a decoder to reconstruct images after oversampling, preserving essential features.
• SMOTE-based Oversampling in Latent Space: Applies SMOTE interpolation between minority-class latent vectors to produce diverse synthetic samples within the latent representation.
• Dedicated Loss Function with Penalty Term: Uses a tailored loss function that includes a penalty term to optimize the realism and informativeness of generated synthetic images.
• Latent-space Reconstruction: Reconstructs oversampled latent vectors back to image space via the decoder to obtain usable synthetic images for training.
• No GAN Discriminator: Operates without a generative adversarial discriminator, distinguishing it from GAN-based oversampling methods.

Scientific Applications:

• Medical Imaging: Balances underrepresented classes in medical imaging datasets by generating synthetic minority-class images.
• Remote Sensing: Augments minority-class samples in remote sensing image datasets to mitigate class imbalance.
• Deep Learning Model Training: Enhances representation of minority classes to improve training robustness and generalization of deep learning models on imbalanced image data.

Methodology:

Encode raw images into a latent space, apply SMOTE interpolation between minority-class latent vectors to generate synthetic latent samples, reconstruct synthetic latent samples via the decoder, and optimize generation with a dedicated loss function that includes a penalty term.