Fast and Efficient Image Generation Using Variational Autoencoders and K-Nearest Neighbor OveRsampling Approach

In this paper, we propose a novel approach to image generation that combines Variational Autoencoders (VAEs) with the K-Nearest Neighbor OveRsampling (KNNOR) technique. Traditional Generative Adversarial Networks (GANs), despite their popularity, face challenges such as convergence issues, mode collapse, and image distortion. Our method addresses these problems by leveraging the strengths of VAEs and the robustness of KNNOR to produce high-quality, life-like images with significantly reduced distortion.

We conducted experiments to compare our method against Deep Convolutional GANs (DCGANs) and demonstrated that our approach not only generates more convincing images but also achieves this with models that are half the size of DCGANs. Additionally, we compared the efficacy of KNNOR with the Synthetic Minority Oversampling Technique (SMOTE), further highlighting the advantages of our proposed method. The code for this research has been made publicly available on GitHub, facilitating broader adoption and experimentation.

Recommended citation: Islam, A., & Belhaouari, S. B. (2023). "Fast and Efficient Image Generation Using Variational Autoencoders and K-Nearest Neighbor OveRsampling Approach." IEEE Access, 11, 28416--28426. https://doi.org/10.1109/ACCESS.2023.3259236
Download Paper | Download Slides